Automated Detection of Complex

Tactical Patterns in Football

Using Machine Learning Techniques to Identify Tactical
Behavior

Doctoral Thesis
in order to obtain the title of Doctor from the Faculty of Economics and
Social Sciences at the University of Tübingen

presented by
M. Sc. Pascal Bauer
Sankt Ingbert

Tübingen, 2021
1st supervisor: Prof. Dr. Oliver Höner
2nd supervisor: Prof. Dr. Augustin Kelava
3rd supervisor: Prof. Dr. Ulf Brefeld

Date of the oral defense: 13.01.2022

Dean: Prof. Dr. Josef Schmid
1st supervisor: Prof. Dr. Oliver Höner
2nd supervisor: Prof. Dr. Augustin Kelava
3rd supervisor: Prof. Dr. Ulf Brefeld
Acknowledgements

This dissertation is a part of a broader research program con-

ducted by Eberhard Karls University Tübingen, DFB-Akademie,
Deutsche Fußball-Liga (DFL) and Sportec Solutions AG. Another
major pillar of the project is the thesis of Gabriel Anzer, which
has been conducted in close collaboration.
First, I would like to thank Prof. Dr. Oliver Höner as the main
supervisor of both theses guiding our research with critical and
valuable feedback and for supporting us whenever necessary.
Second, I would like to thank all co-authors involved, especially
Gabriel Anzer, Joshua Wyatt Smith (PhD) and Prof. Dr. Ulf
Brefeld, for informative collaborations and discussions building
a central component of the work presented. I also want to thank
Prof. Dr. Augustin Kelava for guiding our research in various
discussions.
Additionally, I would like to thank DFL, DFB-Akademie and
Sportec Solutions AG for providing the positional and event data
for the studies, and supporting the dissertations.
This work would also not have been possible without the per-
spective of professional match-analysts and coaches from world
class teams who helped us to define relevant features and spend
much time evaluating (intermediate) results. I would also like
to cordially thank Dr. Stephan Nopp, Christofer Clemens (head
match-analysts of the German mens National team), Jannis Scheibe
(head match-analyst of the German U21 mens national team),
Leonard Höhn (head match-analyst for the womens national team)
as well as Sebastian Geißler (former match-analyst of Borussia
Mönchengladbach).

i
Contents

Contents ii

List of Publications viii

1 Introduction 1

2 State of the Art 7

2.1 Positional and Event Data in Football . . . . . . . . 7
2.1.1 Event Data . . . . . . . . . . . . . . . . . . . . 7
2.1.2 Positional Data . . . . . . . . . . . . . . . . . 9
2.1.3 Dataset of Bundesliga and German National
Team . . . . . . . . . . . . . . . . . . . . . . . 11
2.2 Machine Learning Basics for Sports Applications . 11
2.3 Data-Driven Detection of Tactical Patterns in Sport 17
2.3.1 Tactical Patterns in Invasion Sports . . . . . 17
2.3.2 The Detection of Tactical Patterns in Inva-
sion Sports . . . . . . . . . . . . . . . . . . . . 18
2.3.3 The Detection of Tactical Patterns in Football 21
2.3.4 Phases of Play in Football . . . . . . . . . . . 30

3 Empirical Studies 33
3.1 Study I: Constructing Spaces and Times for Tacti-
cal Analysis in Football (Andrienko et. al. 2019) . . 33
3.2 Study II: The Origins of Goals in the German Bun-
desliga (Anzer, Bauer, & Brefeld 2021) . . . . . . . . 35
3.3 Study III: Data-Driven Detection of Counterpress-
ing in Professional Football (Bauer & Anzer 2021) . 36
3.4 Study IV: Putting Team Formations in Association
Football into Context (Bauer, Anzer, & Shaw 2021) . 40

ii
 3.5 Study V: Individual Role Classification for Play-
ers defending Corners in Football (Bauer, Anzer,
& Smith 2021) . . . . . . . . . . . . . . . . . . . . . . 43
3.6 Study VI: Torward Automatically Labeling Situa-
tions in Football (Fassmeyer et. al. 2021) . . . . . . . 46

4 Discussion 47
4.1 Limitations and Future Work . . . . . . . . . . . . . 52

5 Conclusion 55

References 56

A Appendix—Study I: Constructing Spaces and Times for

Tactical Analysis in Football 76

B Appendix—Study II: The Origins of Goals in the Ger-

man Bundesliga 96

C Appendix—Study III: Data-Driven Detection of Coun-

terpressing in Professional Football 120

D Appendix—Study IV: Putting Team Formations in As-

sociation Football into Context 162

E Appendix—Study V: Individual Role Classification for

Players Defending Corners in Football (Soccer) 182

F Appendix—Study VI: Toward Automatically Labeling

Situations in Soccer 200

iii
List of Publications

The following publications/submissions are in the core of the

cumulative dissertation:

(I) Andrienko, G., Andrienko, N., Anzer, G., Bauer, P., Budziak,
G., Fuchs, G., Hecker D., Weber H., Wrobel, S. (2019). Con-
structing Spaces and Times for Tactical Analysis in Foot-
ball. IEEE Transactions on Visualization and Computer Graph-
ics, 27(4), 2280–2297. https://doi.org/10.1109/TVCG.2019
.2952129

(II) Anzer, G., Bauer, P., & Brefeld, U. (2021). The Origins of
Goals in the German Bundesliga. Journal of Sport Science.
https://doi.org/10.1080/02640414.2021.1943981

(III) Bauer, P., Anzer, G. (2021). Data-Driven Detection of Coun-

terpressing in Professional Football—A Supervised Machine
Learning Task based on Synchronized Positional and Event
Data with Expert-Based Feature Extraction. Data Mining
and Knowledge Discovery, 35(5), 2009–2049.
https://doi.org/10.1007/s10618-021-00763-7

(IV) Bauer, P., Anzer, G., Shaw, L. (2022). Putting Team Forma-
tions in Association Football into Context. Journal of Sports
Analytics (submitted).

(V) Bauer, P., Anzer, G., Smith, J. W. (2022). Individual role

classification for players defending corners in football (soc-
cer)—Categorisation of the defensive role for each player in
a corner kick using positional data. Journal of Quantitative
Analysis in Sports (submitted).

(VI) Fassmeyer, D., Anzer, G., Bauer, P., Brefeld, U. (2021). To-
ward Automatically Labeling Situations in Soccer. Frontiers

iv
 in Sports and Active Living, 3(November). https://doi.org/
10.3389/fspor.2021.725431

Additional, publications/submissions conducted within the scope

of this research program:

(i) Anzer, G., Bauer, P. (2021). A Goal Scoring Probability

Model based on Synchronized Positional and Event Data.
Frontiers in Sports and Active Learning (Special Issue: Using
Artificial Intelligence to Enhance Sport Performance), 3(0), 1–18.
https://doi.org/10.3389/fspor.2021.624475

(ii) Anzer, G., Bauer, P. (2022). Expected Passes—Determining

the Difficulty of a Pass in Football (Soccer) Using Spatio-
Temporal Data. Data Mining and Knowledge Discovery, Springer
US. https://doi.org/10.1007/s10618-021-00810-3.

(iii) Herold, M., Goes, F., Nopp, S., Bauer, P., Thompson, C., &
Meyer, T. (2019). Machine Learning in Men’s Professional
Football: Current Applications and Future Directions for
improving Attacking Play. International Journal of Sports Sci-
ence and Coaching, 14(6).
https://doi.org/10.1177/1747954119879350

(iv) Herold, M., Kempe, M., Bauer, P., & Meyer, T. (2021). At-
tacking Key Performance Indicators in Soccer: Current Prac-
tice and Perceptions from the Elite to Youth Academy Level.
Journal of Sports Science and Medicine, 20(1), 158–169. https://
doi.org/10.52082/jssm.2021.158

(v) Anzer, G., Bauer, P., Höner, O. (2021). The Identification of

Counterpressing in Football. In D. Memmert (Ed.), Match
Analysis—How to Use Data in Professional Sport (1st Edi-
tio, pp. 228–235). New York: Routledge. https://doi.org/
https://doi.org/10.4324/9781003160953

v
(vi) Ric, A., Bradley, P., Shaw, L., Thies, H., Sumpter, D., López-
felip, M. A.,Ade J. Dixon D., James A., Evans M., Gómez-
díaz A., Harrison H., Laws A., Petersen M., Seirul P., Robert-
son S., Pollard, R., Bransen, L., Kempe M., & Bauer, P.
(2021). Football Analytics 2021: The Role of Context in
Transferring Analytics to the Pitch. Barça Innovation Sum-
mit 2020, Barcelona.

vi
Abstract

Football tactics is a topic of public interest, where decisions

are predominantly made based on gut instincts from domain-
experts. Sport science literature often highlights the need for
evidence-based research on football tactics, however the limited
capabilities in modeling the dynamics of football has prevented
researchers from gaining usable insights. Recent technological
advances have made high quality football data more available
and affordable. Particularly, positional data providing player and
ball coordinates at every instance of a match can be combined
with event data containing spatio-temporal information on any
event taking place on the pitch (e.g. passes, shots, fouls). On
the other hand, the application of machine learning methods to
domain-specific problems yields a paradigm shift in many in-
dustries including sports.
The need for more informed decisions as well as automat-
ing time consuming processes—accelerated by the availability
of data—has motivated many scientific investigations in football
analytics. This thesis is part of a research program combining
methodologies from sports and data science to address the fol-
lowing problems: the synchronization of positional and event
data, objectively quantifying offensive actions, as well as the de-
tection of tactical patterns. Although various basic insights from
the overall research program are integrated, this thesis focuses
primarily on the latter one.
Specifically, positional and event data are used to apply ma-
chine learning techniques to identify eight established tactical
patterns in football: namely high-/mid-/low-block defending, build-
up/attacking play in the offense, counterpressing and counterattacks
during transitions, and patterns when defending corner-kicks,

vii
e.g. player-/zonal- or post-marking. For each pattern, we consol-
idate definitions with football experts and label large amounts
of data manually using video recordings. The inter-labeler relia-
bility is used to ensure that each pattern is well-defined. Unsu-
pervised techniques are used for the purpose of exploration, and
supervised machine learning methods based on expert-labeled
data for the final detection. As an outlook, semi-supervised
methods were used to reduce the labeling effort. This thesis
proves that the detection of tactical patterns can optimize ev-
eryday processes in professional clubs, and leverage the domain
of tactical analysis in sport science by gaining unseen insights.
Additionally, we add value to the machine learning domain by
evaluating recent methods in supervised and semi-supervised
machine learning on challenging, real-world problems.

viii
1 Introduction

In 1987, Bate declared (association) football as a ’game of opin-

ions’ in which ’coaches and managers base strategy and tactics on their
own opinions’. Hence, he motivated that team strategy should
be based on something more substantial rather than opinions
and instincts. More than 30 years later, Kuper (2018) main-
tains a disproportionately high number of decisions are made
based on gut instincts in professional football. Aiming to build
more evidence around football tactics, Reep and Benjamin (1968)
were the first to systematically annotate data from professional
football matches. Since then, various studies on tactical per-
formance analysis used manually acquired statistics to conduct
studies on a substantial basis (Camerino, Chaverri, Anguera, &
Jonsson, 2012; Borrie, Jonsson, & Magnusson, 2002; Gould &
Gatrell, 1979). The major research question for such investiga-
tions in football was to study the efficiency of tactics and strategy
(Sarmento et al., 2018, 2014) more objectively in order to support
decision-making (Desporto, 2009). Following the idea of Reep
and Benjamin (1968), various sport science researchers utilized
coding systems to manually annotate match logs. By doing so,
they captured a rough extraction of a football match to answer
pre-defined research questions (Camerino et al., 2012; Alcock,
2010; Borrie et al., 2002).
The growing public interest in football and recent advances
in technology has enabled exhaustive data collection across all
professional football leagues (Beal, Norman, & Ramchurn, 2019;
Seidl, 2019). Companies1 commercialized the systematic collec-
tion of so called event data across several professional compe-
titions (Lucey, Oliver, Carr, Roth, & Matthews, 2013). Based
1 such as Sportec Solutions AG, Statsperform, Statsbomb or Wyscout

1
on a dedicated definition catalog, defining basic events (e.g.,
passes, shots, duels, . . .) in detail, manual operators annotate
each event with the support of software-tools (Pappalardo et al.,
2019). Even though event data comprise the central actions of a
football match, they only contain information regarding the few
players directly involved in a ball action (Borrie et al., 2002). This
problem was solved through the introduction of global position-
ing systems (Hennessy & Jeffreys, 2018) and the latest improve-
ments in image processing and computer vision (Manafifard,
Ebadi, & Abrishami Moghaddam, 2017; D’Orazio & Leo, 2010;
Barris & Button, 2008). So called positional data (often also re-
ferred to as movement or tracking data), containing the position of
every player and the ball across the whole match, became avail-
able and affordable in sports (Manafifard et al., 2017; Stein et al.,
2017).
Not limited to football, the combination of positional and
event data (often also referred to as play-by-play data) enabled a
change in paradigm for sport science (Sarmento et al., 2018; Link,
2018; Rein & Memmert, 2016) as well as in the everyday business
of clubs and federations (Herold, Kempe, Bauer, & Meyer, 2021;
Andrienko et al., 2017; Herold et al., 2019). Besides the well-
known Moneyball-story (MacLennan, 2005)—gaining notoriety
whereby a statistician provided evidence for transfer decisions to
Oakland Athletics’s general manager Billy Beane—a whole com-
munity of statistical researchers emerged in baseball (so called
Sabermetrics) (Albert, 2010; Baumer & Zimbalist, 2014) and later
in basketball analytics (so called APBRmetrics) (Stephanos, Husari,
Bennett, & Stephanos, 2021; Schumaker, Solieman, & Chen, 2010).
More recently, researchers tried to transfer the learnings to other
invasion sports such as American football (Atmosukarto, Ghanem,
Ahuja, Ahuja, & Muthuswamy, 2013), Australian rules football

2
(Sampaio & Maças, 2012), team handball (Pfeiffer & Perl, 2015),
rugby (Bunker, Fujii, Hanada, & Takeuchi, 2020) and (associa-
tion) football (F. R. Goes, Meerhoff, et al., 2020; Lucey, Oliver,
et al., 2013), establishing the research domain of sports analytics
(Araújo, Couceiro, Seifert, Sarmento, & Davids, 2021; Beal et al.,
2019; Link, 2018; Morgulev, Azar, & Lidor, 2018; Schumaker et
al., 2010).
This dissertation is not only motivated from a football per-
spective. Recent success in machine learning, a research domain
established since the 1950’s (Samuel, 1959) warrants application
domains—notably practical tasks a machine learning algorithm
can learn from data—to explore and evaluate new methods. In
this context, the dynamic nature of football, often referred to as
the most complex of invasion sports (Tuyls et al., 2021) provides
various problems challenging machine learning researchers. In
a position paper, Tuyls et al. (2021) outlines how applications of
artificial intelligence in football—due to its modeling complexity,
dynamics and the ubiquitous human component—can serve as
a valuable playing ground for machine learning research.
We focus on tactical (rather than physical) performance anal-
ysis in football, and on investigations to improve decision mak-
ing when applying tactics (rather than result-predictions). Rele-
vant studies include probabilistic models to quantify goal scor-
ing probabilities using expected goals (xG) values (Anzer & Bauer,
2021; Robberechts & Davis, 2020; Lucey, Bialkowski, Monfort,
Carr, & Matthews, 2014), pass completion probabilities using
expected pass (xPass) values (Anzer & Bauer, 2022; Spearman,
Basye, Dick, Hotovy, & Pop, 2017), or the goal scoring proba-
bility at any time-point in the match using expected possession
values (Fernández, Bornn, & Cervone, 2021; Spearman, 2018).
The above described metrics allow more contextual evaluation of

3
the key-events in football (e.g., passes, shots), facilitating to over-
come otherwise limited, binary evaluations (e.g., passes: com-
pleted or not; shots: successful or not). Expected pass approaches
are often combined with reward quantification of passes (Steiner,
Rauh, Rumo, Sonderegger, & Seiler, 2019; Power, Ruiz, Wei, &
Lucey, 2017; Rein, Raabe, & Memmert, 2017; Chawla, Estephan,
Gudmundsson, & Horton, 2017; F. Goes, Schwarz, Elferink-Gemser,
Lemmink, & Brink, 2021; Gómez-Jordana, Milho, Ric, Silva, &
Passos, 2019; F. R. Goes, Kempe, Meerhoff, & Lemmink, 2019)
allowing to quantify pass decisions considering risk and reward
of pass options compared to alternatives. Another continuously
addressed concept is the control of space and the development
of movement models (Martens, Dick, & Brefeld, 2021; Brefeld,
Lasek, & Mair, 2019; Fernandez & Bornn, 2018; Fujimura &
Sugihara, 2005; Taki, Hasegawa, & Fukumura, 1996). Teams’
playing styles (Beal, Chalkiadakis, Norman, & Ramchurn, 2020;
Kempe, Vogelbein, Memmert, & Nopp, 2014; Vogelbein, Nopp,
& Hökelmann, 2014) or formations (Gudmundsson, Laube, &
Wolle, 2017) were studied across the course of a match or a whole
season.
Various studies from the sports science domain (F. R. Goes,
Meerhoff, et al., 2020) used data to validate hypotheses via de-
ductive reasoning (Pappalardo et al., 2019; Sarmento et al., 2018;
Duarte et al., 2013; Bartlett, Button, Robins, Dutt-Mazumder,
& Kennedy, 2012). On the other hand, more data-driven ap-
proaches aim to investigate interesting patterns directly from the
high-dimensional, spatio-temporal data. Accordingly, two major
approaches can be distinguished: the supervised detection of pre-
defined patterns (Müller-Budack, Theiner, Rein, & Ewerth, 2019;
Chawla et al., 2017), as well as the unsupervised exploration of
patterns (Decroos, Van Haaren, & Davis, 2018; Knauf, 2014).

4
 The detection of pre-defined tactical patterns conducted by
teams in specific situations is a central issue in basketball and
American football analytics, nevertheless, it is not adequately ad-
dressed in football. In the following we define a tactical pattern
as a well-specified, repeatable and coordinated movement of a
team (or a group of members) conducted in a specific situation
of a match. One relevant aspect of tactical patterns in our con-
text is that they are uniquely identifiable by experts (see Section
2.3 for details). Another established concept in invasion sports is
game-states, splitting each match into offense and defense on the
highest level (Gréhaigne, Bouthier, & David, 1997). In football,
transitions to offense or defense as well as set-pieces are often
considered as separate game-states (Wei, Sha, Lucey, Morgan, &
Sridharan, 2013). An overview of the game-states in football is
shown in Figure 1. A team falls into the transition to offense
phase after winning the ball (and vice versa), whereas set-pieces
are separated since they start with a stoppage in the match.
In 2002, Borrie et al. stated that behavior in team sports con-
tains more patterns than the the human eye can observe. For a
single match, Laird and Waters (2008) showed that coaches have
a limited recall in reconstructing relevant scenes from the match.
Hence, the objective analysis of tactical patterns is of high inter-
est for practitioners and for sport science research (Gudmundsson
et al., 2017; Stein et al., 2017). Annotating tactical patterns by in-
specting video-footage is time-consuming (Perse, Kristan, Perš,
& Kovacic, 2006; Gudmundsson et al., 2017) and often subjective
(Perše, Kristan, Kovačič, Vučkovič, & Perš, 2009). As a conse-
quence, sample sizes that can be investigated (e.g., for trend anal-
ysis over multiple seasons) are limited due to time-constraints.
Consequently, the main objective of this thesis is to automate
this process using positional and event data, i.e. to answer the

5
following research question: (How) can tactical patterns be detected
automatically using machine learning algorithms based on positional
and event data?
To answer this question, the remainder of this thesis is struc-
tured as follows: Section 2 describes the basic preliminaries for
all investigations, starting with a detailed description of posi-
tional and event data (Section 2.1). The basics in machine learn-
ing methodologies for sports applications are provided in Sec-
tion 2.2. Finally, Section 2.3 describes related work on the detec-
tion of tactical patterns in order to derive the above presented
definition of a tactical pattern and to motivate the experimen-
tal studies presented in Section 3. In studies I & II (Sections
3.1 and 3.2) interesting tactical patterns are explored using un-
supervised techniques, while studies III–VI aim to detect tactical
patterns along the five established game-states shown in Figure
1 (offensive, defensive, transition to offense, transition to defense
and set-pieces). Study III (Section 3.3) focuses on transitions to
defense, i.e. counterpressing. The detection of team-tactical pat-
terns is extended to offensive (i.e., build-up and attacking play)
and defensive patterns (i.e., low-/mid- and high-block) in study IV
(Section 3.4). Tactical patterns during set-pieces (i.e., player- or
zonal-marking during corner kicks) are detected in study V (Sec-
tion 3.5). Study VI (Section 3.6) presents an outlook of offensive
transitions, i.e. counterattacks. All results are discussed in Section
4.
Besides the core contributions presented in this thesis (stud-
ies I–VI in Section 3), further publications have been achieved
within the scope of this research program: In Herold et al. (2021)
and Herold et al. (2019) the relevance of machine learning ap-
plications for football were explored from a practical perspec-
tive. In Anzer and Bauer (2021) and Anzer and Bauer (2022)

6
Figure 1: Overview of game-states in football.

expected goals (xG) and expected pass (xPass) models have been
built which are used as a fundamental component in all empir-
ical studies of this thesis. Additionally, two book contributions
provide a more general overview on machine learning applica-
tions in football (Anzer, Bauer, & Höner, 2021), as well as on
defensive organisation (Ric et al., 2021).

2 State of the Art

2.1 Positional and Event Data in Football

2.1.1 Event Data

Event data are logs of well-defined basic events describing a

football match, such as passes, shots, fouls, substitutions, corners
etc. Besides the time-stamp, the involved players (e.g., passer,
pass receiver), as well as several sub-attributes for each event
(e.g., high/low pass, completed/not completed), an estimated
position on the where on the pitch the event happened is usually
contained (Pappalardo et al., 2019; Stein et al., 2017). For a long
time, event data were collected explicitly to answer pre-defined
research questions (F. R. Goes, Meerhoff, et al., 2020). For exam-

7
ple, several approaches focused on annotating passes (Reep &
Benjamin, 1968; Camerino et al., 2012; Alcock, 2010; Borrie et al.,
2002), others annotated corners manually (Pulling, 2015; Casal,
Maneiro, Ardá, Losada, & Rial, 2015; Schmicker, 2013). Follow-
ing more of a bottom-up approach, event data are nowadays col-
lected systematically across matches and seasons (Pappalardo et
al., 2019) describing a match as an ordered sequence of events
(Stein et al., 2017). Companies like Stats Perform2 , Sportec Solu-
tions3 , and Statsbomb4 developed their own event data catalog
respectively, defining events and several attributes in great de-
tail.5 Basic event data catalogs are also described in Bialkowski
et al. (2016); Pappalardo et al. (2019) or Stein et al. (2017). Stein
et al. (2017) pointed out that the accordance of definitions dif-
fer per event. For example, events like throw-ins, kick-offs or
corner kicks are well-defined, whereas different interpretations
of tackles/duels, crosses or even successful passes cause huge
differences qua definition and/or subjective annotations.
A major flaw with primarily manually annotated event data
is inaccuracies due to human errors (Stein et al., 2017), especially
for the position on the pitch (typically drawn on a digital co-
ordinate system) and the timestamp (usually delayed due to the
human reaction time). Making use of automatically collected po-
sitional data to correct timestamp and location of event data is,
consequently, a relevant issue. The issue of synchronizing po-
sitional and event data is only rarely addressed as a limitation
(Spearman et al., 2017), current approaches typically use either
2 Statsperform LLC, Chicago, https://www.statsperform.com/.; former AMISCO, Pro-
zone and Opta.
3 Sportec Solutions AG, Munich, Germany https://www.sportec-solutions.de/index

.html.
4 Statsbomb Services Limited https://statsbomb.com/.
5 Note that event data are also collected on a systematic basis in other sports and of-

ten referred to as play-by-play data in other sports, e.g. basketball (Vračar, Štrumbelj, &
Kononenko, 2016).

8
positional or event data, or rely on exhaustive manual annota-
tions aligning event data with other data sources. Within the
scope of this research program, the issue of synchronizing event
data and positional data is addressed in Anzer and Bauer (2021)
for shots and in Anzer and Bauer (2022) for passes or more gen-
erally in Anzer (2021). All event data used in this dissertation are
spatio-temporally synchronized with the positional data, allow-
ing us to effectively use complementary information from both
data-sources.

2.1.2 Positional Data

Event data focuses on events with the ball, whereas Link and
Hoernig (2017) pointed out that a player possesses the ball on
average only for less than three minutes per match (depend-
ing on the position).6 Accordingly, literature has expressed a
haunting need to gather more information on off-the-ball activ-
ities (F. R. Goes, Meerhoff, et al., 2020; Vilar, Araújo, Davids, &
Travassos, 2012; Borrie et al., 2002). Positional data, often also re-
ferred to as tracking or movement data capture the exact positions
of players at any time within a match (typically with a frequency
of 10 or 25 Hz). Positions are transformed into a two-dimensional
Cartesian coordinate system determined by the pitch surround-
ings (Stein et al., 2017; Andrienko et al., 2017). For the ball,
positional data typically contain a third dimension—the height
relative to the pitch surface.
For the acquisition of positional data, one can differentiate be-
tween sensor-based solutions (i.e., global or local positioning sys-
tems) and optical tracking systems. Optical tracking systems ap-
6 Average possession times according to Link and Hoernig (2017): central forwards (0:49
± 0:43 min), central defenders (1:38 ± 1:09 min), central midfielders (1:27 ± 1:08 min) and,
surprisingly, the longest for goalkeepers (1:38 ± 0:58 min)

9
ply computer vision algorithms to video footage collected from
dedicated tracking cameras covering multiple perspectives of a
football match (Taberner et al., 2020; Stein et al., 2017). In the
past, GPS tracking data were predominantly used in sport sci-
ence literature, especially for physical performance analysis (see
Low et al. (2020) for an overview). For tactical analysis, various
shortcomings limited their usage: GPS data provide longitude
and latitude coordinates, which describes an object’s position
relative to the earth surface (McNeff, 2002). The transformation
to the pitch-centered coordinate system as well as the stadium
infrastructure (disturbing the connection between GPS receiver)
cause significant inaccuracies on player positions (Sathyamorthy,
Shafii, Amin, Jusoh, & Ali, 2015; Pons et al., 2019). Addition-
ally, various practical limitations makes it hard to consistently
use sensor-based data (e.g., missing data of opponent and ball,
players conceiving the devices as disturbance, single devices can
break during a match) (F. R. Goes, Meerhoff, et al., 2020; Low
et al., 2020; Buchheit et al., 2014; Hennessy & Jeffreys, 2018).
With recent developments in computer vision and video pro-
cessing, optical tracking systems turned out to be an alternative
with sufficient and increasing accuracy (Linke, Link, & Lames,
2020; Taberner et al., 2020; Linke, Link, & Lames, 2018). A wide
range of research has been conducted on optical tracking systems
(see Manafifard et al. (2017) for an overview). Even though, the
evaluation of positional data accuracy is an ill-posed problem
due to missing ground-truth information, experimental evalu-
ation studies are conducted in great detail (Linke et al., 2020;
Taberner et al., 2020; Linke et al., 2018; Cardinale, 2006), stating
that cutting-edge systems track player positions with an error of
less then 10 cm (Linke et al., 2020).

10
2.1.3 Dataset of Bundesliga and German National Team

For the remainder of this thesis, we make use of event data

collected by Sportec Solutions following the official event data
catalog for German Bundesliga.7 Positional data are generated
via different generations of the Chyronhego TRACAB system,8
which has been validated in Linke et al. (2020). The dataset, con-
sisting of several seasons of Bundesliga data as well as national
team matches is described in more detail in the respective pub-
lications, as well as in Anzer (2021). A detailed description of
the annotation of shots can be found in Anzer and Bauer (2021),
more details on passing events are pointed out in Anzer and
Bauer (2022).
The data used in this study are property of the Deutsche
Fußball-Bundesliga,9 as well as the Deutsche Fußball-Bund10 and
cannot be shared publicly. However, open-source positional data
(Pettersen et al., 2014)11 and event data (Pappalardo et al., 2019)12
can be used for reproduction. These open source data-sets pro-
vide the scientific community the option to reconstruct and eval-
uate the approaches presented in this thesis.

2.2 Machine Learning Basics for Sports Applications

Goodfellow, Bengio, and Courville (2016) defines machine learn-

ing as the ability of algorithms to learn from data. Although ma-
chine learning methods have been researched for decades (Samuel,
7 https://s.bundesliga.com/assets/doc/10000/2189_original.pdf
8 https://tracab.com/products/tracab-technologies/ or https://chyronhego.com/
wp-content/uploads/2019/01/TRACAB-PI-sheet.pdf
9 https://www.dfl.de/de/
10 https://www.dfb.de/index/
11 Non-scientific open-source positional data sets can be accessed from Skillcorner

(https://github.com/SkillCorner/opendata) or Metrica sports (https://github.com/

metrica-sports/sample-data).
12 Non-scientific open-source data sets can be accessed from Skillcorner (https://github

.com/SkillCorner/opendata), Metrica sports (https://github.com/metrica-sports/

sample-data) or Statsbomb (https://github.com/statsbomb/open-data).

11
1959), its recent success was only enabled through the avail-
ability of vast amounts of data and the affordability of com-
puting power needed to process such volumes of data (Jordan
& Mitchell, 2015). A major methodological distinction in ma-
chine learning is made between supervised and unsupervised
learning (Alloghani, Al-Jumeily, Mustafina, Hussain, & Aljaaf,
2020). While unsupervised machine learning explores patterns in
data with little or no human guidance (Gentleman & Carey,
2008), supervised machine learning uses human annotations to im-
itate experts performing tasks (Sing, Thakur, & Sharma, 2016).
The biggest category of unsupervised machine learning approach-
es solve clustering problems, while supervised machine learning
algorithms are typically applied to classification or regression
tasks (Goodfellow et al., 2016). Unsupervised algorithms are
thus purely data-driven, unbiased by expert assumptions, well-
suited to explore unknown data, and have the ability to disclose
unknown patterns in data (Goodfellow et al., 2016; Alloghani et
al., 2020). On the contrary, supervised algorithms require con-
siderable human effort (i.e., for data labeling) and a clearly de-
fined task. However, once a supervised machine learning algo-
rithm performs a given task with sufficient accuracy, they can be
used in production to assume repetitive tasks that human experts
would perform otherwise. To reduce the human effort required
for labeling in supervised machine learning, semi-supervised ap-
proaches emerged as a third category (Goldberg, 2009). The basic
idea is that human expertise is still used for labeling, but sup-
ported by more automated approaches to reduce the amount of
labels required. For example, weak supervision (applied in study
V, Section 3.5) uses rule-based approaches to create weakly la-
beled data, which are used in combination with expert-labeled
data to train supervised machine learning algorithms (Ratner et

12
al., 2017). Another method to increase the amount of labeled
data without human effort is data augmentation (Dyk & Meng,
2001), used also in studies V & VI (Sections 3.5 & 3.6). Here,
labeled data are slightly modified to create artificial data.
On the highest level, supervised machine learning algorithms
aim to learn a function f taking input data x 2 X to predict
pre-defined labels y 2 Y: f ( x ) ⇡ y. Hereby, x can be seen as
a reproduction of the real-world phenomena (e.g., a sequence
of a football match), typically transformed into a lower dimen-
sional space in order to reduce the complexity (e.g., through
expert-driven feature extraction from raw data). y describes the
label (target variable) in a pre-defined label space Y. For re-
gression tasks y is a continuous variable, for classification task y
is discrete. Compared to traditional statistics, inductive reason-
ing (i.e., generalisation) is performed via using three different
sub-sets of data: training data Xtrain , Ytrain , test data Xtest , Ytest ,
and new data Xnew . The training and testing labels (Ytrain and
Ytest ) are available through expert-annotation, whereas the target
variables for Xnew are unknown. Mathematically, f is trained as
an optimization task learning to chose the best free parameters
p 2 P in f p to minimize the prediction error on the training data:
min p  xi ,yi 2Xtrain ,Ytrain | f p ( xi ) yi |. With sufficient flexibility in f
(i.e., a high dimensional parameter space P), this optimization
task can perform well by overfitting f on the training data. Thus,
generalisation in supervised machine learning is evaluated on
test data Xtest , Ytest , that are not used in the training process. The
error on the test data is described by  xi ,yi 2Xtest ,Ytest | f p ( xi ) yi | (p
optimized via model training) and typically used to evaluate the
accuracy of the model.
For f , a plethora of algorithm families is established (Mohri,
Rostamizadeh, & Talwalkar, 2014; Goodfellow et al., 2016). Re-

13
lying typically on a feature extraction, extreme gradient boosting
(XGBoost) classifiers, from the family of tree-based algorithms, achieve
good results in various domains (T. Chen & Guestrin, 2016). XG-
Boost is used in Anzer and Bauer (2022, 2021), as well as in
study III of this thesis (Bauer & Anzer, 2021). More details on
the methodology can also be found in Anzer (2021). A closely
related methodology, random forests, has also been applied in
sports analytics (Karsten et al., 2017).
Besides tree-based classifiers, support vector machines are pre-
dominantly used to perform (binary) classification tasks (Pisner
& Schnyer, 2019), and have also been applied to sports analytics
(B. W. Hosp, Schultz, Honer, & Kasneci, 2021; Atmosukarto et al.,
2013). Support vector machines are used to detect counterattacks
in study VI of this dissertation.
Artificial neural networks (Shinde & Shah, 2018)—another type
of machine learning algorithm that can be used for various types
of tasks—achieved groundbreaking results in image (and video)
classification tasks (Krizhevsky, Sutskever, & Hinton, 2012), nat-
ural language processing (Otter, Medina, & Kalita, 2021), au-
tonomous driving (Shinde & Shah, 2018), playing video (Mnih
et al., 2013) or board games (Ghory, 2004), and many more (e.g.,
Najafabadi et al. (2015)). Although deep architectures of neu-
ral networks (typically referred to as deep learning) require vast
amounts of (labeled) data and sizeable computing power, their
largest benefit is that they can often be fed with raw data, i.e.
without the need for exhaustive feature engineering. In this dis-
sertation, convolutional neural networks, a special family of neural
networks optimized to handle images (see Zhang et al. (2019) for
an overview), are used in studies IV & V (Sections 3.4 & 3.5).
In study V we also apply the concept of long short-term mem-
ory networks introduced by Hochreiter and Schmidhuber (1996)

14
to efficiently handle temporal data. A combination of convolu-
tional layers with long-short term memory units has been used
in football to perform classification tasks on eye-tracking data
(B. Hosp, Schultz, Kasneci, & Höner, 2021). Grunz, Memmert,
and Perl (2009, 2012) applied self organising maps, a modifica-
tion of neural networks that are able to approach unsupervised
problems, to different sports.
Further machine learning algorithms have been successfully
applied to detect tactical patterns in invasion sports. Logistic re-
gressions (Chawla et al., 2017; Mcintyre, Brooks, Guttag, Wiens,
& Arbor, 2016), Bayesian classifiers, multi-model densities (Li,
Chellappa, & Zhou, 2009), pattern templates (Intille & Bobick,
1999; Siddiquie, Yacoob, & Davis, 2009; Li & Chellappa, 2010;
Perše et al., 2009, 2009), bag-of-word algorithms (Montoliu, Martín-
Félez, Torres-Sospedra, & Martínez-Usó, 2015) and k-nearest neigh-
bours (Bialkowski et al., 2015) were used to solve supervised
problems.
T-pattern analysis (Borrie et al., 2002; Hernández-mendo, 2006),
multi-scale matching (Hirano & Tsumoto, 2004), multiple modifi-
cations of hierarchical clustering (Bialkowski et al., 2014), convo-
lutional kernels (Knauf, 2014) and the Delauany method (Narizuka
& Yamazaki, 2019) addressed unsupervised tasks.
Hierarchical clustering, the most established unsupervised tech-
nique, is used in studies II & IV.
A general introduction on supervised and unsupervised ma-
chine learning applications in football is presented in Anzer,
Bauer, and Höner (2021). Fujii (2021) and Araújo et al. (2021)
outline an overview of machine learning methodologies applied
to sports. In Herold et al. (2019) we give an overview of ma-
chine learning applications applied to quantify offensive play in
football. Technical details on the above presented algorithms can

15
be found in Goodfellow et al. (2016) and (Han, Kamber, & Pei,
2012).
In order to train a machine learning algorithm (either su-
pervised or unsupervised) the underlying data has to be mod-
eled into an appropriate structure, which poses technically de-
manding challenges. Positional data are a high-dimensional,
spatio-temporal abstraction of a football match (Stein et al., 2017).
Typically, the dimensionality of the input data X has to be re-
duced (Lucey, Bialkowski, et al., 2013) to simplify the problem.
A traditional approach is to use subject-matter expertise to ex-
tract features and perform the prediction task from a lower-
dimensional feature space (Duarte et al., 2013). Second, the
movement of players for the same tactical pattern can vary dra-
matically for different scenarios (Perse et al., 2006; Li & Chel-
lappa, 2010; Stracuzzi et al., 2011), which makes the definition
of similarity metrics challenging. A third major problem when
dealing with positional data is to project player positions in a
permutation-invariant space (Wei et al., 2013). The ordering
problem in invasion sports is addressed by describing positional
data as images (Dick & Brefeld, 2019; S. Zheng, Yue, & Lucey,
2016; K.-C. Wang & Zemel, 2016), artificial orderings using heuris-
tics (Le, Yue, Carr, & Lucey, 2017; Fujii, 2021) or, again, ex-
tracting features from the raw data. More recently, different
approaches show that graph neural networks can serve as an
efficient procedure to solve the permutation-invariance problem
(Dick, Tavakol, & Brefeld, 2021; Sun, Karlsson, Wu, Tenenbaum,
& Murphy, 2019; Yeh, Schwing, Huang, & Murphy, 2019).

16
2.3 Data-Driven Detection of Tactical Patterns in Sport
2.3.1 Tactical Patterns in Invasion Sports

Gréhaigne, Godbout, and Bouthier (1999) separated two of-

ten confounded terms: strategy—the plan of a team made be-
fore a match—and tactics—decisions that are conducted during a
match as a response to the dynamic environment. In this context,
they presented a sub-category of tactics, called schemas of play,
describing organized, collective and repeated patterns. How-
ever, Rein and Memmert (2016) claimed that a clear distinction
between tactics and strategy is challenging, since any real-time
interaction will be prone by the a priori strategy. Even though
the concept was originally designed for football, tactics and strat-
egy is often used synonymously for invasion sports in general.
A team’s collective movement in invasion sports is clearly co-
ordinated and contains patterns and aspects of synchronicity
(see Sarmento et al. (2018) or Courel-Ibáñez, McRobert, Toro,
and Vélez (2017) for reviews). Rather than following one fixed
team or game tactic during a whole match, the current game-
state (i.e., offensive, defensive, transition, set-piece) and further
sub-categories (e.g., counterattack or counterpressing for transi-
tions in football) significantly influence team behavior in inva-
sion sports (Alexander, Spencer, Sweeting, Mara, & Robertson,
2019; Wei et al., 2013). Thus, team behavioral patterns have to
be investigated in finer granularity, especially along game-states
and the respective sub-categories.
For tactical patterns, literature lacks an established definition,
rather a plethora of namings have been used for one and the
same concept: schemas of play (Gréhaigne et al., 1999), coop-
erative plays (Hojo, Fujii, Inaba, Motoyasu, & Kawahara, 2018),
game-phases (Perse et al., 2006; Lucey et al., 2014), events (Pfeiffer

17
& Perl, 2015), group motion patterns (Li & Chellappa, 2010),
movement patterns (Stein et al., 2017; Gudmundsson & Horton,
2017), or tactical patterns (Q. Wang, Zhu, Hu, Shen, & Yao, 2015;
Kempe, Grunz, & Memmert, 2015; Knauf, 2014; Grunz et al.,
2012, 2009). While most authors used the respective term, with-
out referring to an explicit definition, Q. Wang et al. (2015) de-
fined tactical patterns as a series of frequently used ball-passing
combinations, stating that this definition fits their purpose of an-
alyzing passing patterns best. In football, Lucey et al. (2014)
used the term game-phase for the same phenomena (less focused
on passing behavior)—patterns conducted by the whole team in
specific situations of a match—and analyzed seven exemplary
categories (corners, free-kicks, penalties, set-pieces, open play,
counterattacks). To be a good candidate for such a pattern that
can be detected using a supervised machine learning approach,
sundry authors highlight that the patterns must be well-defined
and widely used (Hojo et al., 2018; Bialkowski et al., 2015; Mon-
toliu et al., 2015; K.-C. Wang & Zemel, 2016).
In the following, we define a tactical pattern as a repeatable,
coordinated movement of a team (or a group of members) con-
ducted in specific situations, which can be uniquely identified
by experts (Kempe et al., 2015; Q. Wang et al., 2015; Grunz et al.,
2012).

2.3.2 The Detection of Tactical Patterns in Invasion Sports

The relevance of automatically detecting such tactical patterns

in team-sports is highlighted in Desporto (2009) or Gudmundsson
et al. (2017). All known studies following the goal of automat-
ically detecting tactical patterns in invasion sports using posi-
tional data, event data and/or video footage are listed in Table

18
1.13 In the following we pose the related work on tactical pattern
detection in invasion sports along Table 1.
In 1999, Intille and Bobick introduced a pioneering approach
to detect a pre-defined offensive play in American football using
(manually drawn) movement trajectories of 29 attacking plays.
The task of recognizing attacking plays remains a predominant
problem in American football analytics (Stracuzzi et al., 2011;
Li & Chellappa, 2010; Siddiquie et al., 2009; Li et al., 2009).
High-level game-phases (offense, defense, kickoff, punt, field-
goal plays) were detected directly from video-footage by S. Chen
et al. (2014). Atmosukarto et al. (2013). Hochstedler and Gagnon
(2017) detected the static formations (i.e., player roles) at the start
of each play.
Many studies have also been conducted in basketball (see
Stephanos et al. (2021) or Courel-Ibáñez et al. (2017) for reviews),
focusing on the primary objective of automatically detecting lower
level tactics performed by groups of players, i.e. screening14
(Hojo et al., 2018; Mcqueen, Wiens, & Guttag, 2014; Perše et al.,
2009; Perse et al., 2006), defensive counter-strategies on screening
(Tian, De Silva, Caine, & Swanson, 2020; Mcintyre et al., 2016) or
cutting15 (Perse et al., 2006). Those group-tactical behaviors were
often gathered as a sequence of actions to analyse whole attack-
ing plays (Kempe et al., 2015; J. Wang & Zhang, 2015; H. T. Chen,
13 Studies analyzing rather technical than tactical patterns (Schmidt, 2012) or studies an-
alyzing physical patterns for the purpose of injury prediction (Kelly, Coughlan, Green, &
Caulfield, 2012; Cai et al., 2018) were excluded, since those do not suit our definition of tac-
tical patterns. We also exclude studies focusing purely on basic event detection from video
footage (Pouyanfar & Chen, 2016; Kolekar, Palaniappan, Sengupta, & Seetharaman, 2009;
Ekin, Tekalp, & Mehrotra, 2003) or from positional data (Stein et al., 2019; Richly, Bothe,
Rohloff, & Schwarz, 2016; Motoi et al., 2012; Gudmundsson & Wolle, 2010; M. Zheng &
Kudenko, 2010).
14 A group-tactical pattern, where an offensive player (not in possession of the ball) legally

blocks a defender in order to pretend him/her from defending himself/herself or his/her

team-mate in possession of the ball.
15 Cutting describes a sudden off-ball movement of a player in order to get rid of his de-

fender.

19
Chou, Fu, Lee, & Lin, 2012; Perše et al., 2009; Perse et al., 2006)
or to identify players or teams just by recognizing their playing
styles (Mehrasa, Zhong, Tung, Bornn, & Mori, 2018).
From various approaches investigating screen plays in bas-
ketball, Hojo et al. (2018) presented an extended study improv-
ing the state-of-the-art results. First, they properly define differ-
ent types of screen plays. Based on this definition, data were
expert-labeled in a controlled experiment with staged screens,
and later in real game situations. In both experiments positional
data were recorded. For the binary classification (screen play
or not) their thorough experimental set-up exceeds prior results
with an area under the curve of 0.941 for on-ball and 0.855 for
off-ball screens. Although different labelers were involved, no
proper inter-labeler reliability study was conducted.
The data-driven detection of tactical patterns has also been in-
vestigated in ice-hockey (Mehrasa et al., 2018), Australian rules
football (Rennie et al., 2020; Alexander et al., 2019), team hand-
ball (Pfeiffer & Perl, 2015) as well as in rugby (Bunker et al., 2020;
Karsten et al., 2017).16
In Basketball (Hojo et al., 2018; K.-C. Wang & Zemel, 2016;
Mcintyre et al., 2016; Mcqueen et al., 2014) and in American foot-
ball (Hochstedler & Gagnon, 2017; Atmosukarto et al., 2013; Li
et al., 2009; Siddiquie et al., 2009), the application of supervised
machine learning methods using manual expert-annotations is
established (see also Table 1). The relevance of clearly recogniz-
able patterns is highlighted in various studies, nevertheless, the
inter-labeler reliability is only analyzed as an indicator for agree-
ment in football (Chawla et al., 2017) and rugby (Bunker et al.,
2020). Unsupervised approaches often aim for the creation of
16 Again, this dissertation focuses on pattern detection for invasion sports. Nevertheless,
several studies analyzed patterns in other sports, for example volleyball in Van Haaren,
Shitrit, Davis, and Fua (2016).

20
new, unknown insights, whereas process automation for coaches
and match analysts is mentioned as the primary aim of super-
vised approaches.

2.3.3 The Detection of Tactical Patterns in Football

The first known approach analyzing tactical patterns stems

from football: As early as in 1968, Reep and Benjamin (1968)
manually annotated the passing sequences across 54 matches in
the English Premier league as well as 47 matches from other
competitions. Building on that pioneering work, a majority of
approaches presented in football focus on the exploratory detec-
tion of attacking patterns (Decroos et al., 2018; Hobbs, Power,
Sha, Ruiz, & Lucey, 2018; Van Haaren, Dzyuba, Hannosset, &
Davis, 2015; Montoliu et al., 2015; Fernando, Wei, Fookes, Srid-
haran, & Lucey, 2015; Niu, Gao, & Tian, 2012; Borrie et al., 2002),
build-up patterns (Knauf, 2014; Grunz et al., 2012) and patterns
in pass sequences (Chawla et al., 2017; Brooks, Kerr, & Guttag,
2016; J. Wang & Zhang, 2015; Hernández-mendo, 2006; Hirano
& Tsumoto, 2004). Defensive corner kick strategies have been in-
vestigated in Shaw and Gopaladesikan (2021) and Power, Hobbs,
Ruiz, Wei, and Lucey (2018). More frequently addressed (com-
pared to basketball and American football) are studies on team
formations (Narizuka & Yamazaki, 2019; Müller-Budack et al.,
2019; Shaw & Glickman, 2019; Bialkowski et al., 2014, 2015, 2016;
Wei et al., 2013). Note that many studies aim to detect one team
formation aggregated over a whole match, which does not fit
our definition of tactical patterns (see Bauer, Anzer, and Shaw
(2022) for more details). However, due to the complexity of the
game, the analysis of formation patterns in football is a unique
characteristic compared to American football (where only static
formations at the beginning of each play are of interest) or to

21
Basketball (where fewer variations on formations exist).
In football, only a few studies built their work on expert-
labeled data of pre-defined patterns (frequently used in basket-
ball and American football) Montoliu et al. (2015) hand-labeled
five types of attacking plays, Shaw and Gopaladesikan (2021) as
well as Power et al. (2018) detected defensive tactics of defending
corner kicks. In Chawla et al. (2017), two human experts anno-
tated the reward of in total 2, 932 passes on a six-point Lickert
scale (very good, good, marginally good, marginally bad, bad,
very bad). The alignment among the labelers was monitored us-
ing Cohen’s kappa—after labeling two matches, disagreements
were consolidated. Finally, different classifiers were trained to
detect three types of pass ratings (good, ok, bad)—multinominal
logistic regression performed best on various metrics with an F1 -
score of 0.748.
This thesis uses the idea of supervised machine learning to
detect identifiable tactical patterns as well as established phases
of play. By addressing team-tactical patterns in all game-states
(offensive, defensive, transition and set-pieces) and by integrat-
ing football experts closely in our experiments, we guarantee a
practical applicability of the results. As in Chawla et al. (2017),
we make use of both positional and event data and emphasize
the importance of a thorough labeling process.

22
 Table 1: Overview of tactical pattern detection in invasion sports.

Study Sport Data Methods Pattern

Positional data
American (manually Supervised ML (bayesian classifiers, Attacking play (one particular group
Intille and Bobick (1999)
Football drawn pattern template) tactical pattern called p51curl-play)
trajectories)
Attacking plays (combo dropback,
American Supervised (expert-labeled,
Li et al. (2009) Video footage HITCH dropback,
Football multi-model densities)
middle/wideleft/wideright run)
Attacking plays
American Supervised (expert-labeled, feature
Siddiquie et al. (2009) Video footage (left-/middle-/right-runs,
Football extraction)
option-/short-/rollout-/deep-pass)

23
American
Li and Chellappa (2010) Positional data Supervised Attacking plays
Football
American (1) Individual running patterns; (2)
Stracuzzi et al. (2011) Video footage Supervised (pattern template)
Football Attacking plays
American Supervsised (expert-labeled, support
Atmosukarto et al. (2013) Video footage Team formations
Football vector machines)
American Game-states (offense, defense, kickoff,
S. Chen et al. (2014) Video footage Supervised (rule-based)
Football non-punt, field goal plays)
Hochstedler and Gagnon American (1) Supervised; (2) Supervised (1) Team formations; (2) Offensive
Positional data
(2017) Football (expert-labeled, feature extraction) routes
 Table 1: Overview of tactical pattern detection in invasion sports.

Study Sport Data Methods Pattern

(1) Game-states (offensive, defensive,
Perse et al. (2006) Basketball Positional data Supervised (pattern template) time-out); (2) Attacking pattern
(screen, move, player formation)
(1) Game-state
Video footage, (offensive/defensive/time-out); (2)
Perše et al. (2009) Basketball Supervised (pattern template)
positional data Attacking pattern (screen plays, moves,
starting formation); (3) Attacking plays
Basketball,
Grunz et al. (2009) Positional data Unsupervised (self-organising maps) Attacking plays
Football
Positional and Supervised (expert-labeled, feature

24
Mcqueen et al. (2014) Basketball Screen plays
event data extraction)
(1) Detection of plays (fastbreak, horns,
high-pick); (2) Tactical behaviours (3x
Kempe et al. (2015) Basketball Positional data (1) Supervised; (2) Unsupervised
defensive, 6x offensive, 1x transition, 2x
set-piece)
K.-C. Wang and Zemel Supervised (expert-labeled, neural Attacking plays (11 offensive plays with
Basketball Positional data
(2016) network) all players involved)
Supervised (expert-labeled, logistic
Mcintyre et al. (2016) Basketball Positional data Defensive counter on screen plays
regression)
 Table 1: Overview of tactical pattern detection in invasion sports.

Study Sport Data Methods Pattern

Basketball, (1) Hockey-events (pass, dump in/out,
Mehrasa et al. (2018) Ice- Positional data Supervised (neural networks) shot, carry, puck protection); (2)
Hockey Basketball team classification
Subtypes of screen plays (off-ball:
Supervised (expert-labeled, feature
Hojo et al. (2018) Basketball Positional data down, flare, pin, black, flex, cross;
extraction)
on-ball: pick and roll, H and off )
Defensive counter on screen plays
Tian et al. (2020) Basketball Positional data Supervised
(switch and trap)
Annotated pass
Reep and Benjamin (1968) Football Descriptive statistics Passing patterns
sequences

25
Event data
Borrie et al. (2002) Football (purpose Unsupervised (T-pattern analysis) Attacking plays
specific)
Hirano and Tsumoto Unsupervised (multi-scale matching,
Football Event data Passing patterns
(2004) clustering)
Hernández-mendo (2006) Football Event data Unsupervised (T-Pattern analysis) Passing patterns
(1) Team identification using passing
(1), (2) Supervised (feature patterns; (2) Prediction whether
Brooks et al. (2016) Football Event data
extraction); (3) Descriptive statistics passing sequence ends in a shot; (3)
Passing patterns
 Table 1: Overview of tactical pattern detection in invasion sports.

Study Sport Data Methods Pattern

Unsupervised (self-organising maps;
Build-up patterns (long game initiation
Grunz et al. (2012) Football Positional data extended labeling conducted for
versus short game opening)
evaluation)
Attacking patterns (6 pre-defined
Niu et al. (2012) Football Video footage Supervised (rule-based) patterns; ground versus air attack from
different starting points)
(1) Game-phase (in-play, stoppages,
(1) Supervised; (2) Unsupervised
Wei et al. (2013) Football Positional data highlights, different set-pieces); (2) Team
(hierarchical clustering)
formations per game-phase
Vilar, Araújo, Davids, and

26
Football Positional data Descriptive statistics Teams covered areas
Bar-Yam (2013)
Unsupervised (agglomerative
Bialkowski et al. (2014) Football Positional data Team formations
clustering)
Knauf (2014) Football Positional data Unsupervised (convolution kernels) Build-up types
Attacking plays (ball possession; quick
Supervised (expert-labeled, attacks, i.e., switching the attack and fast
Montoliu et al. (2015) Football Video footage
bag-of-words) break; set-pieces, i.e., direct/indirect
freekick, penalty, corner kick)
Q. Wang et al. (2015) Football Event data Unsupervised Passing patterns
Fernando et al. (2015) Football Positional data Unsupervised Attacking plays
Van Haaren et al. (2015) Football Event data Unsupervised Attacking plays
 Table 1: Overview of tactical pattern detection in invasion sports.

Study Sport Data Methods Pattern

(1) Supervised (k-nearest neighbours (1) Team formations; (2) Team
Bialkowski et al. (2015) Football Positional data
regression); (2) Supervised identities
Feuerhake (2016) Football Positional data Unsupervised Movement patterns
Bialkowski et al. (2016) Football Positional data Unsupervised Team formations
Positional and Supervised (expert-labeled, feature Passing patterns (i.e., reward good, ok,
Chawla et al. (2017) Football
event data extraction, inter-labeler accordance) bad)
Hobbs et al. (2018) Football Positional data Hybrid Counterattacking patterns
Supervised (expert-labeled, neural
Power et al. (2018) Football Positional data Defensive corner roles (team-level)
networks)
Decroos et al. (2018) Football Event data Unsupervised Attacking plays

27
Andrienko et al. (2019) (1) Team formations; (2)
Football Positional data Visual analytics
(study I) Counterpressing
Müller-Budack et al.
Football Positional data Supervised Team formations
(2019)
Narizuka and Yamazaki Unsupervised (Delaunay method,
Football Positional data Team formations
(2019) hierarchical clustering)
Shaw and Glickman
Football Positional data Unsupervised Team formations (per game-state)
(2019)
Shaw and Gopaladesikan Supervised (expert-labeled, feature
Football Positional data Defensive corner roles (player-level)
(2021)17 extraction, XGBoost)

17 Note that a slightly different version of that that paper can be found in Haaren et al. (2013).
 Table 1: Overview of tactical pattern detection in invasion sports.

Study Sport Data Methods Pattern

Anzer, Bauer, and Brefeld Positional and Unsupervised (hierarchical
Football Goal scoring patterns
(2021) (study II) event data clustering)
Bauer and Anzer (2021) Positional and Supervised (expert-labeled, feature
Football Counterpressing
(study III) event data extraction, XGboost)
(1) Supervised (expert-labeled,
Bauer, Anzer, and Shaw Positional and convolutional neural networks); (2) (1) Phases of play; (2) Team
Football
(2022) (study IV) event data Unsupervised (hierarchical formations
clustering)
Supervised (expert-labeled,
Bauer, Anzer, and Smith Positional and Defensive corner roles (player-level
Football convolutional neural networks,

28
(2022) (study V) event data and player-marking assignment)
long-short term memory networks)
Fassmeyer, Anzer, Bauer, Semi-supervised (expert-labeled,
Positional and (1) Events (corner kicks, crosses); (2)
and Brefeld (2021) (study Football variational autoencoder, support
event data Counterattacks
VI) vector machines)
Vilar et al. (2012) Futsal Video footage Descriptive statistics Attacker-defender dyads
Lucey, Bialkowski, et al. Field-
Positional data Supervised (formation templates) Team formations
(2013) Hockey
Rugby Positional data Supervised (expert-labeled, random
Karsten et al. (2017) Scrum events
union (GPS) forest)
Supervised (rule-based,
Bunker et al. (2020) Rugby Event data Attacking plays
inter-observer consistency)
 Table 1: Overview of tactical pattern detection in invasion sports.

Study Sport Data Methods Pattern

Event data
Team Supervised (expert-labeled, neural
Pfeiffer and Perl (2015) (purpose- Attacking plays
handball networks)
specific)

29
2.3.4 Phases of Play in Football

The concept of tactical patterns is used in different invasion

sports. On the basis of Gréhaigne et al. (1997), we embed fre-
quently occurring tactical patterns in football in a game-model.
Figure 2 shows an overview of phases of play, further defined as
tactical patterns that are established as sub-categories of game-
states (i.e. offensive, defensive, transition to offense, transition
to defense and set-pieces) in professional football.18 Although
minor differences occur due to different playing philosophies,
the taxonomy in Figure 2 is consolidated in several discussions
with professional coaches and match analysts from German Bun-
desliga teams and the German national teams (see Acknowledge-
ments). Figure 2 differentiates between common game-states (of-
fensive, defensive and transition) on the highest level. Set-pieces,
uncontrolled possessions and situations where the ball is out of
play are listed as a separate game-state respectively due to their
specific characteristics (Wei et al., 2013). For transitions we fur-
ther distinguish between those to offense, and those to defense.
Inter alia, we embed well known strategies like counterattacks
(Hughes & Lovell, 2019; Tenga, Holme, Ronglan, & Bahr, 2010)
or counterpressing (Low et al., 2020; Hobbs et al., 2018) as op-
tional tactics during transitions. In offense, build-up—advancing
possession from the own goal behind the first defending line of
the opponent—is separated from attacking play, in which a team
has possession in the last third of the pitch and purely follows
the objective of scoring a goal (Rein & Memmert, 2016; Plummer,
2013). The defensive game-state is separated based on the height
and activity of a team defending their own goal. Low-block is
a very passive strategy, where a team focuses on the protection
18 Similar frameworks for phases of play has been presented for American football
(Siddiquie et al., 2009) and for football (Wei et al., 2013).

30
of the own goal. On the other hand, teams can block the op-
ponents possession (i.e., the build-up) actively and as close to
the opponents goal as possible, typically declared as high-block
(Power et al., 2017). Midfield-block or mid-block describes a mod-
erate defending tactic. The eleven phases of play presented as
sub-categories of six game-states in Figure 2, again, can contain
further (highly individualized) tactical patterns, which are either
performed by groups of players (i.e., one-twos, overlapping runs
or other passing patterns) or present specific manifestations, for
example different types of counterattacks presented in Hobbs et
al. (2018).
Following the primary objective—the detection of tactical pat-
terns—the contribution of this thesis is outlined with reference to
Figure 2. First, two exploratory studies I & II are presented ana-
lyzing purely data-driven patterns (Sections 3.1 and 3.2), helping
us to derive the taxonomy presented in Figure 2. Study I (Section
3.1) presents an unsupervised approach using visual analytics to
explore differences in team behavior per game-state. Another ex-
ploratory overview on analyzing goal scoring patterns (includ-
ing counterpressing, counterattacks, high-block and mid-block
defending) is presented in study II (Section 3.2). Transition to de-
fense, particularly the detection of counterpressing is addressed
in study III (Section 3.3). Counterattacks, a sub-category of tran-
sition to offense are addressed in study VI (Section 3.6). Study
IV analyses team formations by first detecting the five phases of
offensive and defensive play (Section 3.4). Finally, in study V
(Section 3.5), defending patterns in set-pieces, i.e., corner kicks
are detected.

31
32
Figure 2: Overview of phases of play in football. The contribution of this dissertation is highlighted using different colours, in relation to the game-states
(offensive, defensive, transitions to offense/to defense and set-pieces).
3 Empirical Studies

In the following sections we summarize the empirical stud-

ies. Instead of presenting detailed explanations, the respective
summary focuses on the contribution of each study towards the
above defined research question—details can be found in the re-
spective full text attached in the Appendices.
In Sections 3.1 and 3.2 the importance of investigative, un-
supervised approaches to determine further research questions
is outlined. Building on these results, Sections 3.3, 3.4 and 3.5
present supervised machine learning approaches using close co-
operation with football experts to automatically detect tactical
patterns. Finally, Section 3.6 provides an outlook towards the
potential of semi-supervised learning that can be used effectively
for the purpose of tactical pattern detection.
The contribution to the empirical studies I, II & VI in Sections
3.1, 3.2 and 3.6 is a co-authorship, whereas the articles presented
in studies III, IV & V (Sections 3.3, 3.4 and 3.5) are conducted as
first-author.

3.1 Study I: Constructing Spaces and Times for Tactical Anal-

ysis in Football (Andrienko et. al. 2019)

Complex positional data can only be processed by experts us-

ing aggregations and high-level visualizations depicting interest-
ing patterns. Visual analytics has been a relevant issue in football
analytics (Andrienko et al., 2017; Sacha et al., 2017; Perin, Vuille-
mot, & Fekete, 2013; Wu et al., 2019). Study I (Andrienko et al.,
2019) is the result of a common research project of visual ana-
lytics experts, data-scientists and football experts with the goal
to explore tactical patterns using different methodologies. This
work can thus be seen as a general introduction to the research

33
project, using simple aggregation and query techniques with a
strong emphasis on visualizations in order to explore tactical pat-
terns and team tactics in the data.
Using a substantial portion of domain expertise, data have
been (1) queried and filtered to select time intervals of inter-
est, (2) aggregated to get an overview over different time inter-
vals, and (3) visualized in a way that patterns can be explored
from illustrations. First, we analyzed team formations. Tra-
ditional approaches study the formation of a team, specifically
each player’s role, related to the center of the pitch. By intro-
ducing and visualizing the team-space—player positions related
to their teams’ average position—we help practitioners to cap-
ture the interaction of teammates in greater detail. Second, by
presenting various queries we show the dynamics of team for-
mations and player rules. For example, we visualize differences
in formations per game-state (offensive, defensive, transition),
per score (e.g., teams falling back after goals), and how play-
ers interpret the same role differently. This exploratory analysis
guided the research investigated in study IV (Bauer, Anzer, &
Shaw, 2022) in Section 3.3. Lastly, transition scenarios have been
analyzed and visualized. By proving the practitioner assump-
tion, that the transition to defense strategy varies significantly
per team, we motivate our work on the detection of counter-
pressing (Bauer & Anzer, 2021) presented in study III (Section
3.3).
The major limitation of this approach is that little evidence
or usable insights can be derived by visualizations and/or the
small amount of data used in the experiment (essentially one
match of positional and event data). However, the exploratory
approach and the collaboration of people with different expertise
and perspectives turned out to be very beneficial for our research

34
project. Even though the applied methods fall slightly out of
scope compared with the rest of the thesis, various results of this
project served as a motivation for the rest of the projects.

3.2 Study II: The Origins of Goals in the German Bundesliga

(Anzer, Bauer, & Brefeld 2021)

In professional football, match analysis departments catego-

rize goals scored and received (e.g., open play versus set-pieces
on the highest level) for their own matches (several times in a
season) and for their upcoming opponent (on a weekly basis).
Approaches differ drastically among experts due to a low accor-
dance on well-defined and unique categories of goals scored. In
addition to the time saved using automated categorization, the
problem of goal origins has been studied for decades in the sci-
entific domain (Reep & Benjamin, 1968; González-Ródenas et al.,
2019; Njororai, 2013; Mitrotasios & Armatas, 2012).
In Anzer, Bauer, and Brefeld (2021) we follow an unsupervised
approach to explore objective goal categories using a sample size
that would exceed human capacities. Based on synchronized po-
sitional and event data consisting of 3, 457 goals from two sea-
sons of German Bundesliga and 2nd Bundesliga (2018/20219 and
2019/2020), we devise a rich set of 37 features that can be ex-
tracted automatically, and propose an agglomerative hierarchical
clustering approach to identify group structures. The features
describe the attack leading to a goal and contain, for example,
the duration of the attack, the location of the assist and the shot.
Feature extraction, choosing the number of clusters and contex-
tualizing the clusters based on video footage has been conducted
in close collaboration with football experts. The results consist of
50 interpretable clusters revealing insights into scoring patterns.

35
The clustering found eight highly separated clusters (penalties,
direct freekicks, kick and rush, one-two’s, assisted by header, as-
sisted by throw-in) and nine categories (e.g., corners) combining
more granular patterns (e.g., five subcategories of corner-goals).
Again, the insights motivated further work: One cluster is com-
prised of 124 goals later contextualized as goals after counter-
pressing by the experts (motivating Section 3.3). The clustering
also revealed 93 goals after high-block pressing, as well as 73
goals after mid-block pressing (motivating Section 3.4).
By automating the analysis of goal scoring patterns using
data, with 3, 457 goals, we exceed the sample sizes used in tra-
ditional studies (Reep & Benjamin, 1968; González-Ródenas et
al., 2019; Njororai, 2013; Mitrotasios & Armatas, 2012). Conse-
quently, we are able to reveal patterns that could not have been
detected as such using less goals. The major limitation, typi-
cal for unsupervised machine learning approaches, is that they
can not be used seamlessly in practice. The meaningful results
can rather be used to consolidate and define clear categories of
goals to build a supervised machine learning approach in future
projects.

3.3 Study III: Data-Driven Detection of Counterpressing in

Professional Football (Bauer & Anzer 2021)

After losing the ball, a team conducts counterpressing if at

least one player exerts (spatio and/or temporal) pressure on the
ball carrier, or on the opponents close to the ball. The relatively
young strategy of counterpressing is admittedly touched upon
but never properly investigated in literature (Anzer, Bauer, &
Brefeld, 2021; Low et al., 2020; Andrienko et al., 2019; Hobbs
et al., 2018). Analyzing counterpressing is an important task

36
for any professional match analyst in football, but is being done
exclusively manually by observing video footage.
In study III (Bauer & Anzer, 2021) we present the first ap-
proach in football using supervised machine learning to detect a
complex team-tactical pattern in open-play (see Table 1).19 The
primary purpose of this paper is to automatically identify this
strategy using positional and event data. Together, with pro-
fessional match analysis experts we discussed and consolidated
a consistent definition of counterpressing, extracted 134 features
and manually labeled 20, 928 defensive transition situations from
97 professional football matches. The features describe the con-
stitution of the teams using attributes like team-center, stretch-
index (Santos, Theron, Losada, Sampaio, & Lago-Peñas, 2018),
or pressure on the ball carrier (Andrienko et al., 2017) and were
extracted at three discrete timestamps during a transition to de-
fense (at the time of the ball possession change, plus one and two
seconds after). This provides a drastic reduction of input data
dimensionality. We present a comprehensive inter-labeler relia-
bility with a pair-wise labeling accuracy of 82.01% and provide
rule-based baseline models in order to give an indication of the
task-complexity (area under the curve 60.2%). The trained XG-
Boost model—with an area under the curve of 87.4% on the la-
beled test data—enabled us to judge how quickly teams can win
the ball back with counterpressing strategies, how many shots
they create or allow immediately afterwards, and to determine
what the most important success drivers are. We applied this
automatic detection on all matches from six full seasons of the
German Bundesliga and quantified the defensive and offensive
consequences when applying counterpressing for each team.
19 Power et al. (2018) addressed defensive behaviors during corners as a team-tactical pat-
tern using supervised machine learning techniques.

37
 To capture all counterpressing situations manually, the full
match has to be observed at least once using a dedicated tagging
tool. Further efforts has to be spent to review the labels lead-
ing to an overall manual effort of roughly two hours per match.
In two experimental studies, the effort was reduced by auto-
matically suggesting 15–30 counterpressing scenes per match,
which can be immediately observed. Consequently, automating
the task saves analysts a tremendous amount of time, standard-
izes the otherwise subjective task, and allows to identify trends
within larger data sets. We present an effective way of how the
detection and the lessons learned from this investigation are in-
tegrated effectively into common match analysis processes. For
example, Figure 3 shows how the outcome of the counterpress-
ing detection is used in match-reports for the German national
teams. Instead of manually screening the video footage of all
transitions, the plot overviews the success rate of counterpress-
ing (defined as a regaining possession within the five seconds).
Considering only defensive transitions of the German team, Fig-
ure 3 shows that the detection leads an analyst directly to the
33 scenes of interest out of a total of 164 ball losses in the op-
posing half. One outcome of the study is that counterpressing
is a risky tactic. Figure 3 shows the German U21 national team
conceded two goals after unsuccessful counterpressing attempts
(within 20 seconds after the counterpress). The green/red bul-
lets show the ball losses before successful/unsuccessful counter-
pressing was conducted. In the bottom line, several performance
indicators are benchmarked against Bundesliga average.
A major limitation of the approach is the exhaustive human
endeavor required for such an experiment. In the presented
study, a total of 97 matches (roughly 140 hours of video footage)
had to be observed by at least one football expert. Not only

38
39
Figure 3: Counterpressing. Excerpt from match-report of the U21 national team match Germany against Belgium. This figure is copied from Bauer and
Anzer (2021) and further described in the full text (see Appendix).
labeling, but also feature extraction is an elaborate process. Al-
though many features could be re-used to identify other tactical
patterns, using end-to-end algorithms handling the raw data, as
presented in the next Section 3.4 (study IV), would seem to be
more efficient.

3.4 Study IV: Putting Team Formations in Association Foot-

ball into Context (Bauer, Anzer, & Shaw 2021)

Choosing the right team formation in football is a fundamen-

tal tactical and strategical decision for coaches. The availability
of accurate positional data motivated ample research on team
formations using supervised (Narizuka & Yamazaki, 2019; Shaw
& Glickman, 2019; Bialkowski et al., 2016, 2015, 2014; Wei et al.,
2013) and unsupervised approaches (Müller-Budack et al., 2019).
However (as indicated in study I) formations change dynami-
cally and therefore should not simply be aggregated over an en-
tire match (Andrienko et al., 2019; Shaw & Glickman, 2019; Gud-
mundsson et al., 2017; Bialkowski et al., 2016; Lucey, Bialkowski,
et al., 2013). Past literature focused primarily on aggregating
player positions across all game-states using positional data (Wei
et al., 2013; Bialkowski et al., 2014, 2015). Only Bialkowski et al.
(2016) and Shaw and Glickman (2019) explored differences in
team formations between different game-states, i.e. offensive,
defensive and transition. However, prior work did not consider
formations related to more granular phases of play like build-up
versus attacking play within offense.
To address this gap in literature, in study IV (Bauer, Anzer, &
Shaw, 2022), we first detect those phases of play from Figure 2 in
which formations are of interest for practitioners (namely build-
up, attacking play, high-/mid-/low-block) at each moment of the

40
game using a convolutional neural network with an average F1 -
score of 0.76. To train this model all phases of play of 97 matches
have been labeled manually at each frame (in total 59 hours and
50 minutes of the above listed phases were labeled). Again, for
this supervised machine learning approach, we present a verbose
labeling experiment using definitions consolidated among ex-
perts, different labelers to analyze their inter-labeler accordance
and baseline models indicating the complexity of the classifica-
tion task. Detecting those phases, Figure 4 shows the average po-
sitioning of a team (playing from left to right) in the respective
phases during a match, show that phases of play considerably
influence team formations. The formations have been contextu-
alized (i.e assigned to a 5–3–2 for the low-block phase of play)
in close consultation with football experts. The ellipses show the
80% confidence region for each player.
We then measure and contextualize unique formations per
phase of play by hierarchically clustering the respective sequences
for seven seasons of German Bundesliga (2013/2014–2019/2020).
Instead of over-simplifying a team’s formation across the whole
match into widely established three- or four-digit codes (e.g.,
4–4–2 abbreviating 4 defender, 4 midfielder and 2 attacker), we
provide an objective and granular representation of teams for-
mations per tactical phase of play. Using the most frequently
occurring phases of play, mid-block, we identify and contextu-
alise six unique formations (4–2–3–1, 4–4–2, 4–1–4–1, 4–3–2–1,
5–3–2 and 5–2–3). The definitions of the formations, including
the distinction between three and four lines, has been defined
by the involved experts. A long-term analysis in the German
Bundesliga allows us to quantify the efficiency of each forma-
tion against others, and also to present a helpful scouting tool
to identify how well a coach’s preferred playing style suits to a

41
42
Figure 4: Team formations per phases of play. A similar Figure is shown and further described in Bauer, Anzer, and Shaw (2022).
potential club.
Following our research question—the detection of tactical pat-
terns—this paper contains two major contributions:

(1) The supervised detection of five phases of play, namely low-

/mid-/high block, build-up and attacking play.

(2) The unsupervised exploration of team formations per phase

of play.

By transferring the raw positional data to images, and using con-

volutional neural networks, in (1) we present a major advantage
compared to the method presented in study III (Section 3.3). This
allows us to accurately detect five phases of play with just one
trained algorithm and without manual feature extraction. For (2)
we present an interpretation of team formations as tactical pat-
terns, that should rather be defined on specific sequences than
aggregated over a whole match or over basic game-states (offen-
sive, defensive, transition). In future investigations, more focus
has to be put on the analysis of formation efficiencies, which is
only touched upon in the practical application of our approach.

3.5 Study V: Individual Role Classification for Players de-

fending Corners in Football (Bauer, Anzer, & Smith 2021)

Choosing the right defensive corner-strategy is a crucial task

for each coach in professional football.20 Due to their repeat-
able and relatively static set-up, corners are an obligatory in-
vestigation for pattern analysis using positional data: Power et
al. (2018) detects the defensive strategy on a team-level (player-
marking, zonal-marking, hybrid) using expert-labeled data and
20 Followingmany other teams, the German national squad recently hired a dedicated set-
piece coach purely focusing on set-piece tactics https://www.kicker.de/als-hansi-anrief
-dachte-ich-da-verarscht-mich-jemand-867987/artikel, accessed 28.08.2021.

43
neural networks. Shaw and Gopaladesikan (2021) extended the
automated distinction between player- and zonal-marking to a
prediction on a player-level. For this task, the ordering problem
(mentioned in Section 2.2) becomes challenging. To overcome
this issue, Shaw and Gopaladesikan (2021) extracted features on
a player level and perform an XGBoost model predicting the role
for each player (neglecting player interactions).
Our work in Bauer, Anzer, and Smith (2022) addresses this
problem by combining a convolutional neural network (to han-
dle the ordering-problem) with a long-short-term memory (to
capture the temporal interaction of two players). By doing so,
we identify which of seven well-established roles a defensive
player conducted (player-marking, zonal-marking, placed for coun-
terattack, back-space, near-post, far-post, and short-defender). Fur-
ther, in case of player-marking we detect which attacking player
is marked, which is a relevant extension to Shaw and Gopalade-
sikan (2021). We hand-labeled the role of each defensive player
from 213 corners in 33 matches, where we then employ an aug-
mentation strategy to increase the number of data points. The
model achieves an overall weighted accuracy of 89.3%, and in the
case of player-marking, we are able to accurately detect which
offensive player the defender is marking 80.8% of the time. The
performance of the model is evaluated against a rule-based base-
line model, as well as by an inter-labeler accuracy.
For practical usage, we show three concrete use-cases on how
this approach can support a more informed and fact-based de-
cision making process for defensive corner strategies. Figure 5
shows an excerpt from a match-report using our algorithm. Each
grey or colored line indicates a player’s action for one corner of
that match. For each defensive player (red ellipses) the roles per
corner are shown (link to boxes on the left side). In case of player-

44
marking, the links to the opponents (green ellipses) show which
player was marked. To highlight the insights for coaches and
analysts using the report, the figure also indicates who touched
the ball first for each corner and whether an attacker was able to
create a goal or shot (within 18 seconds after the corner).

Figure 5: Defensive player roles for corners. Excerpt from match-report of the German U21
national team against Denmark. This figure is copied from Bauer, Anzer, and Smith (2022).

The largest limitation is the amount of manual effort required

to annotate the training data. To reduce the labeling time, the
data augmentation presented a massive improvement by multi-
plying the set of training data by a factor of ten without overfit-
ting the model (due to few unique samples). Further, we support

45
the labeling with rule-based suggestions, which are also used to
perform a weak supervision approach, included as an outlook
of the paper. Both for this approach, as well as for the phases
of play detection in Section 3.4, the computing complexity when
handling images using convolutional neural networks is another
relevant limitation. Transferring the spatio-temporal positional
data into a sparsely populated image actually increases the di-
mensionality of the input data. Although image classifications
is a well researched area, where plenty of tweaks in the archi-
tecture of convolutional neural network can handle those high
dimensional data, future work should consider using graph neu-
ral networks as indicated in Dick et al. (2021) or Stöckl, Seidl,
Marley, and Power (2021) to model players interaction and learn
individual classifications.

3.6 Study VI: Torward Automatically Labeling Situations in

Football (Fassmeyer et. al. 2021)

Using semi-supervised methodologies to reduce the expert-

labeling time required has become a relevant issue in machine
learning research. Previous studies showed that there is also a
need to apply such strategies when detecting tactical patterns
in football. Study VI, Fassmeyer et al. (2021) can be seen as an
outlook of the thesis in order to explore further semi-supervised
approaches towards the detection of tactical patterns.
We split the problem into two parts and learn (1) a meaningful
feature representation using variational autoencoders on unla-
beled data at large scales, and (2) a large-margin classifier acting
in this feature space but using only a few (manually) annotated
examples of the situation of interest. Both a static (one fixed time
frame) and a temporal (sequence of positional data) autoencoder

46
is implemented using the transformation of positional data into
an image also presented in studies IV & V. As a proof of concept
for the novel methodology, corner kicks and cross events are de-
tected by support vector machines applied to the encoded data
in a lower-dimensional space. The detection is compared against
event data available abundantly. Since the results are sufficient,
even compared against other approaches aiming to derive event
data from player positions (Stein et al., 2019; Richly et al., 2016;
Motoi et al., 2012; Gudmundsson & Wolle, 2010; M. Zheng & Ku-
denko, 2010), we further applied the approach to counterattacks,
a more complex tactical pattern that is of high interest for prac-
titioners (Hobbs et al., 2018). For counterattacks only 60 positive
examples (27 training data; 33 test data) of a single match had to
be hand-labeled to achieve a sufficient accuracy (area under the
curve: 0.912; F1 -score: 0.730) using the sequential variational au-
toencoder in combination with the support-vector machine clas-
sifier. For interpretability of the approaches, we investigate dif-
ferent false positives and false negatives of the detection in de-
tail, showing that even the misclassifications are reasonable for
experts.
The presented study suggests the potential of semi-supervised
methods for the identification of tactical patterns. The approach
focused on evaluating a new methodology, however, for practical
usage further work has to be conducted, i.e. a generalisation to
other patterns or the integration into an application.

4 Discussion

As shown in Table 1 and Figure 2, we present a substantial

contribution towards the detection of tactical patterns in football.
The first part of our research question—whether complex tactical

47
patterns in football can be detected automatically—can be affirmed
looking at Figure 2. All five standard offensive and defensive
phases of play are detected in study IV (Section 3.4), counter-
pressing (Section 3.3) and counterattacks (Section 3.6) are de-
tected in studies III & VI. Further, tactical patterns during cor-
ners, as an important example of a set-piece are investigated in
study V (Section 3.5). Referring to the introductory statements
of Richard Bate (Bate, 1987) and Simon Kuper (Kuper, 2018), we
focus on two primary values for football (rather than claiming to
provide groundbreaking insights that can fundamentally change
tactics and strategy): On the one hand, we believe that the ap-
proaches presented can provide evidence for coaches’ opinions
and thus support decision making on tactics and strategy. Precisely, a
simple frequency analysis on tactical patterns of interest (shown
in Figures 3 and 5) can help to objectify otherwise biased opin-
ions (Borrie et al., 2002), or to reconstruct all situations of interest
entirely (Laird & Waters, 2008). Second, we present an effec-
tive integration of the process automatization into the everyday-
business of professional clubs or federations, especially in stud-
ies III, IV & V. We show how repeated tasks, being conducted
typically by dedicated match analysts observing hours of video
footage, can be supported or fully realized by machine learning
algorithms. By saving time in various use-cases (e.g., opponent
analysis, team analysis, scouting of players or coaches) the appli-
cations presented can support the experts so that they can focus
on qualitative analysis of scenes of interest rather than repeat-
edly annotating those scenes in the video footage.
For the second part of the research question—the how of tac-
tical pattern detection—we present multiple ways using different
methodologies: exploratory visual analytics (study I), unsuper-
vised machine learning methods (study II, study IV), supervised

48
methods (study III, study IV, study V), as well as semi-supervised
approaches (study VI) and other labeling support methods (i.e.,
rule-based support, data augmentation and weak supervision
in study V). For the detection of tactical patterns, the interplay
between unsupervised or other exploratory approaches (e.g., visual
analytics) and supervised techniques has proven to be a fruitful
combination on the analysis of tactics and strategy in football.
This dissertation indicates that the related learnings can be trans-
ferred to various domains, and aligns with Tuyls et al. (2021),
claiming that football analytics can offer huge potential for the
research domain of machine learning and artificial intelligence
Transforming spatio-temporal multi-agent data into a permutation-
invariant space is not only a problem in sports analytics (Battaglia
et al., 2018). Recent improvements in graph neural networks
can solve this problem, whereby positional data in football and
basketball are favored testbeds due to the vividness of results
(Dick et al., 2021; Games, 2019; Yeh et al., 2019; Sun et al., 2019;
Kipf, Fetaya, Wang, Welling, & Zemel, 2018). A similar con-
cept—stating that theoretical research domains can benefit from
application areas and vice versa—has been proposed by Höner
(2008) at the intersection of sport science and psychology. As
mentioned in the introduction, various work has been conducted
outside this thesis bridging the gap between sport and data sci-
ence: In Herold et al. (2021) and Herold et al. (2019) the aware-
ness of football practitioners and data-driven metrics is analyzed.
In Anzer, Bauer, and Höner (2021), the basic concept of machine
learning (explained with references to football examples) is pre-
sented in a match analysis textbook. Anzer and Bauer (2022)
and Anzer and Bauer (2021) use granular metrics to objectively
quantify passing and shooting performance. These metrics are
not only established in sport science research (Herold et al., 2021)

49
but also support performance evaluation in practice.
The methodological approach, specifically the rigorousness
while creating and evaluating labeled data in our studies, is
comparable to Chawla et al. (2017). For example, Chawla et al.
(2017) hand-labeled in total 2, 932 passes manually, whereas in
study III we use 3, 196 expert-annotated counterpressing scenes
for our detection. As in Chawla et al. (2017)21 we show a high
inter-labeler accordance in study III (pairwise labeler accuracy
82.01%), study IV (e.g., inter-labeler average F1 -score for mid-
block 0.78) and study V (e.g., pairwise accuracy of player-marking
detection 94.3%). However, the application on team-tactical pat-
terns presented in this thesis is unique. For each tactical pattern
detected in the supervised set-up (studies III, IV, V) we consol-
idate definitions in close collaboration with experts, create an-
notated data from different expert-labelers and steadily monitor
their labeling accordance. A central learning of this thesis is that
the inter-labeler reliability, as a measure of how well-defined a
pattern is, should be monitored as an integral part of such stud-
ies. Related to this, the close collaboration with practitioners can be
seen as another major strength of this dissertation. F. R. Goes,
Brink, Elferink-Gemser, Kempe, and Lemmink (2020); Herold et
al. (2019) and Rein and Memmert (2016) allege that collabora-
tions among machine learning and sport science experts is key to
success for the area of sports analytics. This is even more appre-
ciable for manual expert-labeling, hand-crafted feature extrac-
tion, the contextualisation of the results and, last but not least,
the concrete problem definition—all fundamental constituents of
supervised learning. Hence, for projects using supervised ma-
21 Theinter-labeler reliability is presented as the Cohen’s kappa. For the six class prediction
(very good, good, marginally good, marginally bad, bad, very bad) the Cohen’s kappa among
the two labelers is 0.393, which is drastically improved with the three-class classification
(good, bad, ok) yielding a Cohen’s kappa of 0.697.

50
chine learning, we claim that the beneficiaries of the approach
must be involved in project teams. Further, we maintain that
there is a need for data-literate match analysts and coaches with a
basic understanding of machine learning methodologies in order
to transfer insights to practice. This will allow them to contribute
to projects and to draw informed recommendations for actions
based on data-driven results.
Whenever possible we compared our results against published
benchmarks in the respective studies, especially in study III, IV
& V (see Appendices). A comparison across sports and differ-
ent patterns (with different complexities) warps reality, however,
similar accuracies are achieved as presented in literature. With
an area under the curve of 0.874, counterpressing—a complex
team-tactical pattern in football—is detected in study III. The re-
sults are comparable to the detection of screen plays —a group-
tactical pattern in basketball—in Hojo et al. (2018) (area under
the curve: 0.855 off-ball, 0.941 on-ball). The F1 -score of 0.748, pre-
sented in Chawla et al. (2017) classifying passes according their
reward is comparable to our counterpressing detection (study II;
F1 -score: 0.67), the frame-wise detection of mid-blocks (F1 -score:
0.80) or build-up plays (F1 -score: 0.83) in study IV. Again, these
comparisons have major flaws due to way different complexities
of the patterns and confounding factors like data-quality, inter-
labeler reliability, and many more, and should serve as a rough
orientation only. To provide an idea of the task complexity, stud-
ies III, IV & V contain rule-based baseline models helping to give
context to the accuracy of the machine learning model. Accord-
ing to this thesis, understandable baseline models should also be
standardized for future work on the detection of tactical patterns.

51
4.1 Limitations and Future Work

Despite the great level of detail in which data in football are

available, remaining inaccuracies, objectiveness, human error (i.e.,
for event data collection), and missing information (e.g., player
pose) can still be seen as a pitfall of all data-driven approaches in
football. The major shortcoming of optical tracking systems is,
on the one hand, that they aggregate complex movements of a
player to a two-dimensional position, neglecting players orienta-
tion and pose (Arbués-Sangüesa, Haro, Ballester, & Martín, 2019;
Arbues-Sanguesa, Martin, Fernandez, Ballester, & Haro, 2020).
On the other hand, they require an exhaustive set-up on-site
with cameras from different, demanding positions (Manafifard
et al., 2017). Thus, the acquisition of accurate positional data
from basic video-footage (e.g., television signals), as well as ap-
propriate signal processing to compare positional data from dif-
ferent tracking systems (Taberner et al., 2020) are relevant issues
for future research. Another pertinent issue is the automated
creation of event data from video footage (Pouyanfar & Chen,
2016; Kolekar et al., 2009; Ekin et al., 2003) or from positional
data (Stein et al., 2019; Richly et al., 2016; Motoi et al., 2012;
Gudmundsson & Wolle, 2010; M. Zheng & Kudenko, 2010). Al-
though event data collection uses sophisticated software-support
to make the hand-crafted annotation process as efficient as pos-
sible, automating and objectifying this manual task would lever-
age sports analytics to another level. However, given the recent
success in data collection and data quality enabling sports ana-
lytics to emerge to a growing research area, we expect substantial
further development on the accuracy and granularity of data in
all sports.
Nevertheless, positional and event data will always be an ab-

52
straction of the real world proceedings, which does not capture
all facets of a football match. Thus, more interdisciplinary ap-
proaches are required to cover confounding factors (e.g., psycho-
logical components) overlooked in the digital reproduction. Es-
tablishing the interplay between qualitative and quantitative method-
ologies poses a great potential for future investigations. In the
studies presented, the formulation of definitions is often prag-
matic and purpose-driven in order to perform the detection task
with preferably high accuracy. We state that more sophisticated
qualitative studies, grounded solidly on match analysis literature
and rigorous expert interviews, should be established to derive
proper definitions (e.g., for tactical patterns in general or specific
tactical patterns like counterpressing) and also to develop and
consolidate taxonomies as presented in Figure 2.
As a third limitation, the exhaustive labeling effort required to
perform supervised machine learning does not meet the practi-
cal requirements of club-, team-, or coach-specific interpretations
of tactical patterns. Evolving patterns and the high fluctuation
of key-roles in professional football (i.e., coaches and managers)
means integrating new philosophies and setting new priorities
in a short amount of time. Although Grunz et al. (2012) affirmed
that it is hard to formulate precise definitions for tactical pat-
terns, and consequently, that rule-based definitions are inappro-
priate, study V (Bauer, Anzer, & Smith, 2022) shows that heuris-
tics can add value to supervised machine learning approaches
in different aspects. They can serve as baseline models (see also
studies III & IV), simplify the human labeling process in com-
plex scenarios (like individual player annotations at corners in
study V), and even create weakly supervised training data. Stud-
ies V & VI also shows that data augmentation (Bauer, Anzer,
& Smith, 2022) and variational autoencoders (Fassmeyer et al.,

53
2021) can drastically reduce labeling efforts without loosing ac-
curacy. However, labeling support methodologies such as ac-
tive learning (Druck, Settles, & McCallum, 2009), weak supervi-
sion (Ratner et al., 2017), semi-supervised learning (Cholaquidis,
Fraiman, & Sued, 2020) or transfer learning (Panigrahi, Nanda,
& Swarnkar, 2021) are of high relevance for the area of football
(and sports) analytics and should guide future work of the de-
tection of tactical patterns.
Another general problem in sports analytics is comparability
and reproducibility. Positional and event data are typically the
confidential property of leagues or clubs. As a consequence,
only researches working for these organisations or closely re-
lated research groups can access the respective data. In the-
ory, the increasing availability of open-source datasets (see Sec-
tion 2.1) allows for reproducibility. However, in practice, dif-
ferent definitions for event data and tactical patterns, varying
data-accuracies and other factors constrain comparisons. To en-
sure reproducibility, a wide range of researchers need to ac-
cess the exact same dataset to transparently compare their re-
sults. Recently, in American football22 , basketball23 and foot-
ball24 positional and event data were open-sourced within (pub-
lic) competitions. These competitions not only attracted plenty
of (machine learning) research groups (which would rarely ac-
cess sports data otherwise), but also laid the foundation for solid
scientific progress in the area of sports analytics. In football, an
objective comparison of the top European leagues, as well as the
22 The NFL Big Data Bowl is hosted annually since 2018/2019 by the National Football
League (NFL) endowed with a price of up to 100.000$ (https://operations.nfl.com/
gameday/analytics/big-data-bowl/).
23 In 2019, the National Basketball Association (NBA) hosted a closed hackathon (https://

hackathon.nba.com/).
24 The DFB-Akademie, Eintracht Frankfurt and the Sportec Solutions AG hosted a

Hackathon with selected participants in 2020 (https://www.dfb-akademie.de/hackathon2/

-/id-11009109).

54
big international competitions on a team and federation level,
exploits a huge potential for future work.

5 Conclusion

Concluding the overall research program, positional and event

data can help practitioners to make informed decisions instead
of purely relying on their gut instincts. Methods of machine
learning can further help to capture and model the dynamics
and complexity in football. Purely applying these methods to
positional and event data will not suffice, they should rather
be integrated into interdisciplinary approaches combined with
qualitative research and domain-experts.
For the detection of tactical patterns, this dissertation shows
these patterns across all phases of play are well-defined, can be
identified by experts with a high accordance and can automat-
ically be detected when applying machine learning algorithms
on positional and event data. Additionally, this interdisciplinary
approach does not only combine different methodologies (i.e.
supervised machine learning with exploratory approaches), but
attests to the effectiveness of integrating the results into the ev-
eryday business of professional football clubs. By doing so, this
thesis shows how match analysts and coaches can save time for
the otherwise manual task of annotating tactical patterns, and
provides a baseline for evidence-based decisions on tactics and
strategies.

55
References
Albert, J. (2010). Sabermetrics: The past, the present, and the future. Mathe-
matics and Sports, 3–14. doi: 10.5948/UPO9781614442004.002
Alcock, A. M. (2010). Analysis of direct shots at goal from free kicks in elite
women ’ s football Analysis of Direct Shots at Goal from Free Kicks.
Dissertation School of Health and Human Sciences, Southern Cross University,
Lismore, Australia.
Alexander, J. P., Spencer, B., Sweeting, A. J., Mara, J. K., & Robertson, S.
(2019). The influence of match phase and field position on collective
team behaviour in Australian Rules football. Journal of Sports Sciences,
37(15), 1699–1707. Retrieved from https://doi.org/10.1080/02640414
.2019.1586077 doi: 10.1080/02640414.2019.1586077
Alloghani, M., Al-Jumeily, D., Mustafina, J., Hussain, A., & Aljaaf, A. J. (2020).
A Systematic Review on Supervised and Unsupervised Machine Learning Al-
gorithms for Data Science. Springer, Cham. Retrieved from https://link
.springer.com/chapter/10.1007/978-3-030-22475-2_1 doi: 10.1007/
978-3-030-22475-2-1
Andrienko, G., Andrienko, N., Anzer, G., Bauer, P., Budziak, G., Fuchs, G., . . .
Wrobel, S. (2019). Constructing Spaces and Times for Tactical Analysis in
Football. IEEE Transactions on Visualization and Computer Graphics, 27(4),
2280–2297. Retrieved from https://ieeexplore.ieee.org/document/
8894420 doi: 10.1109/TVCG.2019.2952129
Andrienko, G., Andrienko, N., Budziak, G., Dykes, J., Fuchs, G., von Lan-
desberger, T., & Weber, H. (2017). Visual analysis of pressure in foot-
ball. Data Mining and Knowledge Discovery, 31(6), 1793–1839. Retrieved
from https://link.springer.com/article/10.1007/s10618-017-0513
-2 doi: 10.1007/s10618-017-0513-2
Anzer, G. (2021). Large Scale Analysis Offensive Performance in Football—Using
Synchronized Positional and Event Data to Quantify Offensive Actions, Tac-
tics, and Strategies (Unpublished doctoral dissertation). Eberhard Karls
University Tübingen.
Anzer, G., & Bauer, P. (2021). A Goal Scoring Probability Model based on Syn-
chronized Positional and Event Data. Frontiers in Sports and Active Learn-
ing (Special Issue: Using Artificial Intelligence to Enhance Sport Performance),
3(0), 1–18. Retrieved from https://www.frontiersin.org/articles/
10.3389/fspor.2021.624475/full doi: 10.3389/fspor.2021.624475
Anzer, G., & Bauer, P. (2022). Expected Passes—Determining the Difficulty

56
 of a Pass in Football (Soccer) Using Spatio-Temporal Data. Data Min-
ing and Knowledge Discovery, Springer US. Retrieved from https://link
.springer.com/article/10.1007/s10618-021-00810-3 doi: 10.1007/
s10618-021-00810-3
Anzer, G., Bauer, P., & Brefeld, U. (2021). The origins of goals in the Ger-
man Bundesliga. Journal of Sport Science. Retrieved from https://www
.tandfonline.com/doi/full/10.1080/02640414.2021.1943981 doi:
10.1080/02640414.2021.1943981
Anzer, G., Bauer, P., & Höner, O. (2021). The Identification of Counter-
pressing in Football. In D. Memmert (Ed.), Match analysis—how to use
data in professional sport (1st Editio ed., pp. 228–235). New York: Rout-
ledge. Retrieved from https://doi.org/10.4324/9781003160953 doi:
10.4324/9781003160953
Araújo, D., Couceiro, M., Seifert, L., Sarmento, H., & Davids, K. (2021). Arti-
ficial Intelligence in Sport Performance Analysis (1st Editio ed.). New York:
Routledge. Retrieved from https://doi.org/10.4324/9781003163589
doi: 10.4324/9781003163589
Arbués-Sangüesa, A., Haro, G., Ballester, C., & Martín, A. (2019). Head,
Shoulders, Hip and Ball... Hip and Ball! Using Pose Data to Lever-
age Football Player Orientation. Barça sports analytics summit, Barcelona
(Spain), 1–13.
Arbues-Sanguesa, A., Martin, A., Fernandez, J., Ballester, C., & Haro, G.
(2020). Using player’s body-orientation to model pass feasibility in
soccer. IEEE Computer Society Conference on Computer Vision and Pattern
Recognition Workshops, 2020-June, 3875–3884. doi: 10.1109/CVPRW50498
.2020.00451
Atmosukarto, I., Ghanem, B., Ahuja, S., Ahuja, N., & Muthuswamy, K. (2013).
Automatic recognition of offensive team formation in american football
plays. IEEE Computer Society Conference on Computer Vision and Pattern
Recognition Workshops, 991–998. doi: 10.1109/CVPRW.2013.144
Barris, S., & Button, C. (2008). A review of vision-based motion analysis
in sport. Sports Medicine, 38(12), 1025–1043. doi: 10.2165/00007256
-200838120-00006
Bartlett, R., Button, C., Robins, M., Dutt-Mazumder, A., & Kennedy, G. (2012).
Analysing team coordination patterns from player movement trajecto-
ries in soccer: Methodological considerations. International Journal of
Performance Analysis in Sport, 12(2), 398–424. doi: 10.1080/24748668.2012
.11868607

57
Bate, R. (1987). Football Chance: Tactics and Strategy. Science and Football:
Proceedings of the first World Congress of Science and Football, 13–17.
Battaglia, P. W., Hamrick, J. B., Bapst, V., Sanchez-Gonzalez, A., Zambaldi, V.,
Malinowski, M., . . . Pascanu, R. (2018). Relational inductive biases, deep
learning, and graph networks.
Bauer, P., & Anzer, G. (2021). Data-driven detection of counterpressing
in professional football—A supervised machine learning task based
on synchronized positional and event data with expert-based feature
extraction. Data Mining and Knowledge Discovery, 35(5), 2009–2049.
Retrieved from https://link.springer.com/article/10.1007/s10618
-021-00763-7 doi: 10.1007/s10618-021-00763-7
Bauer, P., Anzer, G., & Shaw, L. (2022). Putting Team Formations in Associa-
tion Football into Context. Journal of Sports Analytics (submitted).
Bauer, P., Anzer, G., & Smith, J. W. (2022). Individual role classification
for players defending corners in football (soccer)—Categorisation of the
defensive role for each player in a corner kick using positional data.
Journal of Quantitative Analysis in Sports (submitted).
Baumer, B., & Zimbalist, A. (2014). The Sabermetric Revolution. University of
Pennsylvania Press. doi: 10.9783/9780812209129
Beal, R., Chalkiadakis, G., Norman, T. J., & Ramchurn, S. D. (2020). Op-
timising game tactics for football. Proceedings of the International Joint
Conference on Autonomous Agents and Multiagent Systems, AAMAS, 2020-
May(May), 141–149.
Beal, R., Norman, T. J., & Ramchurn, S. D. (2019). Artificial intelligence for
team sports: A survey. Knowledge Engineering Review, 34, 1–37. doi:
10.1017/S0269888919000225
Bialkowski, A., Lucey, P., Carr, P., Matthews, I., Sridharan, S., & Fookes, C.
(2016). Discovering team structures in soccer from spatiotemporal data.
IEEE Transactions on Knowledge and Data Engineering, 28(10), 2596–2605.
doi: 10.1109/TKDE.2016.2581158
Bialkowski, A., Lucey, P., Carr, P., Yue, Y., Sridharan, S., & Matthews, I.
(2014). Large-Scale Analysis of Soccer Matches Using Spatiotemporal
Tracking Data. IEEE International Conference on Data Mining, ICDM (Pro-
ceeding)(January), 725–730. doi: 10.1109/ICDM.2014.133
Bialkowski, A., Lucey, P., Carr, P., Yue, Y., Sridharan, S., & Matthews, I. (2015).
Identifying team style in soccer using formations learned from spa-
tiotemporal tracking data. IEEE International Conference on Data Mining
Workshops, ICDMW(January), 9–14. doi: 10.1109/ICDMW.2014.167

58
Borrie, A., Jonsson, G. K., & Magnusson, M. S. (2002). Temporal pat-
tern analysis and its applicability in sport: An explanation and ex-
emplar data. Journal of Sports Sciences, 20(10), 845–852. doi: 10.1080/
026404102320675675
Brefeld, U., Lasek, J., & Mair, S. (2019). Probabilistic movement models
and zones of control. Machine Learning, 108(1), 127–147. Retrieved
from https://doi.org/10.1007/s10994-018-5725-1 doi: 10.1007/
s10994-018-5725-1
Brooks, J., Kerr, M., & Guttag, J. (2016). Developing a data-driven player
ranking in soccer using predictive model weights. Proceedings of the ACM
SIGKDD International Conference on Knowledge Discovery and Data Mining,
49–55. doi: 10.1145/2939672.2939695
Buchheit, M., Allen, A., Poon, T. K., Modonutti, M., Gregson, W., & Di Salvo,
V. (2014). Integrating different tracking systems in football: mul-
tiple camera semi-automatic system, local position measurement and
GPS technologies. Journal of Sports Sciences, 32(20), 1844–1857. Re-
trieved from http://dx.doi.org/10.1080/02640414.2014.942687 doi:
10.1080/02640414.2014.942687
Bunker, R., Fujii, K., Hanada, H., & Takeuchi, I. (2020). Supervised sequential
pattern mining of event sequences in sport to identify important pat-
terns of play: an application to rugby union. (October). Retrieved from
http://arxiv.org/abs/2010.15377 doi: 10.31236/osf.io/g2bj8
Cai, Y., Wu, S., Zhao, W., Li, Z., Wu, Z., & Ji, S. (2018). Concussion classi-
fication via deep learning using whole-brain white matter fiber strains.
PLoS ONE, 13(5), 1–21. doi: 10.1371/journal.pone.0197992
Camerino, O. F., Chaverri, J., Anguera, M. T., & Jonsson, G. K. (2012). Dy-
namics of the game in soccer: Detection of T-patterns. European Journal
of Sport Science, 12(3), 216–224. doi: 10.1080/17461391.2011.566362
Cardinale, M. (2006). Validation of Prozone ® : A new video- based per-
formance analysis system. International Journal of Performance Analysis in
Sport, 6(1), 108–116.
Casal, C. A., Maneiro, R., Ardá, T., Losada, J. L., & Rial, A. (2015). Analysis of
corner kick success in elite football. International Journal of Performance
Analysis in Sport, 15(2), 430–451. doi: 10.1080/24748668.2015.11868805
Chawla, S., Estephan, J., Gudmundsson, J., & Horton, M. (2017). Classification
of passes in football matches using spatiotemporal data. ACM Transac-
tions on Spatial Algorithms and Systems, 3(2). doi: 10.1145/3105576
Chen, H. T., Chou, C. L., Fu, T. S., Lee, S. Y., & Lin, B. S. P. (2012). Recognizing

59
 tactic patterns in broadcast basketball video using player trajectory. Jour-
nal of Visual Communication and Image Representation, 23(6), 932–947. Re-
trieved from http://dx.doi.org/10.1016/j.jvcir.2012.06.003 doi:
10.1016/j.jvcir.2012.06.003
Chen, S., Feng, Z., Lu, Q., Mahasseni, B., Fiez, T., Fern, A., & Todorovic,
S. (2014). Play Type Recognition in Real-World Football Video. IEEE
Winter Conference on Applications of Computer Vision, 652–659. doi: 10
.1109/WACV.2014.6836040.
Chen, T., & Guestrin, C. (2016). XGBoost: A scalable tree boosting system.
Proceedings of the ACM SIGKDD International Conference on Knowledge Dis-
covery and Data Mining, 13-17, 785–794. doi: 10.1145/2939672.2939785
Cholaquidis, A., Fraiman, R., & Sued, M. (2020). On semi-supervised learning.
MIT Press, 29(4), 914–937. doi: 10.1007/s11749-019-00690-2
Courel-Ibáñez, J., McRobert, A. P., Toro, E. O., & Vélez, D. C. (2017). Collective
behaviour in basketball: A systematic review. International Journal of
Performance Analysis in Sport, 17(1-2), 44–64. doi: 10.1080/24748668.2017
.1303982
Decroos, T., Van Haaren, J., & Davis, J. (2018). Automatic discovery of tactics
in spatio-temporal soccer match data. Proceedings of the ACM SIGKDD
International Conference on Knowledge Discovery and Data Mining, 223–232.
doi: 10.1145/3219819.3219832
Desporto, J. (2009). Trends of tactical performance analysis in team
sports: bridging the gap between research, training and com-
petition. Revista Portuguesa de Ciências do Desporto, 9(1), 81–
89. Retrieved from http://www.scielo.mec.pt/scielo.php?pid=S1645
-05232009000100008&script=sci_arttext&tlng=es
Dick, U., & Brefeld, U. (2019). Learning to Rate Player Positioning in Soccer.
Big Data, 7(1), 71–82. doi: 10.1089/big.2018.0054
Dick, U., Tavakol, M., & Brefeld, U. (2021). Rating Player Actions in Soccer.
Frontiers in Sports and Active Learning (Special Issue: Using Artificial Intel-
ligence to Enhance Sport Performance), 3, 174. Retrieved from https://
www.frontiersin.org/article/10.3389/fspor.2021.682986 doi: 10
.3389/fspor.2021.682986
D’Orazio, T., & Leo, M. (2010, 8). A review of vision-based systems for
soccer video analysis. Pattern Recognition, 43(8), 2911–2926. doi: 10
.1016/j.patcog.2010.03.009
Druck, G., Settles, B., & McCallum, A. (2009). Active learning by labeling
features. EMNLP 2009 - Proceedings of the 2009 Conference on Empirical

60
 Methods in Natural Language Processing: A Meeting of SIGDAT, a Special
Interest Group of ACL, Held in Conjunction with ACL-IJCNLP 2009, 81–90.
doi: 10.3115/1699510.1699522
Duarte, R., Araújo, D., Folgado, H., Esteves, P., Marques, P., & Davids, K.
(2013). Capturing complex, non-linear team behaviours during com-
petitive football performance. Journal of Systems Science and Complexity,
26(1), 62–72. doi: 10.1007/s11424-013-2290-3
Dyk, D. A., & Meng, X. L. (2001). The art of data augmentation. Journal of Com-
putational and Graphical Statistics, 10(1), 1–50. Retrieved from https://
www.tandfonline.com/doi/abs/10.1198/10618600152418584 doi: 10
.1198/10618600152418584
Ekin, A., Tekalp, A. M., & Mehrotra, R. (2003). Automatic soccer video anal-
ysis and summarization. IEEE Transactions on Image Processing, 12(7),
796–807. doi: 10.1109/TIP.2003.812758
Fassmeyer, D., Anzer, G., Bauer, P., & Brefeld, U. (2021). Toward Auto-
matically Labeling Situations in Soccer. Frontiers in Sports and Active
Living, 3(November). Retrieved from https://www.frontiersin.org/
articles/10.3389/fspor.2021.725431/full doi: 10.3389/fspor.2021
.725431
Fernandez, J., & Bornn, L. (2018). Wide Open Spaces : A statistical technique
for measuring space creation in professional soccer. MIT Sloan Sports
Analytics Conference, Boston (USA), 1–19.
Fernández, J., Bornn, L., & Cervone, D. (2021). A framework for the fine-grained
evaluation of the instantaneous expected value of soccer possessions (Vol. 110)
(No. 6). Springer US. Retrieved from https://doi.org/10.1007/s10994
-021-05989-6 doi: 10.1007/s10994-021-05989-6
Fernando, T., Wei, X., Fookes, C., Sridharan, S., & Lucey, P. (2015).
Discovering Methods of Scoring in Soccer Using Tracking Data.
KDD Workshop on Large-Scale Sports Analytics, 1–4. Retrieved from
https://large-scale-sports-analytics.org/Large-Scale-Sports
-Analytics/Submissions2015_files/paperID19-Tharindu.pdf
Feuerhake, U. (2016). Recognition of Repetitive Movement Patterns—The
Case of Football Analysis. ISPRS International Journal of Geo-Information,
5(11), 208. doi: 10.3390/ijgi5110208
Fujii, K. (2021). Data-Driven Analysis for Understanding Team Sports
Behaviors. Journal of Robotics and Mechatronics, 33(3), 505–514. doi:
10.20965/jrm.2021.p0505
Fujimura, A., & Sugihara, K. (2005, 6). Geometric analysis and quantitative

61
 evaluation of sport teamwork. Systems and Computers in Japan, 36(6),
49–58. doi: 10.1002/scj.20254
Games, M.-a. S. (2019). A Graph Attention Based Approach for Trajectory
Prediction in Multi-agent Sports Games. Preprint (arXiv).
Gentleman, R., & Carey, V. J. (2008). Unsupervised Machine Learning. In Bio-
conductor case studies (pp. 137–157). Springer, New York, NY. Retrieved
from https://link.springer.com/chapter/10.1007/978-0-387-77240
-0_10 doi: 10.1007/978-0-387-77240-0
Ghory, I. (2004). Reinforcement learning in board games. Depart-
ment of Computer Science, University of Bristol, 1–57. Retrieved
from http://scholar.google.com/scholar?hl=en&btnG=Search&q=
intitle:Reinforcement+learning+in+board+games+.#0
Goes, F., Schwarz, E., Elferink-Gemser, M., Lemmink, K., & Brink, M. (2021).
A risk-reward assessment of passing decisions: comparison between po-
sitional roles using tracking data from professional men’s soccer. Science
and Medicine in Football, 1–9. Retrieved from https://doi.org/10.1080/
24733938.2021.1944660 doi: 10.1080/24733938.2021.1944660
Goes, F. R., Brink, M. S., Elferink-Gemser, M., Kempe, M., & Lemmink,
K. A. (2020). The tactics of successful attacks in professional association
football—large-scale spatiotemporal alanalysis of dynamic subgroups
using position tracking data. Journal of Sports Sciences, 39(5), 523–532.
doi: 10.1080/02640414.2020.1834689
Goes, F. R., Kempe, M., Meerhoff, L. A., & Lemmink, K. A. (2019). Not
Every Pass Can Be an Assist: A Data-Driven Model to Measure Pass
Effectiveness in Professional Soccer Matches. Big Data, 7(1), 57–70. doi:
10.1089/big.2018.0067
Goes, F. R., Meerhoff, L. A., Bueno, M. J., Rodrigues, D. M., Moura, F. A.,
Brink, M. S., . . . Lemmink, K. A. (2020). Unlocking the potential of
big data to support tactical performance analysis in professional soccer:
A systematic review. European Journal of Sport Science, 0(0), 1–16. Re-
trieved from https://doi.org/10.1080/17461391.2020.1747552 doi:
10.1080/17461391.2020.1747552
Goldberg, X. (2009, 6). Introduction to semi-supervised learning. Synthesis
Lectures on Artificial Intelligence and Machine Learning, 6, 1–116. doi: 10
.2200/S00196ED1V01Y200906AIM006
Gómez-Jordana, L. I., Milho, J., Ric, , Silva, R., & Passos, P. (2019). Landscapes
of passing opportunities in Football – where they are and for how long
are available ? Barça sports analytics summit, Barcelona (Spain), 1–14.

62
González-Ródenas, J., López-Bondia, I., Aranda-Malavés, R., Tudela Desantes,
A., Sanz-Ramírez, E., & Aranda Malaves, R. (2019). Technical, tactical
and spatial indicators related to goal scoring in European elite soccer.
Journal of Human Sport and Exercise, 15(1), 186–201. doi: 10.14198/jhse
.2020.151.17
Goodfellow, I., Bengio, Y., & Courville, A. (2016). Deep Learning. MIT Press.
Gould, P., & Gatrell, A. (1979). A structural analysis of a game: The Liverpool
v Manchester united cup final of 1977. Social Networks, 2(3), 253–273.
doi: 10.1016/0378-8733(79)90017-0
Gréhaigne, J. F., Bouthier, D., & David, B. (1997). Dynamic-system analysis
of opponent relationships in collective actions in soccer. Journal of Sports
Sciences, 15(2), 137–149. doi: 10.1080/026404197367416
Gréhaigne, J. F., Godbout, P., & Bouthier, D. (1999). The foundations of tactics
and strategy in team sports. Journal of Teaching in Physical Education,
18(2), 159–174. doi: 10.1123/jtpe.18.2.159
Grunz, A., Memmert, D., & Perl, J. (2009). Analysis and simulation of actions
in games by means of special self-organizing maps. International Journal
of Computer Science in Sport, 8(1), 22–37.
Grunz, A., Memmert, D., & Perl, J. (2012). Tactical pattern recognition in soc-
cer games by means of special self-organizing maps. Human Movement
Science, 31(2), 334–343. Retrieved from http://dx.doi.org/10.1016/
j.humov.2011.02.008 doi: 10.1016/j.humov.2011.02.008
Gudmundsson, J., & Horton, M. (2017). Spatio-temporal analysis of team
sports. ACM Computing Surveys, 50(2), 1–34. doi: 10.1145/3054132
Gudmundsson, J., Laube, P., & Wolle, T. (2017). Movement Patterns in Spatio-
Temporal Data (Z. X. Shekhar S. Xiong H., Ed.). Boston: Springer. doi:
10.1007/978-3-319-17885-1
Gudmundsson, J., & Wolle, T. (2010). Towards automated football analysis:
Algorithms and data structures. Proc. 10th Australas. Conf. Math. Comput.
Sport.
Haaren, J. V., Zimmermann, A., Renkens, J., Broeck, G. V. D., Beéck, T. O. D.,
Meert, W., & Davis, J. (2013). Machine Learning and Data Mining for Sports
Analytics (No. September). doi: 10.1007/978-3-030-64912-8
Han, J., Kamber, M., & Pei, J. (2012). Data Mining: Concepts and Techniques.
Elsevier Inc. doi: 10.1016/C2009-0-61819-5
Hennessy, L., & Jeffreys, I. (2018). The current use of GPS, its potential, and
limitations in soccer. Strength and Conditioning Journal, 40(3), 83–94. doi:
10.1519/SSC.0000000000000386

63
Hernández-mendo, A. (2006). Hidden patterns of play interaction in soccer
using SOF-CODER. Behaviour Research Methods(3), 372–381.
Herold, M., Goes, F., Nopp, S., Bauer, P., Thompson, C., & Meyer, T. (2019).
Machine learning in men’s professional football: Current applications
and future directions for improving attacking play. International Journal
of Sports Science and Coaching, 14(6). doi: 10.1177/1747954119879350
Herold, M., Kempe, M., Bauer, P., & Meyer, T. (2021). Attacking key perfor-
mance indicators in soccer: Current practice and perceptions from the
elite to youth academy level. Journal of Sports Science and Medicine, 20(1),
158–169. Retrieved from https://doi.org/10.52082/jssm.2021.158
doi: 10.52082/jssm.2021.158
Hirano, S., & Tsumoto, S. (2004). Finding interesting pass patterns from soccer
game records. Knowledge Discovery in Databases: PKDD 2004; Springer
Berlin Heidelberg, 3202, 209–218. doi: 10.1007/978-3-540-30116-5{\_}21
Hobbs, J., Power, P., Sha, L., Ruiz, H., & Lucey, P. (2018). Quantifying the
Value of Transitions in Soccer via Spatiotemporal Trajectory Clustering.
MIT Sloan Sports Analytics Conference, Boston (USA), 1–11.
Hochreiter, S., & Schmidhuber, J. (1996). Long Short-Term Memory. Hardware-
Software Co-Synthesis of Distributed Embedded Systems, 9(8), 13–39. doi:
10.1007/978-1-4757-5388-2
Hochstedler, J., & Gagnon, P. T. (2017). American Football Route Identifi-
cation Using Supervised Machine Learning. MIT Sloan Sports Analytics
Conference, Boston (USA), 1–11.
Hojo, M., Fujii, K., Inaba, Y., Motoyasu, Y., & Kawahara, Y. (2018). Au-
tomatically recognizing strategic cooperative behaviors in various sit-
uations of a team sport. PLoS ONE, 13(12), 1–15. doi: 10.1371/
journal.pone.0209247
Höner, O. (2008). Basiert die Sportwissenschaft auf unterschiedlichen
„Sorten” von Theorien? Sportwissenschaft, 38(1), 3–23. doi: 10.1007/
bf03356066
Hosp, B., Schultz, F., Kasneci, E., & Höner, O. (2021). Expertise Classifica-
tion of Soccer Goalkeepers in Highly Dynamic Decision Tasks: A Deep
Learning Approach for Temporal and Spatial Feature Recognition of
Fixation Image Patch Sequences. Frontiers in Sports and Active Living,
3(July). doi: 10.3389/fspor.2021.692526
Hosp, B. W., Schultz, F., Honer, O., & Kasneci, E. (2021). Soccer goalkeeper
expertise identification based on eye movements. PLoS ONE, 16(5 May),
1–22. doi: 10.1371/journal.pone.0251070

64
Hughes, M., & Lovell, T. (2019). Transition to attack in elite soccer. Journal of
Human Sport and Exercise, 14(1), 236–253. doi: 10.14198/jhse.2019.141.20
Intille, S. S., & Bobick, A. F. (1999). Framework for recognizing multi-agent
action from visual evidence. Proceedings of the National Conference on
Artificial Intelligence, 518–525.
Jordan, M. I., & Mitchell, T. M. (2015). Machine learning: Trends, perspectives,
and prospects. Science (Special Section on Artificial Intelligence), 349(6245),
255–260.
Karsten, B., Baker, J., Naclerio, F., Klose, A., Antonino, B., & Nimmerichter, A.
(2017). Validity of a Microsensor-Based Algorithm for Detecting Scrum
Events in Rugby Union. International Journal pf Sports Physiology and Per-
formance, 14(2), 156-162. Retrieved from https://www.cochranelibrary
.com/central/doi/10.1002/central/CN-01787161/full
Kelly, D., Coughlan, G. F., Green, B. S., & Caulfield, B. (2012). Automatic
detection of collisions in elite level rugby union using a wearable sensing
device. Sports Engineering, 15(2), 81–92. doi: 10.1007/s12283-012-0088-5
Kempe, M., Grunz, A., & Memmert, D. (2015). Detecting tactical patterns in
basketball: Comparison of merge self-organising maps and dynamic
controlled neural networks. European Journal of Sport Science, 15(4),
249–255. Retrieved from http://dx.doi.org/10.1080/17461391.2014
.933882 doi: 10.1080/17461391.2014.933882
Kempe, M., Vogelbein, M., Memmert, D., & Nopp, S. (2014). Possession vs.
Direct Play: Evaluating Tactical Behavior in Elite Soccer. International
Journal of Sports Science, 4(6A), 35–41. doi: 10.5923/s.sports.201401.05
Kipf, T., Fetaya, E., Wang, K. C., Welling, M., & Zemel, R. (2018). Neural
relational inference for Interacting systems. 35th International Conference
on Machine Learning, ICML 2018, 6, 4209–4225.
Knauf, K. (2014). Spatio-Temporal Convolution Kernels for Clustering Trajec-
tories. KDD Workshop on Large-Scale Sports Analytics, New York (USA).
Kolekar, M. H., Palaniappan, K., Sengupta, S., & Seetharaman, G. (2009).
Semantic concept mining based on hierarchical event detection for soc-
cer video indexing. Journal of Multimedia, 4(5), 298–312. doi: 10.4304/
jmm.4.5.298-312
Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012). ImageNet Classification
with Deep Convolutional Neural Networks. Advances in neural informa-
tion processing systems, 25, 1097–1105. doi: 10.1201/9781420010749
Kuper, S. (2018). Soccernomics: Why England Loses; Why Germany, Spain,
and France Win; and Why One Day Japan, Iraq, and the United States Will

65
 Become Kings of the World’s Most Popular Sport. Nation Books. Re-
trieved from https://books.google.com/books/about/Soccernomics
_2018_World_Cup_Edition.html?hl=de&id=OqS-swEACAAJ
Laird, P., & Waters, L. (2008). Eyewitness Recollection of Sport Coaches.
International Journal of Performance Analysis in Sport, 8(1), 76–84. doi:
10.1080/24748668.2008.11868424
Le, H. M., Yue, Y., Carr, P., & Lucey, P. (2017). Coordinated multi-agent
imitation learning. In 34th international conference on machine learning,
icml 2017 (Vol. 4, pp. 3140–3152).
Li, R., & Chellappa, R. (2010). Group motion segmentation using a spatio-
temporal driving force model. Proceedings of the IEEE Computer Society
Conference on Computer Vision and Pattern Recognition, 2038–2045. doi:
10.1109/CVPR.2010.5539880
Li, R., Chellappa, R., & Zhou, S. K. (2009). Learning multi-modal densities on
discriminative temporal interaction manifold for group activity recogni-
tion. 2009 IEEE Computer Society Conference on Computer Vision and Pat-
tern Recognition Workshops, CVPR Workshops 2009, 2009 IEEE, 2450–2457.
doi: 10.1109/CVPRW.2009.5206676
Link, D. (2018). Sports Analytics: How (commercial) sports data create new
opportunities for sports science. German Journal of Exercise and Sport
Research, 48(1), 13–25. doi: 10.1007/s12662-017-0487-7
Link, D., & Hoernig, M. (2017). Individual ball possession in soccer. PLoS
ONE, 12(7), 1–15. doi: 10.1371/journal.pone.0179953
Linke, D., Link, D., & Lames, M. (2018). Validation of electronic performance
and tracking systems EPTS under field conditions. PLoS ONE, 13(7),
1–20. doi: 10.1371/journal.pone.0199519
Linke, D., Link, D., & Lames, M. (2020). Football-specific validity of TRA-
CAB’s optical video tracking systems. PLoS ONE, 15(3), 1–17. doi:
10.1371/journal.pone.0230179
Low, B., Coutinho, D., Gonçalves, B., Rein, R., Memmert, D., & Sampaio,
J. (2020). A Systematic Review of Collective Tactical Behaviours in Football
Using Positional Data (Vol. 50) (No. 2). Springer International Publishing.
Retrieved from https://doi.org/10.1007/s40279-019-01194-7 doi:
10.1007/s40279-019-01194-7
Lucey, P., Bialkowski, A., Carr, P., Morgan, S., Matthews, I., & Sheikh, Y.
(2013). Representing and discovering adversarial team behaviors us-
ing player roles. Proceedings of the IEEE Computer Society Conference
on Computer Vision and Pattern Recognition, 2706–2713. doi: 10.1109/

66
 CVPR.2013.349
Lucey, P., Bialkowski, A., Monfort, M., Carr, P., & Matthews, I. (2014). "Quality
vs Quantity": Improved Shot Prediction in Soccer using Strategic Fea-
tures from Spatiotemporal Data. MIT Sloan Sports Analytics Conference,
Boston (USA), 1–9. Retrieved from http://www.sloansportsconference
.com/?p=15790
Lucey, P., Oliver, D., Carr, P., Roth, J., & Matthews, I. (2013). Assessing team
strategy using spatiotemporal data. Proceedings of the ACM SIGKDD In-
ternational Conference on Knowledge Discovery and Data Mining, Part F1288,
1366–1374. doi: 10.1145/2487575.2488191
MacLennan, T. (2005, 5). Moneyball: The Art of Winning an Un-
fair Game. The Journal of Popular Culture, 38(4), 780–781. Re-
trieved from https://onlinelibrary.wiley.com/doi/full/10.1111/
j.0022-3840.2005.140_11.xhttps://onlinelibrary.wiley.com/doi/
abs/10.1111/j.0022-3840.2005.140_11.xhttps://onlinelibrary
.wiley.com/doi/10.1111/j.0022-3840.2005.140_11.x doi:
10.1111/j.0022-3840.2005.140-11
Manafifard, M., Ebadi, H., & Abrishami Moghaddam, H. (2017). A survey on
player tracking in soccer videos. Computer Vision and Image Understand-
ing, 159, 19–46. doi: 10.1016/j.cviu.2017.02.002
Martens, F., Dick, U., & Brefeld, U. (2021). Space and control in soccer. Fron-
tiers in Sports and Active Learning (Special Issue: Using Artificial Intelligence
to Enhance Sport Performance), 3, 1–13. doi: 10.3389/fspor.2021.676179
Mcintyre, A., Brooks, J., Guttag, J., Wiens, J., & Arbor, A. (2016). Recognizing
and Analyzing Ball Screen Defense in the NBA Learning to Classify
Defensive Schemes. MIT Sloan Sports Analytics Conference, Boston (USA),
1–10.
McNeff, J. G. (2002, 3). The global positioning system. IEEE Transactions on
Microwave Theory and Techniques, 50(3), 645–652. doi: 10.1109/22.989949
Mcqueen, A., Wiens, J., & Guttag, J. (2014). Automatically Recogniz-
ing on-Ball Screens. MIT Sloan Sports Analytics Conference, Boston
(USA), 1–10. Retrieved from https://pdfs.semanticscholar.org/
6334/05a9abc6c9ac20045365a3f47192f8fa6544.pdf?_ga=2.111994950
.378455875.1556236357-933593754.1556236357
Mehrasa, N., Zhong, Y., Tung, F., Bornn, L., & Mori, G. (2018). Learning
Person Trajectory Representations for Team Activity Analysis. MIT Sloan
Sports Analytics Conference, Boston (USA), 1–8. Retrieved from http://
arxiv.org/abs/1706.00893

67
Mitrotasios, M., & Armatas, V. (2012). Analysis of goal scoring patterns in the
2012 European Football Championship. The Sport Journal(50), 1–9.
Mnih, V., Kavukcuoglu, K., Silver, D., Graves, A., Antonoglou, I., Wierstra,
D., & Riedmiller, M. (2013). Playing Atari with Deep Reinforcement
Learning. Arxiv Preprint, 1–9. Retrieved from http://arxiv.org/abs/
1312.5602
Mohri, M., Rostamizadeh, A., & Talwalkar, A. (2014). Foundations in Machine
learning. In Springerbriefs in computer science (Vol. 0, pp. 39–44).
Montoliu, R., Martín-Félez, R., Torres-Sospedra, J., & Martínez-Usó, A. (2015).
Team activity recognition in Association Football using a Bag-of-Words-
based method. Human Movement Science, 41, 165–178. doi: 10.1016/
j.humov.2015.03.007
Morgulev, E., Azar, O. H., & Lidor, R. (2018). Sports analytics and the big-
data era. International Journal of Data Science and Analytics, 5(4), 213–222.
Retrieved from https://doi.org/10.1007/s41060-017-0093-7 doi: 10
.1007/s41060-017-0093-7
Motoi, S., Misu, T., Nakada, Y., Yazaki, T., Kobayashi, G., Matsumoto, T., &
Yagi, N. (2012). Bayesian event detection for sport games with hidden
Markov model. Pattern Analysis and Applications, 15(1), 59–72. doi: 10
.1007/s10044-011-0238-6
Müller-Budack, E., Theiner, J., Rein, R., & Ewerth, R. (2019). "Does 4-4-
2 exist?" – An analytics approach to understand and classify football
team formations in single match situations. Proceedings Proceedings of the
2nd International Workshop on Multimedia Content Analysis in Sports (Nice,
France), MMSports ’(September), 25–33. doi: 10.1145/3347318.3355527
Najafabadi, M. M., Villanustre, F., Khoshgoftaar, T. M., Seliya, N., Wald, R.,
& Muharemagic, E. (2015). Deep learning applications and challenges
in big data analytics. Journal of Big Data, 2(1), 1–21. doi: 10.1186/s40537
-014-0007-7
Narizuka, T., & Yamazaki, Y. (2019). Clustering algorithm for formations in
football games. Scientific Reports, 9(1), 1–8. Retrieved from http://
dx.doi.org/10.1038/s41598-019-48623-1 doi: 10.1038/s41598-019
-48623-1
Niu, Z., Gao, X., & Tian, Q. (2012). Tactic analysis based on real-world ball
trajectory in soccer video. Pattern Recognition, 45(5), 1937–1947. Re-
trieved from http://dx.doi.org/10.1016/j.patcog.2011.10.023 doi:
10.1016/j.patcog.2011.10.023
Njororai, W. W. (2013). Analysis of goals scored in the 2010 world cup soccer

68
 tournament held in South Africa. Journal of Physical Education and Sport,
13(1), 6–13. doi: 10.7752/jpes.2013.01002
Otter, D. W., Medina, J. R., & Kalita, J. K. (2021). A Survey of the Usages
of Deep Learning for Natural Language Processing. IEEE Transactions
on Neural Networks and Learning Systems, 32(2), 604–624. doi: 10.1109/
TNNLS.2020.2979670
Panigrahi, S., Nanda, A., & Swarnkar, T. (2021). A Survey on Transfer
Learning. Smart Innovation, Systems and Technologies, 194, 781–789. doi:
10.1007/978-981-15-5971-6
Pappalardo, L., Cintia, P., Rossi, A., Massucco, E., Ferragina, P., Pedreschi,
D., & Giannotti, F. (2019). A public data set of spatio-temporal match
events in soccer competitions. Scientific Data, 6(1), 1–15. Retrieved from
http://dx.doi.org/10.1038/s41597-019-0247-7 doi: 10.1038/s41597
-019-0247-7
Perin, C., Vuillemot, R., & Fekete, J. D. (2013). SoccerStories: A kick-off for
visual soccer analysis. IEEE Transactions on Visualization and Computer
Graphics, 19(12), 2506–2515. doi: 10.1109/TVCG.2013.192
Perše, M., Kristan, M., Kovačič, S., Vučkovič, G., & Perš, J. (2009). A
trajectory-based analysis of coordinated team activity in a basketball
game. Computer Vision and Image Understanding, 113(5), 612–621. doi:
10.1016/j.cviu.2008.03.001
Perse, M., Kristan, M., Perš, J., & Kovacic, S. (2006). A Template-Based Multi-
Player Action Recognition of the Basketball Game. CVBASE ’06 - Pro-
ceedings of ECCV Workshop on Computer Vision, 71–82.
Pettersen, S. A., Johansen, D., Johansen, H., Berg-Johansen, V., Gaddam, V. R.,
Mortensen, A., . . . Halvorsen, P. (2014). Soccer video and player position
dataset. Proceedings of the 5th ACM Multimedia Systems Conference, MM-
Sys 2014 (Singapore, March 2014), 18–23. doi: 10.1145/2557642.2563677
Pfeiffer, M., & Perl, J. (2015). Analysis of tactical defensive behavior in team
handball by means of artificial neural networks. IFAC-PapersOnLine,
28(1), 784–785. doi: 10.1016/j.ifacol.2015.05.169
Pisner, D. A., & Schnyer, D. M. (2019). Support vector machine. Machine
Learning: Methods and Applications to Brain Disorders, 101–121. doi: 10
.1016/B978-0-12-815739-8.00006-7
Plummer, B. T. (2013). Analysis of Attacking Possessions Leading to a Goal
Attempt, and Goal Scoring Patterns within Men’s Elite Soccer. Journal
of Sports Science, 1(1), 1–038.
Pons, E., García-Calvo, T., Resta, R., Blanco, H., del Campo, R. L., García, J. D.,

69
 & Pulido, J. J. (2019). A comparison of a GPS device and a multi-camera
video technology during official soccer matches: Agreement between
systems. PLoS ONE, 14(8), 1–12. doi: 10.1371/journal.pone.0220729
Pouyanfar, S., & Chen, S.-C. (2016). Semantic Event Detection Using Ensemble
Deep Learning. IEEE International Symposium on Multimedia (ISM), San
Jose (USA), 203–208. doi: 10.1109/ism.2016.0048
Power, P., Hobbs, J., Ruiz, H., Wei, X., & Lucey, P. (2018). Mythbusting Set-
Pieces in Soccer. MIT Sloan Sports Analytics Conference, Boston, 102(2),
1–12.
Power, P., Ruiz, H., Wei, X., & Lucey, P. (2017). "Not all passes are created
equal:" Objectively measuring the risk and reward of passes in soccer
from tracking data. Proceedings of the ACM SIGKDD International Confer-
ence on Knowledge Discovery and Data Mining, Part F1296, 1605–1613. doi:
10.1145/3097983.3098051
Pulling, C. (2015). Long Corner Kicks in the English Premier League. Kinesi-
ology, 47(2), 193–201.
Ratner, A., Bach, S. H., Ehrenberg, H., Fries, J., Wu, S., & Ré, C. (2017). Snorkel:
Rapid training data creation with weak supervision. Proceedings of the
VLDB Endowment, 11(3), 269–282. doi: 10.14778/3157794.3157797
Reep, C., & Benjamin, B. (1968). Skill and Chance in Association Football
Author. Journal of the Royal Statistical Society, 131(4), 581–585. Retrieved
from https://www.jstor.org/stable/2343726?seq=1
Rein, R., & Memmert, D. (2016). Big data and tactical analysis in elite soccer:
future challenges and opportunities for sports science. SpringerPlus, 5(1).
doi: 10.1186/s40064-016-3108-2
Rein, R., Raabe, D., & Memmert, D. (2017). “Which pass is better?” Novel
approaches to assess passing effectiveness in elite soccer. Human Move-
ment Science, 55, 172–181. Retrieved from http://dx.doi.org/10.1016/
j.humov.2017.07.010 doi: 10.1016/j.humov.2017.07.010
Rennie, M. J., Kelly, S. J., Bush, S., Spurrs, R. W., Austin, D. J., & Watsford,
M. L. (2020). Phases of match-play in professional Australian Football:
Distribution of physical and technical performance. Journal of Sports
Sciences, 38(14), 1682–1689. Retrieved from https://doi.org/10.1080/
02640414.2020.1754726 doi: 10.1080/02640414.2020.1754726
Ric, A., Bradley, P., Shaw, L., Thies, H., Sumpter, D., López-felip, M. A., . . .
Bauer, P. (2021). Football Analytics 2021: The role of context in trans-
ferring analytics to the pitch. Barça sports analytics summit, Barcelona
(Spain).

70
Richly, K., Bothe, M., Rohloff, T., & Schwarz, C. (2016). Recognizing com-
pound events in spatio-temporal football data. IoTBD 2016 - Proceedings
of the International Conference on Internet of Things and Big Data(March
2018), 27–35. doi: 10.5220/0005877600270035
Robberechts, P., & Davis, J. (2020). How data availability affects the ability
to learn good xG models. Communications in Computer and Information
Science, Springer, Cham, 1324, 17–27. doi: 10.1007/978-3-030-64912-8
Sacha, D., Al-Masoudi, F., Stein, M., Schreck, T., Keim, D. A., Andrienko, G.,
& Janetzko, H. (2017). Dynamic Visual Abstraction of Soccer Movement.
Computer Graphics Forum, 36(3), 305–315. doi: 10.1111/cgf.13189
Sampaio, J., & Maças, V. (2012). Measuring tactical behaviour in football.
International Journal of Sports Medicine, 33(5), 395–401. doi: 10.1055/s
-0031-1301320
Samuel, A. L. (1959). Some Studies in Machine Learning. IBM Journal
of Research and Development, 3(3), 210–229. Retrieved from https://
ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=5392560
Santos, A. B., Theron, R., Losada, A., Sampaio, J. E., & Lago-Peñas, C. (2018).
Data-driven visual performance analysis in soccer: An exploratory pro-
totype. Frontiers in Psychology, 9, 1–12. Retrieved from https://www
.frontiersin.org/article/10.3389/fpsyg.2018.02416 doi: 10.3389/
fpsyg.2018.02416
Sarmento, H., Clemente, F. M., Araújo, D., Davids, K., McRobert, A., &
Figueiredo, A. (2018). What Performance Analysts Need to Know About
Research Trends in Association Football (2012–2016): A Systematic Re-
view. Sports Medicine, 48(4), 799–836. doi: 10.1007/s40279-017-0836-6
Sarmento, H., Marcelino, R., Anguera, M. T., CampaniÇo, J., Matos, N., &
LeitÃo, J. C. (2014). Match analysis in football: a systematic review.
Journal of Sports Sciences, 32(20), 1831–1843. doi: 10.1080/02640414.2014
.898852
Sathyamorthy, D., Shafii, S., Amin, Z. F. M., Jusoh, A., & Ali, S. Z. (2015).
Evaluation of the accuracy of global positioning system (GPS) speed
measurement via GPS simulation. Defence S and T Technical Bulletin,
8(2), 121–128.
Schmicker, R. H. (2013). An application of satscan to evaluate the spatial
distribution of corner kick goals in major league soccer. International
Journal of Computer Science in Sport, 12(2), 70–79.
Schmidt, A. (2012). Movement pattern recognition in basketball free-throw
shooting. Human Movement Science, 31(2), 360–382. doi: 10.1016/j.humov

71
 .2011.01.003
Schumaker, R. P., Solieman, O. K., & Chen, H. (2010). Sports Data Mining
(Integrated ed.; R. Sharda & S. Voß, Eds.). Springer US. Retrieved from
https://www.springer.com/de/book/9781441967299 doi: 10.1007/978
-1-4419-6730-5
Seidl, T. (2019). Radio-based Position Tracking in Sports—Validation, Pattern
Recognition and Performance Analysis. Dissertation thesis, TU München.
Shaw, L., & Glickman, M. (2019). Dynamic analysis of team strategy in pro-
fessional football. Barça sports analytics summit, Barcelona (Spain), 1–13.
Shaw, L., & Gopaladesikan, S. (2021). Routine inspection: A playbook for
corner kicks. MIT Sloan Sports Analytics Conference, Boston (USA).
Shinde, P. P., & Shah, S. (2018, 7). A Review of Machine Learning and Deep
Learning Applications. In Proceedings - 2018 4th international conference
on computing, communication control and automation, iccubea 2018. Institute
of Electrical and Electronics Engineers Inc. doi: 10.1109/ICCUBEA.2018
.8697857
Siddiquie, B., Yacoob, Y., & Davis, L. S. (2009). Recognizing Plays in
American Football Videos. Technical Report, University of Maryland,
1(1), 1–8. Retrieved from http://www.researchgate.net/publication/
228519111_Recognizing_Plays_in_American_Football_Videos
Sing, A., Thakur, N., & Sharma, A. (2016). A review of supervised machine
learning algorithms. International Conference on Computing for Sustainable
Global Development (INDIACom), 1310–1315. Retrieved from https://
ieeexplore.ieee.org/abstract/document/7724478
Spearman, W. (2018). Beyond Expected Goals. MIT Sloan Sports Analytics Con-
ference, Boston (USA), 1–17. Retrieved from https://www.researchgate
.net/publication/327139841
Spearman, W., Basye, A., Dick, G., Hotovy, R., & Pop, P. (2017).
Physics-Based Modeling of Pass Probabilities in Soccer. MIT
Sloan Sports Analytics Conference, Boston (USA), 1–14. Retrieved
from https://www.researchgate.net/profile/William-Spearman/
publication/315166647_Physics-Based_Modeling_of_Pass
_Probabilities_in_Soccer/links/58cbfca2aca272335513b33c/
Physics-Based-Modeling-of-Pass-Probabilities-in-Soccer.pdf
Stein, M., Janetzko, H., Seebacher, D., Jäger, A., Nagel, M., Hölsch, J., . . .
Grossniklaus, M. (2017). How to Make Sense of Team Sport Data: From
Acquisition to Data Modeling and Research Aspects. Data, 2(1), 2. doi:
10.3390/data2010002

72
Stein, M., Seebacher, D., Karge, T., Polk, T., Grossniklaus, M., & Keim, D. A.
(2019). From Movement to Events: Improving Soccer Match Annota-
tions. Lecture Notes in Computer Science (including subseries Lecture Notes
in Artificial Intelligence and Lecture Notes in Bioinformatics), 11295 LNCS,
130–142. doi: 10.1007/978-3-030-05710-7
Steiner, S., Rauh, S., Rumo, M., Sonderegger, K., & Seiler, R. (2019). Out-
playing opponents—a differential perspective on passes using position
data. German Journal of Exercise and Sport Research, 49(2), 140–149. doi:
10.1007/s12662-019-00579-0
Stephanos, D. K., Husari, G., Bennett, B. T., & Stephanos, E. (2021). Ma-
chine learning predictive analytics for player movement prediction in
NBA: Applications, opportunities, and challenges. Proceedings of the 2021
ACMSE Conference - ACMSE 2021: The Annual ACM Southeast Conference,
2–8. doi: 10.1145/3409334.3452064
Stöckl, M., Seidl, T., Marley, D., & Power, P. (2021). Making Offensive Play
Predictable - Using a Graph Convolutional Network to Understand De-
fensive Performance in Soccer. MIT Sloan Sports Analytics Conference,
Boston (USA).
Stracuzzi, D. J., Fern, A., Ali, K., Hess, R., Pinto, J., Li, N., . . . Shapiro, D.
(2011). An application of transfer to American football: From observa-
tion of raw video to control in a simulated environment. AI Magazine,
32(2), 107–125. doi: 10.1609/aimag.v32i2.2336
Sun, C., Karlsson, P., Wu, J., Tenenbaum, J. B., & Murphy, K. (2019). Stochastic
prediction of multi-agent interactions from partial observations. Seventh
International Conference on Learning Representations (ICLR), New Orleans
(USA), 1–15.
Taberner, M., O’Keefe, J., Flower, D., Phillips, J., Close, G., Cohen, D. D., . . .
Carling, C. (2020). Interchangeability of position tracking technologies;
can we merge the data? Science and Medicine in Football, 4(1), 76–81. Re-
trieved from https://doi.org/10.1080/24733938.2019.1634279 doi:
10.1080/24733938.2019.1634279
Taki, T., Hasegawa, J. i., & Fukumura, T. (1996). Development of mo-
tion analysis system for quantitative evaluation of teamwork in soccer
games. IEEE International Conference on Image Processing, 815–818. doi:
10.1109/icip.1996.560865
Tenga, A., Holme, I., Ronglan, L. T., & Bahr, R. (2010). Effect of playing
tactics on goal scoring in norwegian professional soccer. Journal of Sports
Sciences, 28(3), 237–244. doi: 10.1080/02640410903502774

73
Tian, C., De Silva, V., Caine, M., & Swanson, S. (2020). Use of machine learning
to automate the identification of basketball strategies using whole team
player tracking data. Applied Sciences (Switzerland), 10(1), 1–16. doi:
10.3390/app10010024
Tuyls, K., Omidshafiei, S., Muller, P., Wang, Z., Connor, J., Hennes, D., . . .
Hassabis, D. (2021). Game plan: What AI can do for football, and what
football can do for AI. Journal of Artificial Intelligence Research, 71(2020),
41–88. doi: 10.1613/JAIR.1.12505
Van Haaren, J., Dzyuba, V., Hannosset, S., & Davis, J. (2015). Automati-
cally discovering offensive patterns in soccer match data. International
Symposium on Intelligent Data Analysis, 9385, 286–297. doi: 10.1007/
978-3-319-24465-5
Van Haaren, J., Shitrit, H. B., Davis, J., & Fua, P. (2016). Analyzing vol-
leyball match data from the 2014 world championships using machine
learning techniques. Proceedings of the ACM SIGKDD International Con-
ference on Knowledge Discovery and Data Mining, 13-17, 627–634. doi:
10.1145/2939672.2939725
Vilar, L., Araújo, D., Davids, K., & Bar-Yam, Y. (2013). Science of win-
ning soccer: Emergent pattern-forming dynamics in association foot-
ball. Journal of Systems Science and Complexity, 26(1), 73–84. doi: 10.1007/
s11424-013-2286-z
Vilar, L., Araújo, D., Davids, K., & Travassos, B. (2012). Constraints on com-
petitive performance of attacker-defender dyads in team sports. Journal
of Sports Sciences, 30(5), 459–469. doi: 10.1080/02640414.2011.627942
Vogelbein, M., Nopp, S., & Hökelmann, A. (2014). Defensive transition in
soccer - are prompt possession regains a measure of success? A quan-
titative analysis of German Fußball-Bundesliga 2010/2011. Journal of
Sports Sciences, 32(11), 1076–1083. Retrieved from http://dx.doi.org/
10.1080/02640414.2013.879671 doi: 10.1080/02640414.2013.879671
Vračar, P., Štrumbelj, E., & Kononenko, I. (2016). Modeling basketball play-
by-play data. Expert Systems with Applications, 44, 58–66. doi: 10.1016/
j.eswa.2015.09.004
Wang, J., & Zhang, J. (2015). A win-win team formation problem based on the
negotiation. Engineering Applications of Artificial Intelligence, 44, 137–152.
Retrieved from http://dx.doi.org/10.1016/j.engappai.2015.06.001
doi: 10.1016/j.engappai.2015.06.001
Wang, K.-C., & Zemel, R. (2016). Classifying NBA Offensive Plays Using
Neural Networks. MIT Sloan Sports Analytics Conference, Boston (USA).

74
Wang, Q., Zhu, H., Hu, W., Shen, Z., & Yao, Y. (2015). Discerning tactical
patterns for professional soccer teams: An enhanced topic model with
applications. Proceedings of the ACM SIGKDD International Conference
on Knowledge Discovery and Data Mining, 2015-Augus(April 2016), 2197–
2206. doi: 10.1145/2783258.2788577
Wei, X., Sha, L., Lucey, P., Morgan, S., & Sridharan, S. (2013). Large-scale
analysis of formations in soccer. 2013 International Conference on Digital
Image Computing: Techniques and Applications, DICTA 2013. doi: 10.1109/
DICTA.2013.6691503
Wu, Y., Xie, X., Wang, J., Deng, D., Liang, H., Zhang, H., . . . Chen, W. (2019,
1). ForVizor: Visualizing Spatio-Temporal Team Formations in Soccer.
IEEE Transactions on Visualization and Computer Graphics, 25(1), 65–75.
doi: 10.1109/TVCG.2018.2865041
Yeh, R. A., Schwing, A. G., Huang, J., & Murphy, K. (2019). Diverse generation
for multi-agent sports games. Proceedings of the IEEE Computer Society
Conference on Computer Vision and Pattern Recognition, 2019-June, 4605–
4614. doi: 10.1109/CVPR.2019.00474
Zhang, Q., Zhang, M., Chen, T., Sun, Z., Ma, Y., & Yu, B. (2019). Recent
advances in convolutional neural network acceleration. Neurocomputing,
323, 37–51. doi: 10.1016/j.neucom.2018.09.038
Zheng, M., & Kudenko, D. (2010). Automated event recognition for football
commentary generation. International Journal of Gaming and Computer-
Mediated Simulations, 2(4), 67–84. doi: 10.4018/jgcms.2010100105
Zheng, S., Yue, Y., & Lucey, P. (2016). Generating long-term trajectories using
deep hierarchical networks. Advances in Neural Information Processing
Systems (NIPS), 1551–1559.

75
A Appendix—Study I: Constructing Spaces and Times
for Tactical Analysis in Football

In the following, we present Andrienko et al. (2019), an ac-

cepted manuscript of an article published by IEEE Transactions
on Visualization and Computer Graphics on April 1st 2021, avail-
able online: https://ieeexplore.ieee.org/document/8894420

76
THIS ARTICLE IS PUBLISHED IN IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS 1

Constructing Spaces and Times for Tactical

Analysis in Football
Gennady Andrienko, Natalia Andrienko, Gabriel Anzer, Pascal Bauer, Guido Budziak, Georg Fuchs, Dirk
Hecker, Hendrik Weber, and Stefan Wrobel

Abstract—A possible objective in analyzing trajectories of multiple simultaneously moving objects, such as football players during a
game, is to extract and understand the general patterns of coordinated movement in different classes of situations as they develop. For
achieving this objective, we propose an approach that includes a combination of query techniques for flexible selection of episodes of
situation development, a method for dynamic aggregation of data from selected groups of episodes, and a data structure for
representing the aggregates that enables their exploration and use in further analysis. The aggregation, which is meant to abstract
general movement patterns, involves construction of new time-homomorphic reference systems owing to iterative application of
aggregation operators to a sequence of data selections. As similar patterns may occur at different spatial locations, we also propose
constructing new spatial reference systems for aligning and matching movements irrespective of their absolute locations. The approach
was tested in application to tracking data from two Bundesliga games of the 2018/2019 season. It enabled detection of interesting and
meaningful general patterns of team behaviors in three classes of situations defined by football experts. The experts found the
approach and the underlying concepts worth implementing in tools for football analysts.

Index Terms—Visual analytics, movement data, coordinated movement, sport analytics, football, soccer.

1 I NTRODUCTION
Football (soccer) is an exciting sport that attracts millions analysis is still challenging. Data-driven tactical analysis
of players and billions of spectators worldwide. Wikipedia requires understanding of information hidden in large vol-
explains the basics as: “Football is a team sport played with a umes of game tracking data that include frequently sampled
spherical ball between two teams of eleven players. ... The game positions of players and the ball and numerous game events
is played on a rectangular field called a pitch with a goal at such as goal shots and goals, passes, tackles, possession
each end. The object of the game is to score by moving the ball changes, substitutions, fouls etc.
beyond the goal line into the opposing goal” [1]. Although it was Professional football attracts tremendous interest and
sufficient for G.Lineker to use just a single sentence to fully therefore is supported by industry and huge investments
define football as “Twenty-two men chase a ball for 90 minutes into infrastructure, players and coaches. Recent progress
and at the end, the Germans always win”, in reality football in football data collection, processing and analysis [2], [3]
is very complex. 22+ players, the ball and 3 referees move created new opportunities for providing data-driven in-
and act in coordination within the teams and in competition sights into the game and, eventually, supporting a variety
between the teams. The game is defined by voluminous of stakeholders including coaches, medical staff of clubs,
rules, characterized by complex interactions, and requires players, scouts, leagues, journalists and general public. Pro-
specific skills and sophisticated tactics. fessional clubs nowadays intensively hire data scientists and
Team managers (coaches) define team tactics and select a some major clubs already have their own data analysis
plan for each game that needs to be carefully implemented departments [4], [5]. Several companies develop software
by the players. For winning a game, a team needs (1) skilled for supporting data collection, processing and statistical
players in excellent physical conditions and (2) sophisticated analysis and provide services to clubs delivering data and
tactics intelligently defined by coaches, effectively taught to analysis results, including visualizations, which are mainly
players, trained, and carefully implemented in the game. of illustrative nature. Analytical visualizations and visual
While training is covered well by sport science, tactical analytics at large are still seeking their way to this domain.
This paper results from joint research and co-authorship
• Gennady Andrienko and Natalia Andrienko are with Fraunhofer IAIS and of a group involving visual analytics researchers, data sci-
City, University of London. E-mail: gennady.andrienko@iais.fraunhofer.de entists, and football experts. The research goal of the group
• Gabriel Anzer is with Sportec Solutions GmbH, Germany was to find approaches to extracting and understanding
• Pascal Bauer is with DFB Akademie, Germany
• Georg Fuchs and Dirk Hecker are with Fraunhofer IAIS, Germany
the general patterns of the team behaviors and dynamics
• Guido Budziak is with TU Eindhoven, The Netherlands of changes in relation to events and context in different
• Hendrik Weber is with DFL Deutsche Fussball Liga GmbH, Germany classes of game situations as these situations develop. In
• Stefan Wrobel is with Fraunhofer IAIS and University of Bonn, Germany the context of the paper, the term ’situation’ refers to a
© 2019 IEEE. Personal use of this material is permitted. Permission from IEEE combination of circumstances in which players behave, and
must be obtained for all other uses, in any current or future media, including the term ’episode’ refers to the situation development in
reprinting/republishing this material for advertising or promotional purposes,
creating new collective works, for resale or redistribution to servers or lists, or which the circumstances dynamically change. The overall
reuse of any copyrighted component of this work in other works.. goal involves the following sub-goals:
THIS ARTICLE IS PUBLISHED IN IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS 2

1) enable selection of groups of situations with particular 3. Pattern visualization. To enable perception and in-
characteristics and extraction of data pieces reflecting terpretation of collective behavior patterns by an analyst, a
the development of these situations; set of aggregates (i.e., pseudo-trajectories) generated from
2) derive general patterns of team behaviors from the selected episodes is represented visually. For comparative
extracted data pieces; analysis of sets of aggregates generated for different teams,
3) enable comparison of general patterns corresponding to kinds of situations, and games, the pseudo-trajectories are
different groups of situations. put in a common spatial domain (i.e., the pitch space, team
To achieve these sub-goals, our group has developed a space, or attribute space) and aligned with respect to their
framework including techniques for (1) query and data abstract times.
extraction (filtering), (2) integration of extracted data pieces While the framework makes use of previously existing
into aggregate structures, which may involve (2*) space techniques and approaches, it also incorporates novel ideas,
transformation, and (3) visualization of the aggregates for specifically:
interpretation, exploration, and comparison. These compo- • new primitives for temporal queries allowing specifica-
nents of the framework are briefly described below. tion of relative time intervals (Section 4.1);
1. Query and filtering. This component enables selection • a novel way of aggregating movement data that is
of time intervals containing game episodes with target char- suitable for bringing together temporally disjoint data
acteristics, which may refer to occurrences of specific game pieces (Section 4.3);
events (e.g., shots, passes in a given direction at a certain • a data structure for representing aggregated movement
distance, etc.) and attributes, such as speed or acceleration, data that allows the aggregates to be visualized and
of the ball, teams, and selected players. Once a set of target explored similarly to trajectories (Section 4.3.2);
intervals is selected according to event- or attribute-based • glyphs showing usual relative positions of players in
query conditions, it is possible to further select intervals their teams and providing hints at their roles (Sec-
positioned in time in a specific way in relation to the target tion 4.2 and Fig. 4).
intervals, e.g., starting at a given time distance before or We demonstrate the effectiveness of the proposed frame-
after the beginning or the end of each target interval and work in several case studies using real data from two
having a specified duration. This enables exploration of Bundesliga games [6], [7] of the season 2018-19.
what had happened before and after the target episodes and
The remainder of the paper has the following structure.
at different stages of their development.
Section 2 introduces the main concepts concerning football
2. Aggregation. Trajectory fragments extracted from the
tactics, describes the collection, contents, and structure of
selected intervals are integrated into aggregate structures,
data from a football game, and presents the research prob-
where each structure represents the behavior of one moving
lem we have addressed. After an overview of the related
object (a player, the ball, the mass center of a team, etc.) and
work (Section 3), we present our approach and components
consists of a sequence of generalized positions correspond-
of the analytical framework in Section 4 and describe how
ing to a sequence of interactive selections of time intervals.
we have applied it to three complex scenarios of team tactics
A generalized position is an aggregate of the positions of the
analysis (Section 5). Section 6 discusses the overall approach
object extracted from all selected intervals. It is represented
and outlines directions for further work.
by a central point, which may be the mean, median, or
medoid of the extracted subset of positions, and one or
more convex hulls covering chosen fractions (e.g., 50 and 2 BACKGROUND
75%) of this subset. Each sequence of generalized positions
is represented by a pseudo-trajectory in an abstract temporal 2.1 Football tactics in a nutshell
domain where the sequence of time stamps corresponds to Football tactics depends on multiple factors: which team
the sequence of selections. The resulting sets of pseudo- possesses the ball, in what part of the pitch the ball and
trajectories of all players and the ball provide a generalized the teams are located, and how the players are arranged
representation of collective behaviors in all situations with within their teams and in relation to the opponents. When a
particular properties. team possesses the ball, it aims at scoring a goal by offensive
2*. Space transformation. As an additional means of ab- actions, although in some rarely occurring cases (such as
straction and generalization, this component allows putting a lead close to the end of the game) it can be a teams’
together similar movements that might take place in differ- solely objective to stay in ball possession. When the ball
ent parts of the pitch. The core idea is to replace the positions is possessed by the opponents, a team aims at preventing
of the moving objects in the physical space (i.e., on the pitch) a goal and performs defense. There is an intermediate
by corresponding positions in artificially constructed spaces, turnover stage between offensive and defensive actions.
such as a team space, which represents the relative place- After winning the ball, a team can either counter-attack or
ments of the players within a team, or an abstract space with safeguard and build up. After losing the ball, a team can
dimensions corresponding to some attributes. Generalized either fall back and defend or perform counter-pressing.
positions and pseudo-trajectories can be constructed from In some situations, e.g. after fouls or if the ball goes out of
positions in artificial spaces in the same way as from the the pitch, the game is interrupted and resumes through set
original positions in the pitch space. The pitch space is good pieces, such as corners, free kicks, throw-ins, goalie kicks,
for analyzing the tactics of the team movement while the and penalties.
team space is good for seeing the relative arrangement of the Players in football teams have different roles. An es-
players and how it changes depending on circumstances. tablished term formation means a way how 10 outfield
THIS ARTICLE IS PUBLISHED IN IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS 3

players in a team generally position themselves relative data. Event data include the positions and times of the
to their teammates. Formations typically consist of three events and annotations, i.e., attributes describing the event
or four rows of players and are described, respectively, by types, involved players, outcomes, etc. The event extraction
three or four numbers specifying how many players are in and annotation is done partly manually, though there exist
each row from the most defensive to the most forward [8]. implementations that facilitate manual annotation using
For example, formation 4-3-3 means that the team has 4 machine learning approaches. Major companies doing data
defenders, 3 midfielders and 3 forwards, or strikers. In some acquisition and processing are ChyronHego [17], OPTA [18],
formations, intermediate lines appear denoting attacking or STATS [19], SecondSpectrum [20] and Track160 [21]. Smaller
defensive midfielders or so-called second forwards playing companies (e.g., FootoVision [22]) develop lightweight solu-
slightly behind their partner. tions for extracting data from a single video.
Formations usually differ significantly depending on the A typical data set for one game consist of general
ball possession, so that each team has an offensive formation information (date and location, playing teams, names of
and a defensive formation. Schematic figures in media usu- referees), information about the teams (list of players with
ally show only the offensive formations. First investigations their intended positions on the pitch, list of reserve players
on comparing offensive and defensive formations of the for substitutions), positional data (coordinates of the ball,
same teams were made in [9] where the average line-ups of players and, sometimes, referees in 2D x,y or 3D x,y,z space
two teams were shown both in and out of ball possession. with time references) and events (what happened, when and
When the ball possession changes, teams strive to arrange where, with event-specific characteristics). In average, about
themselves as fast as possible into the respective opposite 140,000 positions for the ball and each player are recorded
(offensive or defensive) formation. Generally, formations as during one game, roughly 3,500,000 positions in total. In
a major component of football tactics are carefully studied in addition to automatically recorded positions, about 1,500
literature [10]. There exist manuals for coaches (e.g. [11]) and events are annotated manually and then validated using
catalogues of offensive formations (e.g. [12]) enumerating computational methods. As any real-world data, football
possible attacking styles and suggesting efficient defense. data require assessment of data quality, plausibility check-
However, not only the chosen formations are important. ing, and evaluation of the coverage in space, time, and the
Football is a highly dynamic game where the players not set of moving objects [23]. Particularly, it is necessary to
just take fixed relative positions but constantly move in a make queries with allowing certain tolerance to potential
coordinated manner, which does not simply mean moving mismatch of times in positional and event data.
in parallel and thereby keeping the same arrangement. In addition to data collection, commercial companies
Both the arrangement of the players and their relative provide basic analytical and visualization services. A typical
movements depend on multiple factors, including which menu of provided visuals includes depiction of individual
team and for how long possesses the ball, where on the events (e.g., positions of fouls and tackles, geometries of
pitch the teams are located, what are the distances to the passes and shots) and aggregated representations of players’
opponents, what events happened recently, what is the positions on the pitch such as density heat maps. Both types
current score of the game, etc. For understanding teams’ of visuals can be filtered by players and times. However,
tactics and their efficiency, it is necessary to see the spatial possibilities for exploration by connecting different aspects
arrangements of the players and the character and dy- are not available.
namics of their changes in response to game events and
other circumstances. In today’s practice, this is a very time- 2.3 Problem statement
consuming process done largely by analysts watching game The formulation of the research goals comes from the foot-
videos and synthesizing information by reasoning. Several ball experts. Basically, their question was: How team tactics
recent research prototypes [13], [14], [15], [16] support this can be understood from data reflecting the movements of
activity by extracting formations and their changes from the the individual players and the ball (i.e., their trajectories)
data. However, changes of formations and, more generally, and the events that occurred during a game? All partners
changes of movement behaviors do not happen instantly. communicated to clarify the concept of team tactics and, on
What is still missing and challenging in supporting game this basis, define and refine the research goals.
analysis is a possibility to analyze the process of change in A team tactic can be defined as a general pattern of
the context of game events and situation characteristics. Our collective behavior in a group of situations with particular
paper intends to fill this very important gap. properties. This definition requires further clarification of
the concepts of situation properties, collective behavior, and
general pattern. Situation properties can be specified in
2.2 Data acquisition, content, and structure terms of various attributes: which team possesses the ball,
Today, detailed data are collected for almost every football how much time has elapsed since the possession change,
game in major professional leagues. Usually positional data which team is winning, where on the pitch is the ball and
are extracted from video recordings. For this purpose, sta- the majority of team players, etc. Collective behavior means
diums are equipped with stationary installations of multi- relative positions, movements, and actions of the players
ple cameras that record games from different viewpoints. with respect to their teammates and the opponents. A
Video analysis software is used for extracting time-stamped general pattern means a synoptic representation integrating
positions of the players, referees, and the ball from video multiple specific instances of collective behavior that were
footage, usually with a sampling rate of 10-25Hz. Addition- practiced in similar situations. The patterns differ depend-
ally, game events are extracted from video and positional ing on the situation properties. Hence, understanding of
THIS ARTICLE IS PUBLISHED IN IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS 4

team tactics requires consideration of groups of situations aware of four such tools: STATS Edge Viewer [37], parts
with different properties. of the SAP Sports One [38], Second Spectrum [20], and
Based on this refinement, the research goals presented in the online match analysis portal offered to the Bundesliga
the introduction section were formulated. clubs by Sportec Solutions [39]. The functionality provided
by these tools can be divided into three categories: calcula-
tion of various statistics, which are visualized in business
3 R ELATED WORK
graphics, search for specific game episodes in video records,
3.1 Major approaches to football analytics and replaying of selected episodes augmented with visual
Several groups of researchers managed to get access to game representation of calculated features, such as control zones
tracking data and developed interesting research proto- and pass opportunities. Hence, it is possible either to an-
types. Often the starting point was an adaptation of methods alyze the overall statistics at the level of a whole game
and tools developed for other purposes (e.g. animal tracking or to explore details of individual episodes. Our research
or transportation) for football data. A prominent example is fills the gap between these two extremes by developing
the famous Soccermatics book by D.Sumpter [4] that builds approaches to extracting general movement patterns from
on his research on collective animal behavior [24]. multiple episodes with some common properties. Impor-
A review [25] observes the state of the art, considering tantly, it is not limited to computing numeric statistics from
the following high-level tasks: playing area subdivision, selected parts of a game, but it produces more complex
network techniques for team performance analysis, specific spatio-temporal constructs representing movements.
performance metrics, and application of data mining meth-
ods for labelling events, predicting future event types and
3.2 Relevant visual analytics approaches beyond foot-
locations, identifying team formations, plays and tactical
ball
group movement, and temporally segmenting the game.
Some of the considered methods actively use visualization Different visual analytics approaches proposed for analyz-
components and thus fall into visual analytics (VA) ap- ing spatio-temporal and movement data [40] are relevant
proaches. Another review [9] takes a different perspective, to football analysis, although some of them have been
emphasizing the works with substantial involvement of vi- developed for specific application domains such as trans-
sualization and identifying the following major approaches: portation [41].
Analysis of game events. A representative example is Querying and filtering. The structure of movement data
SoccerStories [26], which summarizes game episodes using suggests possibilities for selection of subsets based on the
visual primitives for game events such as long ball, turning identities of the moving objects and their attributes, as well
the ball, cross, corner, shot etc. as dynamic data items including locations, times, and move-
Analysis of trajectories and trajectory attributes. A se- ment attributes, such as speed and direction [40]. A query
ries of works from the University of Konstanz proposed may involve a combination of multiple heterogeneous as-
methods for clustering trajectories of players during game pects; thus, Weaver [42] discusses interactive cross-filtering
episodes [27] and segmenting the game, finding interesting across multiple coordinated displays by direct manipulation
game situations [28], [29] and plays of particular config- in the displays. There exist special query devices for tempo-
urations [30], analyzing multiple attributes along trajecto- ral sequences of attribute values (e.g. TimeSearcher [43]) and
ries [31] and computing features of team coordination [32]. for sequences of events.
Analysis of team formations and derived features of The kind of analysis our group aimed to support requires
them. Several papers from Disney Research target at re- selection of groups of time intervals containing game episodes
constructing team formations and player roles from po- with particular characteristics. Database researchers long
sitional data. The proposed methods identify the role of ago proposed time query primitives [44], [45] suitable for
each player at each time moment allowing the analyst to such purposes. Recently, similar ideas were implemented
trace short- and longer-term roles and detect role swaps within an interactive visual analytics environment in a tool
between players. This approach allowed characterization of called TimeMask [46]. Our approach extends this work by
team styles in several games [13], [14]. After enumerating increasing query flexibility, see Section 4.1.
offensive and defensive configurations of players, paper [15] Transformation of space and time provide additional
evaluates pairwise success statistics. ForVizor [16] uses the perspectives for looking at movement data. It may be useful
dynamics of detected formations for segmenting the game. to treat selected pairs of numeric attributes as coordinates
Another approach for game segmentation is clustering of in an abstract space [47]. A polar coordinate system may be
time moments based on features reflecting relative positions used in such a space if the movement directions or cyclic
of players or other team configuration indicators [33]. time attributes are involved [48]. Research on group move-
Computation of football-specific constructs, such as in- ment [49] introduces the idea of a group space consisting
teraction spaces [31], and indicators such as scoring chances, of relative positions in respect to a central trajectory of the
pass options [34], [35], and pressure degrees [9], followed by group. This idea was successfully applied for analyzing the
visual representation of these in spatio-temporal displays. distribution of pressure over team formations in football [9].
An interesting development is including visualization of Analysis of multiple asynchronous trajectories can ben-
computed features directly in video frames [36]. efit from transforming the time stamps of the positions to
Apart from the research prototypes, there are also com- relative time references within relevant time cycles or within
mercial systems and services that support coaches and the individual lifetimes of the trajectories, which brings
match analysts in their work with positional data. We are them to a common temporal reference system and thus
THIS ARTICLE IS PUBLISHED IN IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS 5

supports comparisons and finding general patterns [50]. In The circumstances may include the ball status (in play or out
this work, the ideas of space and time transformation have of play) and possession (one of the two teams), the absolute
been further extended, taking into account the new ways of or relative spatial positions of the players and/or the ball,
time filtering, see Section 4.2. their movement characteristics, such as direction and speed,
Aggregation is one of the most important tools for the events that are happening currently or have happened
spatial abstraction and simplification of massive movement before, the relative time with respect to the game start and
data [51]. Review [25] suggests spatial aggregation over end, etc. These circumstances dynamically change during
Cartesian or polar grids or hand-designed polygons that re- the game. We refer to a sequence of changes happening
flect specific functions of pitch regions. Aggregation results during a continuous time interval as situation development
can be automatically re-calculated in response to changes of and to the corresponding time interval as an episode of
query conditions. situation development.
The existing approaches to aggregation of movement To explore and generalize the behaviors of the players
data produce the following major types of aggregates: and teams in certain situations, one needs to be able to select
density fields [9], [52], place-related attributes reflecting all episodes when such situations happened and developed.
various statistics of the appearances of moving objects in The selection requires appropriate query facilities for (A)
the places [53], flows between places [51], and a central specification of the situations in terms of the circumstances
trajectory of a set of similar trajectories [54]. The former involved and (B) specification of the relative time intervals
two types represent the presence of moving objects rather in which the situation development will be considered. For
than their movement while the latter two represent the example, the circumstances may be “team A gains the ball
movement of a whole group but not the movements of its possession when the ball is in the opponents’ half of the
members. In previous works, relative positions of group pitch”, and the relative time interval may be from one sec-
members were represented by density distributions [49] ond before to five seconds after the situation has arisen. An
or by their average positions [9], but their movements earlier proposed interactive query tool called TimeMask [46]
within the group were not reflected. Hence, none of the supports (A) but not (B). To support both, we propose the
previously existing aggregation methods is well-suited for following extended set of query primitives:
representing collective movement patterns. Therefore, our
(A) Specification of situations
group has devised a novel way of aggregation producing a
Result: set of target time intervals T1 , T2 , ..., TN , where
novel kind of aggregate – a set of pseudo-trajectories, see
Section 4.3. Aggregation operations are combined with time
Ti = [tstart
i , tend
i ]
filtering and time transformation and enable assessment of • Query conditions

variability within aggregates. – Attribute-based: selection of value intervals for

numeric attributes and particular values for cat-
3.3 What is missing egorical attributes
Among the earlier works dealing with collective movement, – Event-based: selection of particular event cate-
some focus on detection of occurrences of specific rela- gories
tionships between moving objects, such as close approach, • Condition modifier: logical NOT

others search for overall patterns of collective behavior. • Minimal duration of a situation

However, the former result in multiple disjoint data pieces (B) Specification of relative intervals
that do not make a general picture while the latter, on Result: set of relative time intervals R1 , R2 , ..., RN ,
the opposite, tend to overgeneralize by neglecting essen- where Ri = [ristart , riend ]
tial differences between situations. Football is a dynamic • Relative interval start ri
start
: reference (tstart
i or tend
i )
phenomenon with high variability of situations, therefore and time shift ± , i.e.,
it is necessary to understand the dynamics of patterns and ristart = tstart
i ± or ristart = tend
i ± .
differentiate individual and collective behaviors depending • Relative interval end ri
end
: reference (tstart
i or tend
i )
on the situational context. Hence, there is a need to de- and time shift ± , i.e.,
velop methods for identifying classes of situations, detecting riend = tstart
i ± or riend = tend i ± .
patterns of coordinated movement in subsets of similar • Relative interval duration D and reference ri
start
or
situations, and comparing patterns across different subsets. ri , i.e.,
end
Another important aspect is a possibility to see relation- riend = ristart + D or ristart = riend D.
ships between individual behaviors in the overall context of The primitives for relative interval specification allow
coordinated movement. There are two important aspects of this to be done in one of two ways: to set both the start and
these relationships: (1) the spatial arrangement of individ- the end (one of the time shifts or may be zero), or to set
uals within a group and (2) how the arrangement changes either the start or the end and the interval duration. Here
in response to different circumstances. Our work aimed at are examples of possible specifications of relative intervals
developing appropriate methods for satisfying these needs. with respect to a target T :
• select X sec after T:
4 A PPROACH
[Rstart , Rend ] [Tend , Tend + X]
4.1 Temporal queries for episode selection • select initial X sec of T:
We use the term situation to denote a particular combination [Rstart , Rend ] [Tstart , Tstart + X]
of circumstances that may take place in the course of a game. • add X sec before and Y sec after T:
THIS ARTICLE IS PUBLISHED IN IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS 6
1. Query: out of play excluded. Result: 102 episodes / 97159 frames

2. Query: only BVB ball possession; time zoom to 15 minutes. Result: 230 episodes / 43607 frames

Fig. 2. A schematic illustration of a transformation from the pitch space

3. Duration threshold: episodes shorter than 1 second ignored. Result: 206 episodes / 43291 frames (left) to the attacking team space (right). The coordinate grid in this and
all further team spaces has 5m resolution.

preceding the target situations (Fig.1.5).

Episode selection works as a temporal filter: only
4. Query: the ball in the attacking third of BVB. Result: 60 episodes / 6341 frames
data from currently selected intervals R1 , R2 , ..., RN (or
T1 , T2 , ..., TN if there is no relative interval specification) are
treated as “active”, being shown in visual displays and used
in computational operations. This filter may be combined
with various other filters applicable to movement data [40].

5. Relative intervals: add 1 second before the selected episodes. Result: 60 episodes / 7841 frames
4.2 Space transformation
The space transformation is based on determining the rel-
ative positions of points of all trajectories in respect to the
corresponding point of a chosen or constructed reference
trajectory and its movement vector [49]. Taking into account
Fig. 1. A sequence of temporal query operations. Yellow vertical stripes the nature of the football game, we usually assume the
mark time intervals selected by queries. Green is used for selected time movement vector to be perpendicular to the opponent’s
intervals after ignoring short intervals. Blue shows interval extensions
goal line, see Fig. 2. This choice can be modified by, for
due to modifiers.
example, treating differently situations when the players are
[Rstart , Rend ] [Tstart X, Tend + Y ] very close to one of the goals, when teams/players often
Figure 1 provides a visual illustration of a sequence of give up their preferred formations.
query operations for episode selection. Here and further on, A reference trajectory may be chosen or constructed in
we use data collected by a commercial service for the game different ways depending on the character of the collective
of Borussia Dortmund and FC Bayern München [6], further movement and analysis goals. For our goals, none of the
called by the abbreviations BVB and FCB, respectively. BVB existing individual trajectories can be used as an adequate
is usually shown in yellow and FCB in red . The data for representative of the movement of a team as a whole. In-
this game span roughly over two hours (2 half times of 45+ stead, we generate a central trajectory of a team by applying
minutes plus a break in between) with 25Hz resolution and an aggregation operator to the positions of all players of a
include about 170,000 frames in total. team, excluding the goalkeeper, at each time moment. The
In the images shown in Fig. 1, the horizontal dimension operator may be the mean, median, or medoid (the medoid
represents time. In the vertical dimension, the images are is the point having the smallest sum of distances to all
divided into sections. Each section shows the variation of others). We have extensively tested all three options using
values of an attribute or a sequence of events. Categorical data from several games and found that the best is to take
attributes are represented by segmented bars, the values the team’s mean after excluding the positions of two most
being encoded in segment colors. Numeric attributes are outlying field players, i.e., the most distant from the mean
represented by line charts. Events are represented by dots of the whole team, excluding the goalkeeper. The central
colored according to event categories. The yellow vertical position computed in this way changes smoothly over time,
stripes mark the target time intervals T1 , T2 , ..., TN selected while the medoid and median positions sometimes change
according to the current situation specification. The blue ver- abruptly, which leads to sharp kinks in the resulting central
tical stripes mark the relative time intervals R1 , R2 , ..., RN . trajectory. Depending on the analysis task, it may be useful
The stripes are semi-transparent; so, the greenish color (a to compute the central trajectories of subgroups of players,
mixture of yellow and blue) appears where relative intervals such as the defenders or midfielders, and investigate the
overlap with target intervals. behaviors of these subgroups.
The following sequence of query operations is shown: Figure 3 demonstrates how movements on the pitch
exclude the periods when the ball was out of play (Fig.1.1); translate to movements in a team space. It presents a short
select the episodes of BVB ball possession (Fig.1.2); exclude episode of a single attack of FCB on the pitch and in the
the episodes when BVB possessed the ball for less than 1 BVB team space. The movements of the players and the ball
second (Fig.1.3); select the episodes in which the ball was in are represented by lines, and the tiny square symbols at the
the attacking third of BVB (Fig.1.4); add 1-second intervals line ends show the positions at the end of the selected time
THIS ARTICLE IS PUBLISHED IN IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS 7

Fig. 3. One attack on the pitch (left) and in the BVB team space (right). In all illustrations, the rendering opacity varies along trajectories so that
earlier segments are more transparent.

interval. The goalkeeper’s trajectory is marked in black. The glyphs is optional. They were intensively used in our cases
pitch map shows mostly parallel movements of all players studies but are rarely present in the illustrations due to the
towards the goal of BVB, but the player 22 of FCB, who length restrictions.
initially ran in the direction of the goal, made a sharp turn There is a possible additional use of the presence statis-
to the right, and the players of BVB who were close to tics by zones. The distance of a player to his average position
the goal or to the player 22 made similar movements. In in the team space can be treated as an indicator of how usual
the team space, the trajectories of 6 defence players of BVB his current position is. These measures can be aggregated
located in the lower part of the team space are very short, over the whole team or a selected group of players (e.g.
which means high synchronization between them. In the the defence line). Based on the aggregated time series, it
upper part, the shapes of the trajectories belonging to the is possible to set query conditions for selecting episodes
remaining 4 players of BVB and 5 players of FCB, show that of unusual or usual team arrangements (Section 4.1). On
the distance of these players from the team center originally the other hand, the presence statistics can be computed for
increased (which means that they moved slower than the different groups of selected episodes, and it is possible to
defenders) and then decreased (as the backward movement choose which set of statistics to use for currently shown
of the defenders slowed down). The diagonal orientation of position glyphs and distance-to-usual-position calculations.
these trajectories corresponds to the movement of the team
center first to the left and then to the right.
4.3 Aggregation
While transformations to the team spaces are primarily
meant for studying relative arrangements and movements We propose a novel method for aggregating movements
of players, they create a useful by-product. By dividing a of an entity under different conditions. The output is a
team space into meaningful zones and aggregating players’ sequence of generalized positions organized in a pseudo-
duration of presence in these zones, we obtain “finger- trajectory along an abstract timeline. Here we describe how
prints” of players’ typical positions in the teams, which we construct the positions and times of pseudo-trajectories.
correspond to their roles in the game. We apply a division
into a central zone (10m around the team center) and 8 areas 4.3.1 Obtaining generalized positions
around it. The fingerprints can be represented by position Aggregation is applied to positions selected by the current
glyphs, as in Fig. 4. Thus, N.Süle in Fig.4, left, was present combination of data filters, including the episode selection
mostly on the back-left and back-center and sometimes in (Section 4.1). It can be done in the pitch space and/or in the
the central zone. Such glyphs facilitate identification of play- team spaces. The currently selected subset of positions of an
ers and spotting their appearances in unusual positions and entity is represented by a generalized position, which can
position swapping. Lines below some glyphs indicate the be the mean, median, or medoid of the selected subset. The
times the players were on the pitch (when it was not the full possibility to switch between these three options can serve
game) to give an idea of what changed after substitutions. as a means for checking the position variability. When the
Thus, N.Süle was a one-to-one substitution, which means variability is small (i.e., the points are compactly clustered),
that he exactly took M.Hummels’ position after getting sub- the mean, median, and medoid positions are very close to
stituted for him. S.Wagner and R.Sanches were 1:1 substitu- each other, and switching between them does not change the
tions to T.Müller and S.Gnabry, respectively, but interpreted general movement pattern obtained through the aggrega-
their roles differently. This is visible from their aggregated tion. Noticeable changes indicate presence of outliers. In this
positions in the team space and different distributions of the case, the analyst may look at the whole set of the original
presence in their fingerprint glyphs. The display of position points and decide whether the outliers can be ignored.
THIS ARTICLE IS PUBLISHED IN IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS 8

Fig. 4. Left: examples of position glyphs for N.Süle (below) who substituted M.Hummels (above). Right: glyphs shown at the average positions of
the FCB players in their team space. The blue dot represents the average position of the ball.

This is possible when the outliers are few, their positions ter with conditions ballInP lay = true & ballP ossession =
are randomly scattered, and the remaining points make a BV B , the second one differs by ballP ossession = F CB .
compact cluster. If the outliers are not negligible, they need To support the comparison, the corresponding points are
to be examined in detail. To see the pattern formed by the connected by lines. The second points (FCB possession)
remaining data, the analyst may switch to using the medoid, are marked by dots. The differences and commonalities of
which is insensitive to outliers (but very costly to compute). players’ arrangement depending on the ball possession can
There is no one-size-fits-all rule for including or ex- be easily seen. On the pitch, both teams move a bit towards
cluding outliers, but the decision may depend on specific the goal line. The most prominent behavioral differences
contexts, such as temporary positional changes, different happen with the wing defenders: they move wide under
roles of the players in set plays, a change of the tactical their own team’s ball possession and narrow under the
system over time, etc. It is crucial that the soccer expert opponent’s possession. In the team space, we see that the
(match analyst, coach, assistant coach) is able to decide those team without the ball gets more compact in both dimensions
things situation-specific and per interrogation of his/her (all players tend to move towards the team center), while the
own. Therefore, the user should be given a high degree of team with the ball gets wider. The defenders of the attacking
flexibility for investigations and the opportunities to exclude team move slower than the other players and thus increase
or include outliers and to change between more and less team’s covered area.
outlier-sensitive aggregates. Another example is shown in the right parts of the Figs. 5
To represent the variation among the positions explicitly, and 6. In the first half of the game, FCB scored a single
we propose to build convex hulls outlining chosen percent- goal at minute 26. We compare the mean positions of the
ages (e.g., 50 and 75%) of the set of original points ordered players and the ball under the BVB possession before and
by their distances to the representative point. Examples can after the goal (the latter are marked by dots), excluding
be seen in Fig.7 and Fig.9. Hence, a generalized position the times when the ball was out of play. The pitch map
of an entity is a combination of a representative point and shows us that both teams have shifted towards the FCB
one or more variation hulls. Such a position is constructed goal and a bit to the right. In the BVB team space, we
for each entity being currently under analysis. The visual see that all BVB players shifted synchronously except two
representation of the variation hulls can be controlled inde- central defenders (#2 and #16) who moved about 3m back
pendently of that of the points; in particular, the hulls can from the team center. This increased the distance between
be temporarily hidden for reducing the display clutter. the lines in the team and thus created opportunities for FCB
When the filter conditions change, aggregation can be counter-attacks. This fact may explain why the FCB player
applied to the new subset of selected positions, which #9 moved about 5m aside in the BVB team space, searching
produces a new set of generalized positions of all entities. for attacking opportunities.
The new generalized positions can be visualized together According to the football experts, such representations
with the previous ones for comparing, as shown in Figs. 5 can be very valuable to coaches by giving an overview
and 6. Figure 5 shows the aggregates in the pitch space, and of behavior changes after any significant game event. For
Figure 6 shows the corresponding aggregates in the team seeing more than a single change and, in general, comparing
space of BVB. In the left parts of the two figures, the first generalized positions in more than two sets of situations, we
aggregation operation was performed using the episode fil- construct abstract timelines for organizing multiple general-
THIS ARTICLE IS PUBLISHED IN IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS 9

Fig. 5. Comparison of the generalized positions on the pitch in different groups of situations. Left: under the ball possession by the different teams;
right: under the BVB possession before and after the goal in the first half of the game.

Thus, the example in Fig. 7 demonstrates pseudo-trajectories

of the players and the ball obtained by queries concerning
the position of the BVB team center on the pitch. We
made a sequence of 10 queries with a common condition
ballP ossession = BV B and the differing conditions refer-
ring to the position of the BVB team center along the X-
axis with respect to the pitch center: x < 40m, 40m 
x < 30m, ..., x 40m. The queries did not include
explicit time constraints, but each query selected a set of
time intervals when the x-coordinate of the team center was
in a specific range.
Two upper images in Fig. 7 show the footprints of the
pseudo-trajectories as lines on the pitch (left) and in the
BVB team space (right). On top of the lines correspond-
ing to the players, the position glyphs of the players are
shown. The glyphs are drawn at the middle positions of
the players’ pseudo-trajectories, which correspond to the
BVB team center position being in the interval [ 10..0)
meters. The relative arrangement of the glyphs of each team
Fig. 6. Comparison of the generalized positions in the BVB team space
for the same groups of situations as in Fig. 5.
reflect the formation used by BVB for preparing an attack
and the defensive formation 4-4-2 of FCB. We can also see
ized positions in pseudo-trajectories. consistent monotonous changes of the player’s positions
from left to right and changes of the teams’ widths along the
4.3.2 Creating virtual times pitch. Complementary to this, the team space demonstrates
changes in the team compactness in both dimensions.
A pseudo-trajectory consists of two or more linearly ordered
generalized positions. A pseudo-trajectory is represented For illustrative purposes, the remaining images in Fig. 7
on a map by a line obtained by connecting consecutive include the 50% variation hulls for selected 5 players of BVB.
positions (more precisely, their representative points). Each The hulls are shown in the pitch space and the BVB team
position of a pseudo-trajectory has an abstract numeric space in 2D maps and 3D space-time cubes. The images
timestamp (1, 2, 3, ...) that equals the ordinal number of the of the space-time cubes are provided for merely illustrative
position. Hence, a pseudo-trajectory has its internal abstract purposes, to demonstrate that the pseudo-trajectories and
timeline made by the sequence of the position timestamps. the hulls are spatio-temporal objects, albeit the time domain
Pseudo-trajectories are generated by successively apply- in which they exist is abstract rather than real. These ob-
ing several query + aggregation operations. Each opera- jects can be treated in the same ways as “normal” spatio-
tion generates one position, which is appended after the temporal objects existing in real time domains.
previously generated position, if any. Hence, the order of The colors of the variation hulls from violet through
the positions in pseudo-trajectories reflects the order of the yellow to orange depict the virtual times of the positions.
query + aggregation operations by which they have been We can observe stable shapes and sizes of the hulls and their
obtained. This very simple idea provides high flexibility very stable locations in the team space, except for the two
for creating position sequences with different semantics, as wing defenders who first moved outwards from the center
demonstrated by an example below and further on in the and then back towards the center. Generally, such stability
case studies. checks are necessary for all aggregates, and they were
The queries used for generating pseudo-trajectories may consistently performed during the analysis. However, the
include conditions of any kind, not necessarily time-based. page limit does not permit having many such illustrations
THIS ARTICLE IS PUBLISHED IN IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS 10

in the paper. (A), the experts described the situations that were interest-
ing for them, and the analysts translated the descriptions
into queries, extracted the corresponding portions of the
4.4 Interaction between the framework components
data, and provided the experts with tools for sampling some
Since our paper aims at presenting the general frame- of the selected situations and reviewing them with the use
work, which can be implemented in different ways, rather of animated maps and corresponding fragments of the game
than a specific implementation, we refrain from describing video. To assess the situation generalization facilities (B), the
techniques for user-computer interaction, which can differ experts were provided with visual displays of the pseudo-
between possible implementations. What we describe here trajectories, which represented the extracted patterns.
is how the components of the general framework are sup- The evaluation was carried out in a series of case studies,
posed to work together within the analysis process. in which the experts set the analysis tasks, the analysts
The process begins with creation of the team spaces (Sec- performed operations according to the framework, and the
tion 4.2). Then, the following sequence of steps is repeatedly football experts interpreted and evaluated the results and
executed: posed further questions.
1. Temporal query (Section 4.1):
1.1. Specify and find situations of interest (Section 4.1(A)). 5 C ASE STUDIES
1.2. Specify a sequence of relative intervals for extracting
the situation development episodes (Section 4.1(B)). The objective of professional football is to win matches and
entertain the public. To win matches, you have to score more
Result: set of episodes.
goals than the opponent. This requires a good balance be-
2. Aggregation (Section 4.3):
tween the offensive and defensive strategies of a team. That
2.1. Automatically aggregate the query result and gener- is why it is important that a team has a tactical plan defining
ate pseudo-trajectories in the pitch space and in the the desired team formations and behavior in different states
team spaces. of the game. Examples of states are own and opponent’s
2.2. Put the pseudo-trajectories as new information layers ball possession, transitions between them after loosing the
in the respective spaces. ball or recovering it, counter attacks, set pieces, such as
Result: set of pseudo-trajectories. corners etc. In the following, we consider several categories
3. Visualization and comparative analysis: of situations that were interesting to analyze for our football
3.1. Represent the pseudo-trajectories within each space experts. We use data from two Bundesliga games [6], [7]
in an interactive visual display using techniques suit- of the season 2018-19. The data from each game consist of
able for ordinary trajectories. The displays need to two parts: (1) trajectories of the players and the ball, i.e.,
be linked through common visual encoding and by sequences of their positions in the pitch recorded every 40
interactive techniques, such as brushing [55]. milliseconds, and (2) records of the game events with their
3.2. Compare the new set of pseudo-trajectories with one attributes, including the event type, time of occurrence, and
or more of the previously obtained sets resulting from players involved; see Section 2.2.
other temporal queries.
Our group has performed this analytical process in the case 5.1 Ball possession change
studies described in Section 5.
The two switching moments between own and opponent’s
possession are getting more and more attention in football.
4.5 Evaluation of the framework The reason is that the desired field occupancy of the players
In our research project, it was not intended to design and team tactics in situations where the team possesses the
and develop software tools according to specific users’ ball is completely different from the ideal field occupation
tasks and requirements. For developing and testing the and tactics in situations in which the opponent has the ball.
components of the framework, the partners specializing As soon as a team loses the ball, the field occupation is often
in visual analytics and data science, further referred to disorganized from a defensive perspective. It takes time for
as the analysts, utilized existing software tools. Among a team to adapt to the new situation (‘switching cost’) in
others, they used a research-oriented software system V- which it has to apply its defensive tactics. This temporary
Analytics (http://geoanalytics.net/V-Analytics/). The ana- ‘chaos’ is something the team with ball possession can take
lysts extended its base functionality by implementing new advantage of.
query, aggregation, and data transformation techniques.
These software developments were necessary for achieving 5.1.1 Checking the five-seconds rule
the research goals, but the key result of the project is the The famous P.Guardiola’s 5 seconds rule for successful
analysis methods and not the tools. counterpressing says [56]: “After losing the ball, the team has
In the course of the research, the methods under devel- five seconds to retrieve the ball, or, if unsuccessful, tactically foul
opment were constantly evaluated by the football domain their opponent and fall back”, that has been key to Manch-
experts according to the following criteria: (A) the possibil- ester City conceding considerably fewer goals. The football
ity to select multiple situations with common properties and experts were interested to see if BVB and FCB applied this
the flexibility in specifying the properties of interest; (B) the tactics in the game [6].
possibility to extract, visualize, and interpret general pat- To extract the episodes of interest, the analysts applied
terns of situation development. To assess the query facilities temporal query operations described in Section 4.1. They
THIS ARTICLE IS PUBLISHED IN IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS 11

Pseudo-trajectories of all players and the ball

Pseudo-trajectories of 5 selected BVB players with 50% variation hulls

The same pseudo-trajectories and hulls in the BVB team space

Fig. 7. Sequences of 10 generalized positions of the players and the ball under the BVB ball possession corresponding to different positions of the
BVB’s team center on the pitch along the X-axis.
THIS ARTICLE IS PUBLISHED IN IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS 12

Fig. 8. Pseudo-trajectories of the players during 5+5 seconds after

loosing the ball. Left: BVB, blue for the starting 5 sec and yellow for
the following. Right: FCB, in cyan and red, respectively.

first selected the moments of change in the ball possession,

excluding those when the ball got out of play. Next, they Fig. 9. The same aggregated data as in Fig.8 are presented in the team
spaces (left: BVB, right: FCB) using the same colors together with the
consecutively applied a series of operations for specification
50% variation hulls.
of relative intervals:
[Rstart , Rend ] [Tend + Xsec, Tend + (X + 1) sec],
game 3:2 FC Bayern [6], they want to compare BVB’s be-
X = 0, 1, 2, ..., 9
havior with their 7:0 game against FC Nürnberg (FCN) [7].
producing 10-steps long pseudo-trajectories from the mean
Following the procedure described in Section 5.1.1, in each
positions and their 50% variation hulls. During the aggre-
game for each player the analysts constructed two 10-steps
gation, irrelevant parts of the original trajectories that were
long pseudo-trajectories summarizing the transitions to the
shorter than 10 seconds (e.g. due to the ball going out of
attacking and defensive formations (Figs. 10 and 11).
play or another change of possession) were discarded by an
The images on the top of Fig. 10 and on the left of Fig. 11
attribute-based query.
correspond to the game against FC Nürnberg, the other
Figure 8 shows the results of the aggregation separately
two images correspond to the game against FCB considered
for BVB (left) and FCB (right). The pseudo-trajectories of the
earlier. The BVB team tactics when loosing and regaining
players are painted in two contrasting colors corresponding
the ball are represented in black and yellow, respectively.
to the first 5 seconds (blue for BVB and cyan for FCB) and to
The pitch images show that, while in the game against FCB
the following 5 seconds (yellow for BVB and red for FCB).
(right) BVB players pressured for 5 seconds after loosing
The images show that almost all BVB players do not move
the ball, in the game against FCN they were falling back
back during the first 5 seconds and gradually fall back in
during the initial 5 seconds and only then put pressure on
the next 5 seconds. This agrees with the fact that BVB is
the opponents. The team was narrowing down in the FCB
known for their pressing style of playing. The patterns of
game but kept constant width in the FCN game. After ball
the FCB players are different. The players continue moving
regaining, BVB players attempted to perform fast counter
forward for the initial 2 seconds on the average and then
attacks against FCB but preferred careful, slowly moving
start moving to the left and back.
forward attack preparations against FCN.
Figure 9 shows the players’ pseudo-trajectories and 50%
The team space images are useful for seeing the synchro-
variation hulls in the team spaces. The colors of the hulls
nization among the players and changes of their arrange-
encode their relative times. For BVB, the hull colors vary
ment. Game against FCN (left): synchronous movement in
from dark blue to dark yellow, so that the shades of blue
the transition to defense, except the central forwards, and
correspond to the first 5 seconds and the shades of yellow to
widening of the team in the transition to offense. Game
the following 5 seconds. For FCB, the hull colors vary from
against FCB: increasing compactness of the team for defense
dark cyan to dark red, respectively. The images show that
and fast expansion in both direction for offense. The team
BVB tended to reduce the team width and depth whereas
depth in the game against FCN was about 30m in contrast
FCB kept the width constant while slightly reducing the
to about 40m against FCB.
depth. The stacks of the hulls with the colors representing
the relative times show that the hulls of the majority of
the players were getting notably smaller over time. This 5.2 Long passes
means that the players were successfully reconstructing One of the most important instruments for advancing the
their planned defensive formations and then were keeping attack, finding open spaces on the pitch, and forcing op-
their intended relative positions. This is in accordance to the ponents to make mistakes, is long passes. After examining
general football philosophy that suggests creativity while the distribution of the pass lengths with respect to the X-
attacking and organization and structure while defending. axis (i.e., how much the ball moved in the direction to the
opponents’ goal) in the game [6], the experts and analysts
5.1.2 Comparison of behaviors in two games jointly chose a query threshold of minimum 20m for select-
For experts it is important to understand how a team adapts ing long passes. After inspecting the selected passes, the
to different opponents and circumstances. To contrast the experts understood that the passes made from the own goal
THIS ARTICLE IS PUBLISHED IN IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS 13

Borussia Dortmund against FC Nürnberg: involved in 3 out of 5 attacks that resulted in goal scoring. To
consider them separately, the analysts added spatial filters
by the pass destinations and obtained aggregates for the
subset of the passes (Fig. 13).
It is interesting to observe that, although the pass targets
were quite widely distributed on the pitch, they were com-
pact in the spaces of the defending teams. All BVB passes
were targeted in the area behind the FCB’s right central
defender J.Boateng, on the average 5 meters behind and 10
meters aside of him. He had to move back during these
passes, breaking the last defensive line. The FCB passes
targeted at a point about 25-30 meters aside of the BVB team
center. These passes forced the defending team to shift left.
It can be concluded that the long forward passes of
BVB were intended to make immediate danger to the goal.
Borussia Dortmund against FC Bayern Munich: The attacking group of the BVB players moved far forward
during these passes. The shape of the team became long but
rather narrow. The long passes of FCB resulted in changes
of the attacking direction with the players moving to the
right. It should be noted that FCB striker R.Lewandowski
was balancing around the offside line at the moment of
the reception of the selected long passes, so it would be
dangerous to pass to him immediately.

5.3 Building up for shots

Goals are scored after successful shots, which require not
only high individual skills of a striker but also work of the
whole team for reaching situations in which good shots
become possible. To help the experts to investigate this
teamwork in the BVB-FCB game, the analysts used queries
to select, first, the moments of the shots and, second, the
Fig. 10. Comparison of movements of Borussia Dortmund players in ball episodes preceding them. They made a series of queries
possession transition periods in games against FC Nürnberg [7] (top)
and FC Bayern Munich [6] (bottom). Only players who were present on
[Rstart , Rend ] [Tstart X sec, Tstart (X 1) sec],
the pitch for at least 30 minutes are shown. Black lines show transi- X = 10, 9, 8, ..., 1, where Tstart is the moment of a shot,
tions to defense: from -1 to +10 seconds after loosing the ball control; and obtained the corresponding sets of pseudo-trajectories
yellow lines represent transitions to offense: from -1 to +10 seconds of the ball, team centers, and all players separately for the
after gaining the ball. Different tactical patterns appear prominently (see
explanations in section 5.1.2 for details). shorts of BVB and FCB. As a measure of position variation,
the analysts also computed the median distances of the
representative points of the generalized positions to the
box need to be excluded. Figure 12 shows the footprints original positions from which they had been derived. Since
of the remaining selected long balls. To produce them, the changes of the ball possession could happen during the
analysts used two queries based on the starting and ending 10-seconds intervals before the shots, the analysts applied
moments of the selected passes: attribute filters discarding irrelevant parts of the episodes in
[Rstart , Rend ] [Tstart , Tstart + ] and which the build up was shorter than 10 seconds.
[Rstart , Rend ] [Tend , Tend + ], The pseudo-trajectories of the players and the ball are
where = 0.2sec was applied for tolerating possible mis- shown in Fig. 14 and the pseudo-trajectories of the team
matches in the times between the position data and manu- centers in Fig. 15. These two figures demonstrate different
ally annotated passes. To include more information about levels of aggregation and abstraction applied to the same
the conditions in which these long balls were made, the data. The position variation indicators associated with the
analysts visualized the pass lines together with aggregated points of the pseudo-trajectories are represented by propor-
positions of all players (colored dots) on top of the density tional sizes of the dots thus marking 1-second segments.
field summarizing the distributions of all players on the Please note that this is an abstract, symbolic representation
pitch during the execution of the selected passes. using the visual variable ’size’ to encode numeric values.
An interesting diagonal configuration of the players The sizes of the symbols are not related to the map scale.
during the long balls of BVB can be seen in Fig. 12, top. This representation is essentially different from the repre-
The passes were sent either along or across this dense sentation of the variation by hulls, which are spatial objects.
concentration. Most of the long passes of both teams were Unlike the dot sizes, the hulls occupy particular areas in
directed to the left flank of BVB and right flank of FCB, so space and have particular shapes.
the same side of the pitch was used actively by both teams. The individual aggregates in Fig. 14 and team aggregates
These passes were very important in this game as they were in Fig. 15 consistently show that the two teams tended to use
THIS ARTICLE IS PUBLISHED IN IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS 14

Fig. 11. Aggregated movements of BVB players in their team space after ball possession changes in games [7] (left) and [6] (right).

Long passes of BVB: pitch and FCB team space details for smaller subsets of similar shots. One option was
grouping by the shot location. However, by inspecting the
episodes preceding the shots, it was found that the variation
of the shot positions does not match the variation of the
trajectories of the ball, teams, and players. Similar build-ups
do not necessarily lead to making shots from similar posi-
tions. Another option was to cluster the shots by similarity
of the last preceding passes or by similarity of particular
trajectories, e.g., of the ball and/or the team centers. We
evaluated several variants of grouping using clustering of
Long passes of FCB: pitch and BVB team space trajectories by relevant parts [54]. They produced either
heterogeneous clusters with too small differences between
them or homogeneous clusters that were too small for
valid generalization. This procedure, however, appears to
have a good potential when applied to a larger number of
situations extracted from multiple games of the same team.

5.4 Conclusions from the use cases

Even if the top leagues and clubs are aware of the necessity
of acquiring event and tracking data, the potential of using
them in the team’s daily business is not yet tapped. Since the
game is very complex and the interpretation of situations is
Fig. 12. Long forward passes of BVB (top) and FCB (bottom) and the
average positions of all players (shown by colored dots) on top of the very subjective even for experts, a lot of data-driven projects
density fields of all players during the passes. fail in terms of communication between data-science and
soccer experts. The football experts concluded that the pro-
different ways to reach their opponents’ goal. FCB mostly posed VA approach is a great step towards making the
used the right flank and then turned towards the center of complex spatio-temporal tracking and event data under-
the penalty box. The overall shape of the BVB attacks looks standable and so usable for professionals.
like an arrow targeted straight at the FCB goal. Application of visual analytics approaches allowed our
The position variation indicators (i.e., the median dis- group to find many interesting patterns that would be very
tances to the representative positions) can be compared difficult or even impossible to detect by means of watching
along and across the pseudo-trajectories. For example, the game footage on TV or an animated visual representation of
variation of the positions of the FCB’s right midfielder the data. We were able to identify patterns, compare them,
is much higher than that of the left midfielder. This can and synthesize further higher-level patterns. Moreover, ob-
be related to the fact that the ball was transferred to the taining interesting and meaningful results motivated many
penalty box mostly from the right flank, and the opponents further analysis scenarios such as identifying formations
were putting more pressure on that side forcing the right during long periods of ball possession, assessing efficiency
midfielder to vary his positions. of counter-pressing efforts, comparing evolution of playing
Our group considered two options for studying further style of the same team over a season and across multiple
THIS ARTICLE IS PUBLISHED IN IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS 15

Selected long passes of BVB: pitch and FCB team space

Selected long passes of FCB: pitch and BVB team space

Fig. 13. Changes in aggregated positions of players and the ball during the selected long passes of BVB (top) and FCB (bottom). Pass targets are
marked by cyan dots.

seasons, searching for impact of changes of team coaches In the experts’ opinion, the power of the framework
and/or key players. Considering such scenarios would be can be further increased by involving Key-Performance-
out of question if only the state-of-the-art techniques were Indicators (KPI) for the soccer game metrics such as ex-
available. pected goals, dangerosity, pass options based either on mea-
suring the difficulty of performing passes or representing
possible gain if a pass is successful (e.g. packing rate),
6 D ISCUSSION indicators of team compactness and structure such as space
This paper presents results of a research project involving occupation, team shape damage, stretch index etc., and
visual analytics researchers, data scientists, and football pressure indicators. These KPIs are on the rise and could
domain experts with the aim to find effective approaches be utilized in making episode queries to both validate and
to gaining practicable knowledge from real complex data. improve the metrics and to select game situations of interest,
The football experts were impressed by the capabilities of enabling further application scenarios.
the visual analytics techniques that were developed. They To put the results of this research into practice, it is nec-
said that the selection of similar game situations based on essary to develop software tools that can be easily utilized
underlying data and extraction of general patterns of the by end users. As there exist different categories of potential
teams’ and players’ behaviors in such situations has been users, as discussed below, user-centered design and devel-
so far an unsolved challenge. Hence, appropriate query opment may need to be done specifically for each category,
and generalization techniques would bring a big benefit taking into account the specific tasks, requirements, and
for experts, especially for scouts and match analysts. The capabilities of the target users. Different classes of users
proposed framework has a very high potential to bring all need different interfaces with different levels of interactivity
the data-insights finally on the pitch and thus produce a and visual complexity, but, irrespective of these, achieving
substantial impact on professional football. high level of automation in making queries, constructing
THIS ARTICLE IS PUBLISHED IN IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS 16

team in previous games and create automatically a cata-

logue of tactical schemes of opponents over a big set of
their previous games. Such automatically acquired tactical
schemes conditioned over different classes of situations
could be a great hint-giving and decision-supporting means.
Medical staff of clubs could examine movements of players
during episodes characterized by fast running at different
times in the games. Scouts could evaluate players’ move-
ments and actions in different classes of situations and spot
their strong and weak abilities. Journalists could present
tactical schemes and compare them in pre-game and post-
game articles or TV shows or even during game breaks.
Leagues could provide services to clubs and also enhance
their media products. Some user categories (particularly,
the latter two) require tools not only for analysis but also
for communication of the insights gained to certain audi-
ences, including the general public. This requires specific
approaches for synthesizing audience-targeted stories from
results of tactical analyses [57].
While the presented framework have been developed
with an orientation to football data and analysis tasks, it
is potentially generalizable to various kinds of coordinated
movements of multiple objects in applications where the
task of extracting general behaviour patterns under different
circumstances is relevant. Examples are movements of play-
ers in other team sports, such as ice-hockey or basketball,
behaviors of animal groups, or movements of people in
crowded environments. We also envision potential appli-
cations in domains of air and sea traffic management. The
Fig. 14. Build-up for the shots by FCB (top) and BVB (bottom). The shot main components of the approach, i.e., the query facilities
positions are marked by black crosses. The cyan line corresponds to for episode selection, the method for generalization and ag-
the ball. The dots with the sizes representing the variation mark the gregation, the data structure for representing generalization
generalized positions, which are separated by 1-second intervals.
results, and the visualization techniques, can be adjusted to
the specifics of various application domains.

7 C ONCLUSION

Our contribution can be summarized as proposing an ana-

lytical framework involving interactive queries, generaliza-
tion and aggregation of query outcomes, and comparative
visual exploration of resulting general patterns. The frame-
work makes use of an interesting and fruitful interplay of
physical and constructed spaces (pitch and team spaces) and
times (absolute and relative times). The query primitives
enable selection of sets of time intervals containing situa-
tions with specified characteristics and, moreover, further
Fig. 15. Build-ups for shots: pseudo-trajectories of the team centers selection of sets of intervals having particular temporal
before the FCB shots (left) and BVB shots (right). relationships to the previously selected intervals. This can
be used, in particular, for considering the episodes of situa-
tion development step-wise or for studying what happened
pseudo-trajectories, and putting them in visual displays is before or after them. The aggregation method produces a
of great importance. There exists a technical possibility to novel type of movement aggregate, pseudo-trajectory, con-
implement the presented framework in the form of auto- sisting of generalized positions arranged along an abstract
mated procedures oriented to specific analysis tasks. Such timeline. The aggregation results are visualized in ways
automated procedures can be used for extracting patterns enabling exploration, comparison, and assessment of the
from large databases containing data from many games. variation of the original movements summarized in the
We anticipate that the following user categories could aggregates. The techniques proved useful for discovery of
benefit from specific applications based on the framework. general patterns of collective movement behavior in diverse
Match analysts could evaluate the efficiency of their own classes of situations.
THIS ARTICLE IS PUBLISHED IN IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS 17

ACKNOWLEDGMENTS [26] C. Perin, R. Vuillemot, and J.-D. Fekete, “Soccerstories: A kick-off

for visual soccer analysis,” IEEE Transactions on Visualization and
This research was supported by Fraunhofer Cluster of Ex- Computer Graphics, vol. 19, no. 12, pp. 2506–2515, 2013.
cellence on “Cognitive Internet Technologies” and by EU in [27] D. Sacha, F. Al-Masoudi, M. Stein, T. Schreck, D. Keim, G. An-
project SoBigData. drienko, and H. Janetzko, “Dynamic visual abstraction of soccer
movement,” Computer Graphics Forum, vol. 36, no. 3, pp. 305–315,
2017.
[28] D. Sacha, M. Stein, T. Schreck, D. A. Keim, O. Deussen et al.,
R EFERENCES “Feature-driven visual analytics of soccer data,” in IEEE Conference
on Visual Analytics Science and Technology. IEEE, 2014, pp. 13–22.
[1] Wikipedia, Association football, 2019 (accessed February 14, 2019),
[29] L. Shao, D. Sacha, B. Neldner, M. Stein, and T. Schreck, “Visual-
https://en.wikipedia.org/wiki/Association football.
interactive search for soccer trajectories to identify interesting
[2] D. Memmert and D. Raabe, Revolution im Profifußball. Mit Big Data
game situations,” Electronic Imaging, vol. 2016, no. 1, pp. 1–10,
zur Spielanalyse 4.0. Springer, 2017.
2016.
[3] D. Link, Data Analytics in Professional Soccer. Springer, 2018.
[30] M. Stein, H. Janetzko, T. Schreck, and D. A. Keim, “Tackling
[4] D. Sumpter, Soccermatics: mathematical adventures in the beautiful
Similarity Search for Soccer Match Analysis: Multimodal Distance
game. Bloomsbury Publishing, 2016.
Measure and Interactive Query Definition,” in Symposium on Visu-
[5] J. Ladefoged, How Data (and Some Breathtaking Soc-
alization in Data Science (VDS) at IEEE VIS 2018, 2018.
cer) Brought Liverpool to the Cusp of Glory, 2019,
[31] M. Stein, H. Janetzko, T. Breitkreutz, D. Seebacher, T. Schreck,
https://www.nytimes.com/2019/05/22/magazine/soccer-
M. Grossniklaus, I. D. Couzin, and D. A. Keim, “Director’s cut:
data-liverpool.html (accessed June 18, 2019).
Analysis and annotation of soccer matches,” IEEE Computer Graph-
[6] Bundesliga, Borussia Dortmund - FC Bayern München 3:2,
ics and Applications, vol. 36, no. 5, pp. 50–60, 2016.
10.11.2018 (accessed February 14, 2019). [Online]. Avail-
able: https://www.bundesliga.com/de/bundesliga/spieltag/ [32] M. Stein, J. Häußler, D. Jäckle, H. Janetzko, T. Schreck, and D. A.
2018-2019/11/borussia-dortmund-vs-fc-bayern-muenchen/stats Keim, “Visual soccer analytics: Understanding the characteristics
[7] ——, Borussia Dortmund - 1. FC Nürnberg 7:0, of collective team movement based on feature-driven analysis and
26.09.2018 (accessed February 14, 2019). [Online]. Avail- abstraction,” ISPRS International Journal of Geo-Information, vol. 4,
able: https://www.bundesliga.com/de/bundesliga/spieltag/ no. 4, pp. 2159–2184, 2015.
2018-2019/5/borussia-dortmund-vs-1-fc-nuernberg/stats [33] G. Andrienko, N. Andrienko, G. Budziak, T. von Landesberger,
[8] Wikipedia, Formation (association foot- and H. Weber, “Coordinate transformations for characterization
ball), 2019 (accessed February 14, 2019), and cluster analysis of spatial configurations in football,” in Joint
https://en.wikipedia.org/wiki/Formation (association football). European Conference on Machine Learning and Knowledge Discovery
[9] G. Andrienko, N. Andrienko, G. Budziak, J. Dykes, G. Fuchs, in Databases. Springer, 2016, pp. 27–31.
T. von Landesberger, and H. Weber, “Visual analysis of pressure in [34] J. Gudmundsson and T. Wolle, “Football analysis using spatio-
football,” Data Mining and Knowledge Discovery, vol. 31, no. 6, pp. temporal tools,” Computers, Environment and Urban Systems,
1793—-1839, 2017. vol. 47, pp. 16–27, 2014.
[10] J. Wilson, Inverting the pyramid: the history of soccer tactics. Nation [35] M. Horton, J. Gudmundsson, S. Chawla, and J. Estephan, “Auto-
Books, 2013. mated classification of passing in football,” in Pacific-Asia Confer-
[11] A. Zauli, Soccer: Modern Tactics. Reedswain Inc., 2002. ence on Knowledge Discovery and Data Mining. Springer, 2015, pp.
[12] M. Lucchesi, Attacking soccer: A tactical analysis. Reedswain Inc., 319–330.
2001. [36] M. Stein, H. Janetzko, A. Lamprecht, T. Breitkreutz, P. Zimmer-
[13] A. Bialkowski, P. Lucey, P. Carr, Y. Yue, S. Sridharan, and I. A. mann, B. Goldlücke, T. Schreck, G. Andrienko, M. Grossniklaus,
Matthews, “Large-scale analysis of soccer matches using spa- and D. A. Keim, “Bring it to the pitch: Combining video and
tiotemporal tracking data,” in 2014 IEEE International Conference movement data to enhance team sport analysis,” IEEE Transactions
on Data Mining, ICDM 2014, Shenzhen, China, December 14-17, 2014, on Visualization and Computer Graphics, vol. 24, no. 1, pp. 13–22, Jan
R. Kumar, H. Toivonen, J. Pei, J. Z. Huang, and X. Wu, Eds. IEEE, 2018.
2014, pp. 725–730. [37] STATSEdgeViewer, https://www.stats.com/edge/, 2019 (accessed June
[14] A. Bialkowski, P. Lucey, G. P. K. Carr, Y. Yue, S. Sridharan, and 18, 2019).
I. A. Matthews, “Identifying team style in soccer using formations [38] SAPSportsOne, https://www.sap.com/germany/products/sports-
learned from spatiotemporal tracking data,” in 2014 IEEE Interna- one.html, 2019 (accessed June 18, 2019).
tional Conference on Data Mining Workshops, ICDM Workshops 2014, [39] SportecSolutions, http://www.bundesliga-datenbank.de/, 2019 (ac-
Shenzhen, China, December 14, 2014, Z. Zhou, W. Wang, R. Kumar, cessed June 18, 2019).
H. Toivonen, J. Pei, J. Z. Huang, and X. Wu, Eds. IEEE, 2014, pp. [40] G. Andrienko, N. Andrienko, P. Bak, D. Keim, and S. Wrobel,
9–14. Visual Analytics of Movement. Springer, 2013.
[15] J. Perl, A. Grunz, and D. Memmert, “Tactics analysis in soccer– [41] G. Andrienko, N. Andrienko, W. Chen, R. Maciejewski, and
an advanced approach,” International Journal of Computer Science in Y. Zhao, “Visual analytics of mobility and transportation: State
Sport, vol. 12, no. 1, pp. 33–44, 2013. of the art and further research directions,” IEEE Transactions on
[16] Y. Wu, X. Xie, J. Wang, D. Deng, H. Liang, H. Zhang, S. Cheng, and Intelligent Transportation Systems, vol. 18, no. 8, pp. 2232–2249, Aug
W. Chen, “Forvizor: Visualizing spatio-temporal team formations 2017.
in soccer,” IEEE Transactions on Visualization and Computer Graphics, [42] C. Weaver, “Cross-filtered views for multidimensional visual anal-
vol. 25, no. 1, pp. 65–75, Jan 2019. ysis,” IEEE Transactions on Visualization and Computer Graphics,
[17] ChyronHego, https://chyronhego.com/, 2019 (accessed June 18, 2019). vol. 16, no. 2, pp. 192–204, March 2010.
[18] OPTA, https://www.optasports.com/, 2019 (accessed February 14, [43] H. Hochheiser and B. Shneiderman, “Dynamic query tools for
2019). time series data sets: Timebox widgets for interactive exploration,”
[19] STATS, https://www.stats.com/, 2019 (accessed February 14, 2019). Information Visualization, vol. 3, no. 1, pp. 1–18, Mar. 2004.
[20] S. Spectrum, https://www.secondspectrum.com/, 2019 (accessed June [44] S. K. Gadia, “A homogeneous relational model and query lan-
18, 2019). guages for temporal databases,” ACM Trans. Database Syst., vol. 13,
[21] Track160, https://track160.com/, 2019 (accessed June 18, 2019). no. 4, pp. 418–448, Oct. 1988.
[22] FootoVision, https://www.footovision.com/, 2019 (accessed February [45] C. S. Jensen, J. Clifford, S. K. Gadia, A. Segev, and R. T. Snodgrass,
14, 2019). “A glossary of temporal database concepts,” SIGMOD Rec., vol. 21,
[23] G. Andrienko, N. Andrienko, and G. Fuchs, “Understanding no. 3, pp. 35–43, Sep. 1992.
movement data quality,” Journal of Location Based Services, vol. 10, [46] N. Andrienko, G. Andrienko, E. Camossi, C. Claramunt, J. M. C.
no. 1, pp. 31–46, 2016. Garcia, G. Fuchs, M. Hadzagic, A.-L. Jousselme, C. Ray, D. Scar-
[24] D. J. Sumpter, Collective animal behavior. Princeton University latti, and G. Vouros, “Visual exploration of movement and event
Press, 2010. data with interactive time masks,” Visual Informatics, vol. 1, no. 1,
[25] J. Gudmundsson and M. Horton, “Spatio-temporal analysis of pp. 25 – 39, 2017.
team sports,” ACM Computing Surveys (CSUR), vol. 50, no. 2, p. 22, [47] T. von Landesberger, S. Bremm, T. Schreck, and D. W. Fellner,
2017. “Feature-based automatic identification of interesting data seg-
THIS ARTICLE IS PUBLISHED IN IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS 18

ments in group movement data,” Information Visualization, vol. 13, Gabriel Anzer is the lead data scientist at
no. 3, pp. 190–212, 2014. Sportec Solutions GmbH, a subsidiary of the
[48] N. Andrienko, G. Andrienko, J. M. C. Garcia, and D. Scarlatti, Deutsche Fußball Liga (DFL). He holds a M.Sc.
“Analysis of flight variability: a systematic approach,” IEEE Trans- in Financial Mathematics and Actuarial Sci-
actions on Visualization and Computer Graphics, vol. 25, no. 1, pp. ences. His research focuses on using spatio-
54–64, Jan 2019. temporal positional data to analyze individual
[49] N. Andrienko, G. Andrienko, L. Barrett, M. Dostie, and P. Henzi, and team based performances of soccer players.
“Space transformation for understanding group movement,” IEEE
Transactions on Visualization and Computer Graphics, vol. 19, no. 12,
pp. 2169–2178, Dec 2013.
[50] G. Andrienko, N. Andrienko, H. Schumann, and C. Tominski, Vi-
sualization of Trajectory Attributes in Space–Time Cube and Trajectory
Wall. Berlin, Heidelberg: Springer Berlin Heidelberg, 2014, pp.
157–163. Pascal Bauer has a background in mathematics
[51] N. Andrienko and G. Andrienko, “Spatial generalization and and data-science in applied research at Fraun-
aggregation of massive movement data,” IEEE Trans. Vis. Comput. hofer and worked as a coach/speaker at Fraun-
Graph., vol. 17, no. 2, pp. 205–219, 2011. hofer Big Data & Artificial Intelligence Alliance.
[52] N. Willems, H. Van De Wetering, and J. J. Van Wijk, “Visualization He holds a UEFA A-level coaching license with
of vessel movements,” Computer Graphics Forum, vol. 28, no. 3, pp. almost nine years of experience as a head
959–966, 2009. coach and has a passion for soccer. At the DFB
[53] G. Andrienko and N. Andrienko, “Spatio-temporal aggregation Academy, he is responsible for a wide field of
for visual analysis of movements,” in Proceedings of the data analysis & machine learning applications
IEEE Symposium on Visual Analytics Science and Technology in soccer, including talent identification, injury
(IEEE VAST) 2008, 2008, pp. 51–58. [Online]. Available: prediction models, tactical match analysis based
https://doi.org/10.1109/VAST.2008.4677356 on positional data, and much more.
[54] G. Andrienko, N. Andrienko, G. Fuchs, and J. M. Cordero-Garcia,
“Clustering trajectories by relevant parts for air traffic analysis,”
IEEE Transactions on Visualization and Computer Graphics, vol. 24,
no. 1, pp. 34–44, Jan 2018. Georg Fuchs is a senior research scientist
[55] A. Buja, J. A. McDonald, J. Michalak, and W. Stuetzle, “Interactive heading the Big Data Analytics and Intelligence
data visualization using focusing and linking,” in Proceedings of the division at Fraunhofer IAIS. His research work is
2Nd Conference on Visualization ’91, ser. VIS ’91. Los Alamitos, CA, focussed on visual analytics and Big Data ana-
USA: IEEE Computer Society Press, 1991, pp. 156–163. [Online]. lytics, with a strong emphasis on spatio-temporal
Available: http://dl.acm.org/citation.cfm?id=949607.949633 and movement data analysis. His further re-
[56] J. Candil, Pep’s five-second rule, the key to City’s success, 2018, search interests include information visualization
https://en.as.com/en/2018/07/26/football/1532614241 079674.html in general, Smart Visual Interfaces, and com-
(accessed February 14, 2019). puter graphics. He has co-authored 55+ peer-
[57] S. Chen, J. Li, G. Andrienko, N. Andrienko, Y. Wang, P. H. Nguyen, reviewed research papers and journal articles,
and C. Turkay, “Supporting story synthesis: Bridging the gap including a best short paper award at Smart
between visual analytics and storytelling,” IEEE Transactions on Graphics 2008 and a VAST challenge award in 2014.
Visualization and Computer Graphics, pp. 1–1, 2019.

Guido Budziak holds a master’s degree in

Computer Science. He played professional foot-
ball for 10 years in the Netherlands, including
Dutch national youth teams. He founded Con-
nected.Football, a football technology company
Gennady Andrienko is a lead scientist respon- aimed at making expert football knowledge ac-
sible for visual analytics research at Fraunhofer cessible to youth academies and amateur foot-
Institute for Intelligent Analysis and Information ball clubs. His research work focuses on commu-
Systems and part-time professor at City Uni- nicating football tactics and tactical performance
versity London. Gennady Andrienko was a pa- analysis.
per chair of IEEE VAST conference (2015–
2016) and associate editor of IEEE Transac-
tions on Visualization and Computer Graphics
(2012–2016), Information Visualization and In-
ternational Journal of Cartography. Dirk Hecker is vice-director of the Fraunhofer
Institute for Intelligent Analysis and Information
Systems IAIS and a member of the board of di-
rectors of Fraunhofer Academy. His current top-
ics of work include spatial analytics, deep learn-
ing and trustworthy AI. He has authored multiple
publications on these topics and works as expert
and auditor in several boards concerned with
artificial intelligence.
Natalia Andrienko is a lead scientist at Fraun-
hofer Institute for Intelligent Analysis and Infor-
mation Systems and part-time professor at City
University London. Results of her research have
been published in two monographs, ”Exploratory Hendrik Weber leads the area of innovation
Analysis of Spatial and Temporal Data: a Sys- and sports technology for the German Foot-
tematic Approach” (2006) and ”Visual Analytics ball League (DFL) and is managing director of
of Movement” (2013). Natalia Andrienko is an the subsidiary Sportec Solutions responsible for
associate editor of IEEE Transactions on Visu- Sport Tech Operations. He holds a PhD in busi-
alization and Computer Graphics. ness administration and is frequent lecturer and
author for performance analysis in sports.
THIS ARTICLE IS PUBLISHED IN IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS 19

Stefan Wrobel is Professor of Computer Sci-

ence at University of Bonn and Director of the
Fraunhofer Institute for Intelligent Analysis and
Information Systems IAIS. His work is focused
on questions of the digital revolution, in particular
intelligent algorithms and systems for the large-
scale analysis of data and the influence of Big
Data/Smart Data on the use of information in
companies and society. He is the author of a
large number of publications on data mining and
machine learning, is on the Editorial Board of
several leading academic journals in his field, and is an elected founding
member of the “International Machine Learning Society”.
B Appendix—Study II: The Origins of Goals in the
German Bundesliga

In the following, we present Anzer, Bauer, and Brefeld (2021),

an accepted manuscript of an article published by Taylor & Fran-
cis in Journal of Sport Science on July 25th 2021, available online:
http://www.tandfonline.com/10.1080/02640414.2021.1943981

96
 The Origins of Goals in the German Bundesliga

Gabriel Anzer1,2 ID
, Pascal Bauer2,3 ID
and Ulf Brefeld4 ID

1
Sportec Solutions AG, subsidiary of the Deutsche Fußball Liga (DFL)
2
Institute of Sports Science, University of Tübingen
3
DFB Akademie, Deutscher Fußball-Bund e.V. (DFB)
4
Leuphana University of Lüneburg, Machine Learning Group

Abstract
We propose to analyze the origin of goals in professional football (soccer) in a purely data-
driven approach. Based on positional and event data of 3, 457 goals from two seasons German
Bundesliga and 2nd Bundesliga (2018/20219 and 2019/2020), we devise a rich set of 37 features
that can be extracted automatically and propose a hierarchical clustering approach to identify
group structures. The results consist of 50 interpretable clusters revealing insights into scoring
patterns. The hierarchical clustering found 8 alone standing clusters (penalties, direct free
kicks, kick and rush, one-two’s, assisted by header, assisted by throw-in) and 9 categories (e.g.
corners) combining more granular patterns (e.g. 5 subcategories of corner-goals). We provide
a thorough discussion of the clustering and show its relevance for practical applications in
opponent analysis, player scouting and for long-term investigations. All stages of this work
have been supported by professional analysts from clubs and federation.

Keywords Sports analytics • Professional football (soccer) • Hierarchical clustering • Tactical analysis

1 Introduction
In the 1960s, Charles Reep began to manually annotate games of Swinden Town FC (Reep et al., 1968).
Though, some of his data-driven conclusions were later questioned (Witts, 2019) , his primitive collection of
detailed game data constituted the birth of football analytics: Today, most international leagues not only
collect manually annotated events from their matches systematically, but also use device- or camera-based
tracking systems in addition. Compared to event logs focusing on ball-actions (e.g. passes, shots, fouls;
often referred to as event data), tracking systems allow to record the positions of all 22 players and the
ball for an entire match (often referred to as positional or tracking data).
Several studies aim to group goals—the most important quantified metric in football—into predefined
categories. For example, González-Ródenas et al. (2019) differentiate between open-play and set-pieces to
categorize 380 goals from of the UEFA Champions League 2016/2017 season. They observe that 75.9% of
all goals occur from open-play, and only 24.1% are scored from set-pieces. A similar study confirms these
numbers on 101 goals taken from the World Cup in 2010 (Njororai, 2013). However, the expressiveness of
manually crafted categories is naturally limited as only a relatively small amount of data can be processed
by hand. Focusing on more detailed features of the goal origin, rather than high-level groups (e.g. whether
the shooter was under pressure), Mitrotasios et al. (2012) investigated factors associated with goal scoring
in 76 matches of the European Championship in 2012. Besides finding a similar dispersion of goal origins
(27.6% after set-pieces; 72.4% from open-play), they show that in more than 50% of the cases the goal-
scorer took his shot without any pressure. Plummer (2013) analyze goal scoring patterns of a lower English
league, pointing out stark differences in the origin of goals between non-professional leagues. In general,
set-pieces are often a decisive factor for winning a game, particularly when teams are equally strong
(Szwarc, 2007; Göral, 2019). Especially for corner-kicks, there exist several studies examining how they
lead to goals: (Taylor et al., 2005; Carling et al., 2006; Armatas et al., 2007; Schmicker, 2013; Pulling
et al., 2013; Pulling, 2015; Casal et al., 2015; Fernández-Hermógenes et al., 2017; Casal et al., 2017).
While these papers are based on manually annotated data, Power et al. (2017) offer an approach using
tracking data. They report a scoring efficiency of 2.1% after corner kicks for the English Premiere League
2016/2017 season. They also found that scoring with the second ball touch after a corner, is even more
likely than converting with the first touch.

1
 Whereas the usage of manually recorded data, acquired for the sole purpose of a single investigation,
is a common practice in sport sciences, the potential of automatically acquired positional data, as well as
off-the-shelf available event data, has not been fully exploited—particularly when it comes to clustering
goals. Hobbs et al. (2018) aim to detect counterattack situations automatically, based on positional and
event data, and derive that it is the most efficient strategy for scoring goals. Sarmento et al. (2014)
propose mixed methods to analyze attacking patterns of 36 games of different European top teams. They
also focused on counterattacks and combined quantitative analyses with expert knowledge to discover team
philosophies, showing that the combination proved to be very beneficial. Several studies highlighted the
relevance of this interplay between sports and computer science (Rein et al., 2016; Herold et al., 2019;
Andrienko et al., 2019; Goes et al., 2020; Marcelino et al., 2020)
Note that a similar approach has been successfully deployed in basketball. Reich et al. (2006) investigate
so-called shot charts, that visualize the location and outcome of every shot, in professional basketball
matches. Similar visualizations are enriched by spatial clustering techniques in (López et al., 2013).
An approach taking different shot characteristics into consideration is presented in Erčulj et al. (2015).
Although these analyses often focus on the location of shots, it led to significant changes in team strategies
and player’s shooting decisions (López et al., 2013; Reich et al., 2006). Simply focusing on shot locations
of goals does of course not translate to football with its complex attacking plays, its low scoring nature,
different shot types (e.g. header, volley) and the additional role of a goalkeeper.
Since only about 1% of all ball possession phases are completed with a goal (Pollard et al., 1997; Tenga
et al., 2010), many studies thus extend the focus to all shots (Fernando et al., 2015) or on proxies such as
carrying the ball into dangerous zones (Njororai, 2013; Merlin et al., 2020) in order to quantify offensive
success on larger sample sizes. Although these approaches may be biased towards successful teams that use
their chances more effectively (Castellano et al., 2012; Delgado-Bordonau et al., 2013; Dufour et al., 2017),
they allowed studies to evaluate processes (i.e. an attacking-play) more granular than just by considering
pure results. Ruiz et al. (2015) analyze the efficiency of shots taken based on the distance and angle to
the goal, Schulze et al. (2018) consider also the set-up of the opposing team during the shot to improve
the expressiveness of the investigation. On the basis of these ideas, a lot of expected goal models exist
that aim to quantify scoring probabilities (Lucey et al., 2014; Rathke, 2017; Ruiz et al., 2017; Robberechts
et al., 2020; Anzer et al., 2021).
Consequently, compared to other team sports, goals in football are rare events and there is a trivial
probability of observing the same goal twice as every goal is sui generis due to the complexity of the game
(Siegle et al., 2013; Salmon et al., 2020). Nonetheless, teams come up with dedicated match plans to
increase the probability of scoring and winning the game. Coaches and video-analysts devise attacking
patterns that ought to exploit weaknesses of the opposing team and result in the creation of chances.
Since these patterns are not random, there must be structure in the creation of goals. In that respect, the
categorization of goals plays an important role in the daily business of professional football clubs. Clubs
typically employ several match-analysts whose role includes to regularly examine scored and conceded
goals, particularly before facing an opponent. Since viewing the video footage is a tedious task and even
experts may disagree on categories (Chawla et al., 2017), it is the objective of this paper to both automatize
and objectify the categorization of goals and support the respective match-analysis departments: Being
able to cluster goals by their origin allows for an unbiased analysis that provides unseen patterns and
discloses trends.
In this paper, we follow a data-driven approach to leverage such data to identify the underlying structure
of the origin of goals in professional football. The contribution of this paper is as follows: First, we propose
a rich set of expert features that can be computed from aligned positional and event data to formally
represent goals as instances in a vector space. Second, we deploy a hierarchical clustering (Murtagh et al.,
2017) to group 3, 457 goals from two seasons of the German Bundesliga and 2nd Bundesliga into meaningful
and interpretable clusters and provide a thorough analysis of the results. Compared to the literature, our
analysis is on a much larger scale, provides rich feature representations, and follows a purely data-driven
approach that renders manual categorization or the definition of rules unnecessary. All quantitative results
have been evaluated qualitatively by professional match-analysts.

2 Methods
2.1 Data
The German Bundesliga and 2nd Bundesliga collects tracking and event data for all their league matches.
The former is captured by optical tracking systems while the latter consists of manual annotations. Tracking
data is recorded automatically using camera-based systems. Optical tracking systems are installed in every
stadium and capture the positions of players, referees and the ball at 25 frames per second. The quality of

2
the tracking data acquired by Chyronhego’s TRACAB system1 is evaluated on a regular basis and presents
sufficient accuracy (Taberner et al., 2020; Linke et al., 2020). However, there remain many events on the
pitch that currently cannot be captured automatically. The event data is therefore collected manually.
Trained operators annotate about 3, 000 basic events per match categorized into different event classes.
There are 30 top-level event classes including passes, crosses, fouls, etc. as well as about another 100
sub-attributes describing the events even in greater detail. The definition of each event follows the official
match data-catalogue designed by German Bundesliga.2 For further processing, the tracking and event
data are synchronized so that the timestamps of the events are aligned to the right frames in the tracking
data as described in Anzer et al. (2021).
We focus in our analysis on 3, 457 goals scored in the Bundesliga and 2nd Bundesliga in the 2017/2018
and 2018/2019 seasons and excluded the 85 own goals due to their often random nature. Every goal is
described by the raw data of all 22 players and the ball in 25Hz as well as all annotated events during
the ball possession phase leading to the goal. We also extracted 8.167 shots of the season 2018/2019
(containing 953 goals). The shots are used for an efficiency analysis of each cluster as described later.

2.2 Mapping goals into feature space

We extract a rich feature set from the synchronized data to turn goals into machine-readable quantities
encoding episodes that end with a successful shot at goal. We mirror the pitch in both dimensions, so that
all goals are scored on the same side of the field. Later on, this transformation remedies the clustering
from having to differentiate between left or right wings.
Besides the location and set-up of the shot itself, football experts (i.e. coaches, match-analysts, . . . ) are
explicitly interested in the complete ball possession phase prior to the goal. However, the fluent invasive
character of football implicates a lot of vagueness in terms of a consistent definition of an attacking play
(Merlin et al., 2020). Particularly very short ball possession phases of defending players during an attacking
play should not be considered as a separate ball possession phase. To establish an appropriate definition,
we reviewed video footage of critical scenes together with experts. Finally, we define the start of such an
episode as either a dead-ball situation (e.g. throw-in, goal-kick, etc.) or a turnover by the opposing team
lasting at least six seconds.
Together with experts—match-analysts with a minimum of five years experience in professional football
teams3 —we define in total 37 features describing the evolution of a goal, from the origin to its finish. The
features are described in detail in the Appendix A. To provide an accurate representation of what leads to
a goal, the features make full use of the synchronization of the positional data with the manual collected
event data. In total we settled on features describing the shot itself (location, type, goalkeeper positioning,
pressure on the goal scorer, . . . ), its assist (location, assist type, . . . ) and features describing the entire ball
possession phase leading up to the goal The latter features include the location and type of the initial gain
of the ball, the number of passes, meters dribbled and bypassed opponents. As a measure of chaos, the
number of opponent touches during the ball possession phase is also counted. Next to prominent scores like
expected goals (xG)4 , describing the probability of a shot being converted, we include several categorical
expert features, such as whether a chance is a sitter, originates from a counterattack, etc. Categorical
features are one-hot encoded in the final representation.
More sophisticated metrics describing the ball possession phase, the assist or the shot itself, present in
the literature were also used. To quantify the average pressure, for instance, we implemented the approach
taken by Andrienko et al. (2017). Additionally, the compactness of both teams is a decisive factor to
differentiate transition situations and counterattacks from other open-play situations. We therefore added
the stretch-index based on the definition in Santos et al. (2018) at the beginning and at the end of
the ball possession phase. Finally, the number of successfully played passes within an attacking play is
complemented with a packing value—describing the number of outplayed opponents per pass as in Steiner
et al. (2019)—to include a notion of the degree of ball control the offensive team had prior to scoring
the goal. All features were discussed, consolidated and steadily improved during workshops and based on
several steps of evaluation.5
1 https://chyronhego.com/products/sports-tracking/tracab-optical-tracking/, accessed 06/20/2020
2 https://www.bundesliga.com/en/news/Bundesliga/noblmd-dfl-subsidiary-sportcast-setting-up-

company-for-official-match-data.jsp, accessed 02/02/2020

3 We provide more information on the experts in the acknowledgements.
4 The xG-value used is calculated as defined in Anzer et al. (2021).
5 A video showing some of the features is available at https://bit.ly/3sa3phw.

3
2.3 Clustering the goals
To accomplish practical needs, it is our primary objective to automatically assign goals to interpretable
categories. We refrained from collecting labeled data from match-analysts for two reasons: On one hand,
categories of goals differ per club, coach and the respective match philosophy, and we prefer to compute
an objective structure that can be augmented in the daily practice irrespectively of the club, analyst or
philosophy. On the other hand, time constraints do allow match-analysts to review only a small amount
of data and the categorization of goals is naturally on a rather high level. Manual inspection of only
a few goals per opponent does not allow for detecting the variety of clusters that a purely data-driven
approach is able to produce at large-scales. A data-driven clustering allows us to reveal and discuss the
hidden structure of goals with our experts. To the best of our knowledge this is the first purely data-driven
approach to clustering goals on synchronized positional and event data and clearly unmatched in terms of
scale.
Agglomerative hierarchical clustering (HCA) provides a conceptually simple framework to compute
interpretable clusterings. HCA works bottom-up by (i) initializing every instance as a singleton cluster,
and (ii) iteratively combining the two most similar clusters, (iii) until only one cluster remains that contains
all instances. The resulting structure is a cluster tree called dendrogram (Murtagh et al., 2017). Different
instantiations of HCA arise by different ways to merge clusters in step (ii). For instance, single-link merges
the two clusters containing the two most similar elements (Sibson, 1973). Hence, single-link often leads to
chain-like structures as only one element of the cluster needs to be similar to one of the other while all other
instances may be very dissimilar. The other extreme is called max- or complete-link and focuses on the
most different elements when merging clusters (Defays, 2015). Max-link leads therefore to more balanced
clusters (Brian, 2011). We do not want to put such a strong prior on the solution and instead leverage a
compromise called average-link that merges clusters that are closest on average (Sokal, 1958). Average-link
is often used in bio- and health-related domains, for instance to infer phylogenetic tress (Felsenstein, 1996),
and serves our needs very well. However, instead of commonly used Euclidean distances, we deploy cosine
distance to meet the characteristics of the data. Recall that numeric elements of the feature representation
encode variables like packing or pass distance. Consider two similar goals, where one has almost twice the
packing score and almost twice the pass distance than the other. Using Euclidean distance, the two goals
would turn out very different. However, the angle between the two vectors in feature space is small and,
hence, the cosine implements the intuition that longer passes may also result in higher packing scores. In
sum, similarity of cluster X 0 and X is computed by

1 X X xT x0
sim(X, X 0 ) = (1)
| X || X | x2X 0 0 kxkkx0 k
0
x 2X

An extensive model selection optimizes the pre-processing pipeline as well as additional parameters like the
number of clusters. The final solution maximized the silhouette measure (Rousseeuw, 1987) and consists
of a z-transformation and a subsequent mapping onto the 20 most informative dimensions corresponding
to the largest eigenvalues identified in a principal component analysis (PCA) (Wold et al., 1987) before
the data is fed into the hierarchical clustering using 50 clusters.
The resulting dendrogram is shown in Figure 1. Starting at the root, the hierarchy differentiates
primarily between the assist type before splitting further into goals arising individualized features per
branch. The tree is evaluated together with professional match-analysts of national teams and a Bundesliga
club to analyze its possible use for practice. Together with the experts, we went through 2D visualizations
of the goals and corresponding video footage, in order to derive a better grasp of the clustering. To reduce
workload, primarily the 2D visualizations were used by the experts to assign names and descriptions to
all nodes in the dendrogram. These characterizations were evaluated on random samples of video footage
manually to verify the solution; in total more than 800 goals were viewed.
After finalizing the contextual description of the clustering, the experts agreed on a simplified version
of the tree they would use in their match-analysis. This simpler version essentially merges small clusters
with close neighbours. We indicate merged clusters by the same colors in Figure 1 and provide a thorough
discussion in the remainder.

3 Results
In this section, we discuss the induced grouping by the dendrogram in Figure 1 and highlight interesting
features of the data-driven solution. In the remainder, we differentiate between goals from open-play, set-
pieces, type of assist, type of shot, and dedicated special goals. Representative goals for each cluster can be
found in the Appendix D. To assess conversion rates per cluster, we classified 8,167 shots from 2018/2019
season into the clustering by assigning every goal to the most similar cluster using the distance metrics

4
Figure 1: The resulting dendrogram with contextual annotations. Numbers show the amount of goals in
the respective branches.

suggested in the previous chapter. Table 5 in the Appendix C provides additional details on conversion
rates per cluster.

3.1 Open-play
A straight forward classification of goals is to differentiate between goals that originate from open-play
and from set-pieces. With in total 2, 231 goals, in-play goals constitute the most frequent type of goals
in our data, with 64.0% of all goals being placed in one of the corresponding clusters. Focusing on the
former, open-play ball possession phases leading to a goal contain 3.6 passes and last 12.8 seconds on
average. Clusters containing goals from open-play are spread throughout the clustering; Figure 2 shows
two-dimensional visualizations of exemplary goals for the largest clusters.
The majority of all in-play goals are contained in the light green and dark green clusters and add up to
a total of 1, 424 goals. The goals can be distinguished by an intended assist from a teammate, without an
opponent touching the ball between assist and shot. The individual clusters further differentiate nuances
of the goal’s origin. For example, LP1 and LP2 in the light green cluster represent prototypical goals
from build-up to a finish. Goals in LP2 however, are typically the greater chances as 98.0% are labeled as
sitters and their goals per shot ratio is 42.0% LP2 compared to LP1’s of 20.0%.
The clustering allows to further dive into the resulting groups and show fine granular differences that
are usually only identified by manual expert inspections. As an example, Figure 3 shows that the clustering
differentiates between different strategies to regain the ball during an opponent’s possession. Coaches and
teams develop complex patterns that involve coordinated actions by many players and we easily identify
goals after successful counterpressing (SP1), midfield-block pressing (SP2) and high-block pressing (SP5).
Figure 3 shows heat maps of the shot location (top), the assist location (center), and the start of the ball
possession phase (bottom). SP1, for example, contains 128 goals scored after regaining the ball in the
opponents half, preferably close to the sideline.
Interestingly, about 40.0% of all shots in SP5 lead to a goal compared to only 5.0% for SP1 and 2.0%
for SP2. The numbers support that excellent goal opportunities are created by a very high pressing. By
contrast, SP2 turns out the most inefficient cluster in terms of goal conversion rate.
In-game crosses that are directly converted into goals are contained in HV3 and HC. Both clusters
encode crucial goal-scoring patterns. In HV3, for example, the ball is gained in the own half and after a
save build-up phase crossed from just inside the box and converted directly with a header (typically labeled
as a sitter). HC distinguishes itself by broader areas where the ball has been won, particularly including
the wings in the opponent’s half and crosses in this cluster are predominantly played from outside the box.
Figure 4 visualizes the differences using heat maps. Note that HC contains more than 10% of all goals

5
Figure 2: Exemplary in-play goals and their clusters. Arrows show the path of the ball leading up to
the successful shot, positions of players at the time of shot are indicated in blue and red,
respectively.

and is by far the largest cluster in the tree.

3.2 Set-Pieces
Roughly a third (in total 1, 227) of all converted ball possession phases in our data begin with a set-piece
in the opponent’s half. Hence, match-analysts dedicate a significant amount of their time to identify
opponent’s strategies and tricks for all sorts of set-pieces. Figure 5 shows exemplary goals from clusters
representing corners and crosses. A total of 7.1% of all goals originate in corners and contained in the cyan
clusters C1–C5. Cluster C1 contains all goals where the ball is touched by at least one opponent before
it is received by the scorer; these situations often end-up in rather uncontrolled ping-pong situations in
the box. Cluster C2 encodes flick-ons, where a target player is positioned at the closest post who slightly
deflects the ball before it can be converted. This cluster is complemented with HA1 that contains header
flick-ons. Goals in C2 show very high xG values with an average of 60.0% and all were rated as sitters by
the experts.
Set-pieces played as crosses into the box follow a similar idea as corner kicks but turn out to be less
effective. In total we count 330 freekick-crosses in cluster S1 and S2 in the data but only 4.7% of them
were converted to goals. The clustering distinguishes between three scenarios: Taking the freekick-cross
directly (FC1, 97 goals), scoring after a resulting ping-pong-situation (FC2, 30 goals), and scoring the
rebound of a freekick-cross in a spectacular way (SO2, 12 goals).
The most straight-forward way of turning set-pieces into goals is through penalties (272 goals, S1) and
direct freekicks (94 goals, S2). Together, the two clusters account for 13.1% of all scored goals in the two
seasons. Unsurprisingly, penalties are the most efficient way of scoring. Even without taking deflected
penalties into consideration, 91 penalties in Bundesliga season 2018/2019 lead to 74 direct goals which
corresponds to a conversion rate of 81.3%. If the goalkeeper initially parries a penalty, but the rebound is
then converted, the goal is not considered as a penalty goal and therefore part of a different cluster SO3
and described in the remainder.
In total 7.2% of all direct attempted freekicks from Bundesliga 18/19 season lead to a goal. Figure 6
shows the shot locations of all directly scored freekicks in S2. Throw-in crosses very rarely lead to goals
(20 goals in total), which are contained in TI.

3.3 Assists
The dendrogram in Figure 1 differentiates between types of assists. Clusters S1 and S2 as described in the
previous section, stem from directly scored set-pieces and trivially do not contain assists. Similarly, BE
represents goals where a pass of the defending team was intercepted and converted by the scorer. Figure 7

6
Figure 3: Goals originating from strategic ball gains in SP1, SP2, and SP5. The figure shows heat maps
for shot (top row), assist location (center row) and start of the ball possession phase (bottom
row). In total 270 goals (7.8% of all) were scored this way.

(top left) shows a heat map of shot locations in BE. While the majority of goals in that cluster are scored
from within the box, there are clearly visible outliers indicating long-ranged shots at goal. On average,
cluster BE is characterized by fatal build-up errors that allow shots from large distances to be converted
due to mispositioning by the goalkeeper.
The clustering further differentiates between types of assists such as an intentionally played final pass
that clearly aims to assist the scorer. This very large group containing 1, 949 goals is further divided
by the clustering into goals from open-play with an intended assist (dendrogram LP1–LP5, SP1–SP6,
HV1–HV3 and OT) and assists in form of crosses. The latter contains directly converted goals by
corners (C3), freekick-crosses (FC1), and open-play crosses (HC) as well as goals arising only after several
opposing ball touches; these ping-pong situations are again separated into corners (C1), freekick-crosses
(FC2), and open-play crosses (LO3 and LO5). Moreover, there are spectacular rebound-volleys where
unsuccessful clearances are scored at large distances (C1 and SO2, see below). Cluster HA1 contains
header assists by flick-ons after crosses and long balls from the own half.
Unintentional assists may arise from regular passes that are completed with outstanding maneuvers
of the scorer and can be found in SA2 and SA1. The contextual analysis of these clusters showed two
different kinds of situations: Either the scorer takes a surprise shot, often at large distance or from difficult
angles (two plots on the right side), or dribbles past several opponents before taking a shot.
Many unintentional assists are simply random and contained in the clusters UA1–UA4. In contrast
to assists by opponents (LO1–SO7), these random assists come in fact from a teammate but without the
direct intention to create a shot. The experts consider goals in this group to be lucky events. Nevertheless,
fortune picks its favorites: in our data, the luckiest teams in every league and season scored about twice as
many random goals as the unluckiest ones. Cluster SA contains shot attempts that are deflected by team
members. Positioning players in the line of shot turns out to be very efficient: almost half of the situations
are converted into goals
The last group in this section constitutes indirect assists from opponents. We already discussed in-
tercepted build-ups in BE, however, compared to BE, indirect assists in LO1–SO7 primarily stem from
uncontrolled and random opponent actions. Additional indirect assists are also contained in the ping-pong
clusters C1 (corners) and FC2 (freekick-crosses). Figure 8 visualizes exemplary goals induced by indirect
assists. For example, Cluster SO3 contains all goals where the opponent’s goalkeeper failed to save a shot
and accidentally assisted the scorer. With 197 goals this cluster contains surprisingly many goals, albeit,
our experts do not consider all these situations as mistakes by the goalkeeper. Although ’flaws’ of the
goalkeepers are often a decisive factor in top leagues, a characteristic trait of excellent strikers is their
sixth sense for these poacher goals.

7
Figure 4: Visualizations of goals scored by headers: example goal (top row), scorer position (2nd row),
assist location (3nd row), begin attacking phase (bottom row).

3.4 Shots
Possibly the most important part of a goal is the shot itself. We differentiate between leg-shots, volleys,
and headers. From our 3, 457 goals, 83.0% are scored by a non-volley leg-shot. The remaining 17.0%
are either headers or volleys and exemplified in Figure 4. Surprisingly, more than the half of these goals
originate from open-play phases (HV3 and HC). For instance, Clusters SO2 and SO1, displayed in Figure
8, contain lovely volley rebounds.
Headers are the predominant way to score after freekicks (74.0%) and corners (87.0%) but play only a
minor role in ping-pong situations (C1 and FC2). Cluster HV2 for example contains spectacular headers,
some of which are also highlighted as triangles in Figure 6.
Cluster HV1 and HV3 are efficient ways of scoring with conversion rates of 26.7% and 31.0%, re-
spectively. As mentioned above, HV3 encapsulates a blueprint worth striving for. Cluster HV1 shows
another constellation: Either a freekick or a corner is cleared by the opponent followed by a second cross
into the box that is then converted in a goal. From the overall 32, 406 shots, 5, 612 headers and volleys led
to 616 goals (11.0%) which is slightly more efficient than leg-shots (10.7%).

3.5 Patterns
Many clusters encode strategic patterns or tricks and by discovering the next opponent’s strategies one
can increase the likelihood of winning. Some of these strategies can be seen in Cluster SA where strikers
cross the line of the shot (likely) on purpose as well as in HA1 with header flick-ons. From the perspective
of a goalkeeper it is crucial to know the locations of freekicks, direct shots as well as crosses into the box.
Figure 6 thus shows the locations of successful long-distance shots depending on the cluster.
The most basic tactical pattern in football is a one-two and encoded by cluster OT. Figure 6 visualizes
a nice example of this pattern.
Cluster K&R represents the kick-and-rush strategy. Goals in this cluster are characterized by a long-
distance pass to the scorer. These passes bridge on average 48.47 m, and are often difficult to control.
Finally, a cluster containing special corner-tricks is C5. Clearly, knowing whether the next opponents
have some corner-tricks in their portfolio is an important piece of information for every coaching staff.

4 Discussion
Analyzing the origin of goals is often limited to small sample sizes due to manual annotation, nevertheless,
studies breaking down scoring patterns are common in sport-science literature (Reep et al., 1968; Njororai,
2013; Mitrotasios et al., 2012). Exploiting the availability of positional and event data can present a change
in paradigm for pattern analysis in football. The automated analysis based on 3, 457 goals, allows us to
put results from recent literature on a sound base: With 64.0% goals scored from open-play, we present
a lower number than previous literature (e.g. Njororai (2013) 75.86% of 145 from several competitions;
Mitrotasios et al. (2012) 72.4% of 76 goals European championship). Njororai (2013) claimed that history

8
Figure 5: Exemplary visualizations of all goals occurring by corners and crosses. Both goals from set-
pieces and from open-play are included representing a total amount of 490 goals (14.17%).

showed a trend towards more open-play goals, which cannot be confirmed by our data-set. Goals occurring
from open-play follow and attacking phase with 3.6 passes on average and the average conversion rate of
shots is 11.67% roughly in line with the original findings from Reep et al., 1968, and later confirmed by
Collet (2013); Sarmento et al. (2014); Vogelbein et al. (2014); González-Ródenas et al. (2019). Another
insight regarding set-pieces, is the lower header-rate of freekicks (74.0%) compared to corners (87.0%),
which can be explained by the additional space behind the offside line, increasing the likelihood of creating
enough separation to finish the cross with the foot. However, the definition which goal still counts as a
converted set piece or when a possession phase starts, varies across the literature, making a comparison
between the results difficult—data-driven studies like the one presented here could overcome this issue by
using consistent definitions, without the need for manual annotations.
Nevertheless, the key benefit of our approach is not the ability to conduct a large-scale descriptive
analysis, but rather to use a hierarchical clustering in order to identify patterns in the origin of goals au-
tomatically and, consequently, to derive meaningful insights for football practitioners from these patterns.
The efficiency of fast ball regains followed by a successful offensive action has been investigated in several
studies (Reep et al., 1968; Hobbs et al., 2018; Vogelbein et al., 2014). Our clustering detects that strategy
as a pattern represented in its own cluster in SP1 (3.7% of all goals). Another useful insight are partic-
ularly high conversion rates of ball-gains after high-blocks (SP5, 40.0%), especially in comparison with
ball gains after counterpressing (SP1, 5.0%) and after mid-blocks (SP2, 2.0%). This finding regarding
the efficiency of counterpressing is in line with Bauer et al. (2021). In the latter case when the ball is won,
the defense is typically quite well organized with many players behind the ball, often leading to long-range
shots. Compared to the usual categorization into corners, direct freekicks, freekick-crosses and penalties,
our approach allows for a much granular view of set-pieces and discloses hidden insights. Cluster HA1
and C2, contain several flick-on goals after corner kicks and confirm the relevance of this sub-category of
corner goals presented in Power et al., 2017. After set-pieces (SO2) and after open-play SO1 a significant
amount of goals were scored through volley rebounds. The high total amount of 2% of all goals, even
surprised the experts. Training shot techniques is a crucial part in professional football, and our analysis
can help to identify the right shooting situations to focus on.
In the following we describe four exemplary use cases of how the insights can support analysts and
coaches in their everyday business:

Use case 1—Automatize and objectify the match-analysts weekly processes: Nowadays, spend-
ing vast amounts of time and resources to perform pre-match-analyses of the next opponents and on the
post-match-analyses of the own performance has become an integral part of professional football. In a
well established process, match-analysts spend hours observing video footage of their upcoming opponent
to figure out what to expect. One of the most crucial questions they need to answer is: how does the

9
Figure 6: Selected special goals. Top left: shot chart for S2. Bottom right: aggregation of extraordinary
shots (circles), volleys (squares), and headers (triangles) from different clusters: shots are
plotted as circles, volleys as quarters and headers as triangles - all in the respective cluster-
color.

opponent score and concede goals? Typically, time constraints allow only to examine the last few goals
from the opponent which are then classified into one of few categories. These categories vary from club
to club, but due to the sample size, the analysis is coarse and expressivity is limited. By contrast, our
fully automated and purely data-driven approach processes arbitrarily long periods and as many goals as
desired and provides detailed clusters that allow for fine-grained analyses. Throw-in crosses, for example,
are rare events and thereafter hard to scout for each upcoming opponent. But some teams actively practice
throw-in crosses6,7 and our clustering automatically discloses whether an opponent uses them effectively.
Our analysis shows that almost half of the goals after long throw-ins are scored by only three teams in
our data-set (Union Berlin, Dynamo Dresden, MSV Duisburg). Our clustering also reveals teams with a
distinct counterpressing strategy (RB Leipzig scored twice as often with SP1 as the runner-up), teams
with dedicated cornertricks (Arminia Bielefeld with several goals in C5), and especially successful teams
after kick and rush plays K&R (TSG Hoffenheim, Bayer 04 Leverkusen and Fortuna Düsseldorf).

Use case 2—Scouting players: Scouting prospective players who will quickly adapt to a teams’ playing-
style or identifying a (near) equal substitute for a leaving or injured player is key to running a professional
club (Radicchi2016a; Pappalardo, 2019). While there already exist many different approaches using
event and/or tracking data, aiming to objectively evaluate players for scouting purposes like expected
goals (Anzer et al., 2021), space-control (Fernandez et al., 2018) or expected possession values (Spearman,
2018; Fernández et al., 2019), these typically only quantify a player’s output. By looking at patterns
instead of the pure outcome, our approach presents a possibility to identify players that not only produce
a high output, but do it in a way that fits a team stylistically.. Figure 9 shows the footprints of the two
famous strikers (Robert Lewandowski and Timo Werner) where line widths are proportional to the number
of scored goals in the respective branch of the tree. To evaluate whether one could substitute another,
we let the data speak and compare their scoring footprints. Since both are strikers, their performance is
measured to a high degree by the number of goals scored per match and the data-driven footprints reveal
whether they score their goals in similar fashions. Another non-trivial aspect in scouting players is to
identify promising talents. If, for instance, a technically skilled talent is needed, an aggregated view on
clusters SA1, SA2 and BE is helpful as they solely contain goals that require technically skilled players
to score. Another data-driven approach to quantify fingerprints of players is presented in Marcelino et al.
(2020). They analyze the directional correlations between players’ movements and find that players whose
movement correlates more strongly with their teammates’ tend to have higher market values. Following
Marcelino et al., 2020 in future studies one could evaluate the connection between players’ footprints and
their market value.

Use case 3—Long term team analysis: Analyzing the dendrogram allows to join clusters that encode
semantically similar goals. As an example consider clusters SP1 and LO2. The former contains classical
counterattacks where the ball is gained and quickly carried forward with determined passes. The latter,
located at a very different branch of the dendrogram, contains similar situations but the (unintentional)
6 https://www.bbc.com/sport/football/46312234, accessed 06/28/2020
7 https://trainingground.guru/articles/leeds-hire-set-piece-specialist-gianni-vio, accessed
06/28/2020

10
Figure 7: The largest 14 in-play clusters shown by three heatmaps each: start of the ball possession phase
(bottom), the assist location (middle) such as the shot location (top).

assist comes from the opponent. Merging the clusters allows to reason about counterattacks in general.
A very traditional category found by our clustering are one-two’s OT. While this pattern is a basic ele-
ment of football, its scarcity leading up to goals (0.8% of all goals) meant, it has not been investigated
by any scientific study. While one-two’s are very effective against men-oriented defensive structures, their
relevance in today’s top leagues seems to shrink significantly. However, are able to detect teams or players
using this strategy more frequently.

Use case 4—Scouting coaches: Moreover, the clustering allows to shed light on many very different
aspects of teams such as the effects of replacing head coaches. While selecting the right coach is a crucial
decision for any club, doing so while making use of positional or event data to support this decision has
not been addressed in literature. Just like players, coaches leave their own footprint in the dendrogram.
Analyzing this footprint can be a massive support when identifying head-coaches with a playing style
suiting their potential new team. Several studies investigated the effect coaching changes had on team
results (Kattuman et al., 2019; Besters et al., 2016), but our method aims to show before a possible change
how a coach would fit stylistically.
Professional football is highly affected by competitional pressure and emotions. Having an objective
and unbiased view on a team’s performance is indispensable for long-term success. By following a purely
data-driven approach, our contribution allows for such an unbiased view on the origin of goals. In order
to overcome biases towards established patterns, we present an exploratory way of analyzing goal scoring
patterns. In future research, the gained insights can be build upon to train supervised machine learning
models that automatically classify goals into pre-defined classes depending on individual club philosophies.
The possibility of (partially) automating regular tasks (e.g. weekly opponent analysis) not only allows to
save time but also to put the human focus on more sophisticated analyses and leave the easy tasks to number
crunching machines. Compared to human analyses, the proposed clustering offers a finer granularity and,
hence, provides a deeper understanding of the origin of goals.
The resulting clustering tree was analyzed and sanity checked by professional match-analysts from
national teams and Bundesliga clubs. The interdisciplinary cooperation with domain experts was of utmost
importance to the project to bridge the gap between computer and sports science and practice (see also
Goes et al., 2020; Herold et al., 2019; Rein et al., 2016). Combining expert opinions with statistical
evaluations (i.e. the Silhouette value of the clustering) turned out to be very beneficial for determining
the number of clusters.
When naming the clusters and discussing the ideal cluster number, in most cases the experts imme-
diately agreed and in the few remaining cases after a brief discussion a conses was found. Nevertheless, a
more systematic evaluation would be desirable for future studies. Since many of the categorical features are

11
Figure 8: Visualizations of goals assisted by the opponent. The 278 goals in this category present 7.8%
of all goals.

derived from manually annotated event data, further studies could also analyze the inter-labeler reliability
of the features. Furthermore, a general limitation of an unsupervised learning task is, that the resulting
clusters are not guaranteed to to make the distinctions a human would make. While, we used experts
opinions to guide us to find the right clustering, and the results satisfied their expectations, it could be of
future interest, to investigate how closely this unsupervised clustering matches an experts clustering.
Besides the reliability of the event data, an improvement of the tracking data quality (e.g. through
limb tracking), could open avenues for even more granular analysis of goals. And, while the data set used
for this study is already one of the largest in the literature, increasing the number of considered goals
would certainly further increase the usefulness of this work (e.g. by identifying very rare types of goals,
like direct corner kick goals). As mentioned earlier, we are excluding own-goals from this analysis, but
investigating how they originate, and what they have in common with "typical" goals could be another
area to explore further in the future.

Figure 9: Footprints of Robert Lewandowski (left) and Timo Werner (right) in the dendrogram.

5 Conclusions
We studied the origin of goals in the German Bundesliga and 2nd Bundesliga. We proposed a rich set of
features that can be extracted from synchronized tracking and event data. The feature representations of
the goals were then processed by an agglomerative clustering algorithm. Using two entire seasons of data,
we showed that the clustering allowed for fine grained differentiations and non-obvious insights that are
approved by professional match-analysts working for national teams and Bundesliga clubs. Our approach
can support professionals in their daily work and renders manual inspection of large amounts of video
footage unnecessary. Moreover, the proposed clustering can objectify pivotal decision making and offers
quantitative solutions to traditionally qualitative domains like scouting players or coaches or analyzing the

12
next opponent.

6 Acknowledgements
This work would not have been possible without the perspective of professional match-analysts from world
class teams who helped us to define relevant features and spend much time evaluating (intermediate) re-
sults. We would cordially like to thank Dr. Stephan Nopp and Christofer Clemens (head match-analysts of
the German mens National team), Jannis Scheibe (head match-analyst of the German U21 mens National
team) as well as Sebastian Geißler (former match-analyst of Borussia Mönchengladbach). Additionally,
the authors would like to thank Dr. Hendrik Weber and Deutsche Fußball Liga

Disclosure Statement The authors report no conflict of interest.

Ethics and Data Sharing By informing all participating players, all tracking is compliant to the
general data protection regulation (GDPR)8 . An ethics approval for wider research program using the
respective data is authorized by the ethics committee of the Faculty of Economics and Social Sciences at
the University of Tübingen. In order to respect the player’s and club’s sensitive information, the data
cannot be shared public.

Additional Material (Confidential) We provide a video with representative goals for each cluster.9

References
Andrienko, Gennady et al. (2017). “Visual analysis of pressure in football”. In: Data Mining and
Knowledge Discovery 31.6, pp. 1793–1839. issn: 1573756X. doi: 10.1007/s10618-017-0513-2
(cit. on pp. 3, 18, 19).
Andrienko, Gennady et al. (2019). “Constructing Spaces and Times for Tactical Analysis in Foot-
ball”. In: IEEE Transactions on Visualization and Computer Graphics, pp. 1–1. issn: 1077-
2626. doi: 10.1109/tvcg.2019.2952129. url: https://ieeexplore.ieee.org/document/
8894420/ (cit. on p. 2).
Anzer, Gabriel & Pascal Bauer (2021). “A Goal Scoring Probability Model based on Synchronized
Positional and Event Data”. In: Frontiers in Sports and Active Learning (Special Issue: Using
Artificial Intelligence to Enhance Sport Performance) 3.0, pp. 1–18. doi: 10.3389/fspor.2021.
624475. url: https://www.frontiersin.org/articles/10.3389/fspor.2021.624475/full
(cit. on pp. 2, 3, 10, 18, 19).
Armatas, Vasilios, Athanasios Yiannakos, & Dimitris Hatzimanouil (2007). “Record and evaluation
of set-plays in european football championship in Portugal 2004”. In: Inquiries in Sport and
Physical Education (cit. on p. 1).
Bauer, Pascal & Gabriel Anzer (2021). “Data-driven detection of counterpressing in professional
football—A supervised machine learning task based on synchronized positional and event
data with expert-based feature extraction”. In: Data Mining and Knowledge Discovery 35.5,
pp. 2009–2049. issn: 1573-756X. doi: 10.1007/s10618-021-00763-7. url: https://link.
springer.com/article/10.1007/s10618-021-00763-7 (cit. on p. 9).
Besters, Lucas M., Jan C. van Ours, & Martin A. van Tuijl (2016). “Effectiveness of In-Season
Manager Changes in English Premier League Football”. In: Economist (Netherlands) 164.3,
pp. 335–356. issn: 15729982. doi: 10.1007/s10645-016-9277-0 (cit. on p. 11).
Brian, S (2011). Cluster analysis Brian S. Everitt ... [et al.] John Wiley & Sons, XII, 330 p. ill.
isbn: 978-0-470-74991-3 (cit. on p. 4).
Carling, Christopher, A. Mark Williams, & Thomas Reilly (2006). “Handbook of Soccer Match
Analysis: A Systematic Approach to Improving Performance”. In: Journal of Sports Science &
Medicine. url: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3818670/ (cit. on p. 1).
8 https://gdpr-info.eu/
9 https://bit.ly/2NAXQcW.

13
Casal, C. A. et al. (2017). “Influence of match status on corner kicks tactics in elite soccer”. In:
Revista Internacional de Medicina y Ciencias de la Actividad Fisica y del Deporte. issn: 1577-
0354. doi: 10.15366/rimcafd2017.68.009. url: http://cdeporte.rediris.es/revista/
revista68/artinfluencia851e.pdf (cit. on p. 1).
Casal, Claudio A. et al. (2015). “Analysis of corner kick success in elite football”. In: International
Journal of Performance Analysis in Sport 15.2, pp. 430–451. issn: 14748185. doi: 10.1080/
24748668.2015.11868805 (cit. on p. 1).
Castellano, Julen, David Casamichana, & Carlos Lago (2012). “Accepted for printing in”. In: Jour-
nal of Human Kinetics 31, pp. 139–147. doi: 10 . 2478 / v10078 - 012 - 0015 - 7. url: http :
//fifa.com/worldcup/index.html (cit. on p. 2).
Chawla, Sanjay et al. (2017). “Classification of passes in football matches using spatiotemporal
data”. In: ACM Transactions on Spatial Algorithms and Systems 3.2. issn: 23740361. doi:
10.1145/3105576 (cit. on p. 2).
Collet, Christian (2013). “The possession game? A comparative analysis of ball retention and team
success in European and international football, 2007-2010”. In: Journal of Sports Sciences 31.2,
pp. 123–136. issn: 02640414. doi: 10.1080/02640414.2012.727455 (cit. on p. 9).
Defays, D. (2015). “An efficient algorithm for a complete link method”. In: url: https://academ
ic.oup.com/comjnl/article-abstract/20/4/364/393966 (cit. on p. 4).
Delgado-Bordonau, Juan Luis et al. (2013). “Offensive and defensive team performance: Relation to
successful and unsuccessful participation in the 2010 Soccer World Cup”. In: Journal of Human
Sport and Exercise 8.4, pp. 894–904. issn: 19885202. doi: 10.4100/jhse.2013.84.02 (cit. on
p. 2).
Dufour, Michel, John Phillips, & Viviane Ernwein (2017). “What makes the difference? Analysis
of the 2014 World Cup”. In: Journal of Human Sport and Exercise 12.3, pp. 616–629. issn:
19885202. doi: 10.14198/jhse.2017.123.06 (cit. on p. 2).
Erčulj, Frane & Erik Štrumbelj (2015). “Basketball shot types and shot success in different levels of
competitive basketball”. In: PLoS ONE 10.6, pp. 1–14. issn: 19326203. doi: 10.1371/journal.
pone.0128885 (cit. on p. 2).
Felsenstein, Joseph (1996). “[24] Inferring phylogenies from protein sequences by parsimony, dis-
tance, and likelihood methods”. In: Methods in Enzymology 266, pp. 418–427. issn: 00766879.
doi: 10.1016/s0076-6879(96)66026-1 (cit. on p. 4).
Fernández-Hermógenes, Daniel, Oleguer Camerino, & Antonio García De Alcaraz (2017). “Set-
piece offensive plays in soccer”. In: issn: 2014-0983. doi: 10.5672/apunts.2014- 0983.es.
(2017/3).129.06. url: https://core.ac.uk/download/pdf/132357632.pdf (cit. on p. 1).
Fernandez, Javier & Luke Bornn (2018). “Wide Open Spaces : A statistical technique for measuring
space creation in professional soccer”. In: MIT Sloan Sports Analytics Conference, pp. 1–19 (cit.
on p. 10).
Fernández, Javier, Luke Bornn, & Dan Cervone (2019). “Decomposing the Immeasurable Sport: A
deep learning expected possession value framework for soccer”. In: MIT Sloan Sports Analytics
Conference, Boston (USA), pp. 1–18. url: https://lukebornn.com/sloan_epv_curve.mp4
(cit. on p. 10).
Fernando, T et al. (2015). “Discovering Methods of Scoring in Soccer Using Tracking Data”. In:
KDD Workshop on Large-Scale Sports Analytics, pp. 1–4. url: https://large-scale-sport
s-analytics.org/Large-Scale-Sports-Analytics/Submissions2015_files/paperID19-
Tharindu.pdf (cit. on p. 2).
Goes, F R et al. (2020). “Unlocking the Potential of Big Data to Support Tactical Performance
Analysis in Professional Soccer: A Systematic Review”. In: European Journal of Sport Science
0.0, pp. 1–16. issn: 1746-1391. doi: 10.1080/17461391.2020.1747552. url: https://doi.
org/10.1080/17461391.2020.1747552 (cit. on pp. 2, 11).
González-Ródenas, Joaquin et al. (2019). “Technical, tactical and spatial indicators related to goal
scoring in European elite soccer”. In: Journal of Human Sport and Exercise. issn: 1988-5202.
doi: 10.14198/jhse.2020.151.17 (cit. on pp. 1, 9).
Göral, Kemal (2019). “The importance of set-pieces in soccer : Russia 2018 FIFA World Cup
analysis Futbolda duran t opların ö nemi : Rusya 2018 FIFA Dünya Kupasının analizi”. In:
16.3. doi: 10.14687/jhs.v16i3.5758 (cit. on p. 1).

14
Herold, Mat et al. (2019). “Machine learning in men’s professional football: Current applications
and future directions for improving attacking play”. In: International Journal of Sports Science
& Coaching, p. 1747954119879350. issn: 1747-9541. doi: 10.1177/1747954119879350. url:
https://doi.org/10.1177/1747954119879350 (cit. on pp. 2, 11).
Hobbs, Jennifer et al. (2018). “Quantifying the Value of Transitions in Soccer via Spatiotemporal
Trajectory Clustering”. In: pp. 1–11 (cit. on pp. 2, 9).
Kattuman, Paul, Christoph Loch, & Charlotte Kurchian (2019). “Management succession and
success in a professional soccer team”. In: PLoS ONE 14.3, pp. 1–20. issn: 19326203. doi:
10.1371/journal.pone.0212634 (cit. on p. 11).
Linke, Daniel, Daniel Link, & Martin Lames (2020). “Football-specific validity of TRACAB’s op-
tical video tracking systems”. In: PLoS ONE 15.3, pp. 1–17. issn: 19326203. doi: 10.1371/
journal.pone.0230179. url: http://dx.doi.org/10.1371/journal.pone.0230179 (cit. on
p. 3).
López, F. A., J. A. Martínez, & M. Ruiz (2013). “Spatial pattern analysis of shot attempts in
basketball; The case of L.A. Lakers”. In: Revista Internacional de Medicina y Ciencias de la
Actividad Fisica y del Deporte 13.51, pp. 585–613. issn: 1577-0354 (cit. on p. 2).
Lucey, Patrick et al. (2014). “"Quality vs Quantity": Improved Shot Prediction in Soccer using
Strategic Features from Spatiotemporal Data”. In: Proc. 8th Annual MIT Sloan Sports Analytics
Conference, pp. 1–9. url: http://www.sloansportsconference.com/?p=15790 (cit. on p. 2).
Marcelino, Rui et al. (2020). “Collective movement analysis reveals coordination tactics of team
players in football matches”. In: Chaos, Solitons and Fractals 138, p. 109831. issn: 09600779.
doi: 10.1016/j.chaos.2020.109831. url: https://doi.org/10.1016/j.chaos.2020.
109831 (cit. on pp. 2, 10).
Merlin, Murilo et al. (2020). “Exploring the determinants of success in different clusters of ball pos-
session sequences in soccer”. In: Research in Sports Medicine 28.3, pp. 339–350. issn: 15438635.
doi: 10.1080/15438627.2020.1716228 (cit. on pp. 2, 3).
Mitrotasios, Michalis & Vasilis Armatas (2012). “Analysis of goal scoring patterns in the 2012
European Football Championship”. In: The Sport Journal 50, pp. 1–9. issn: 15439518. url:
http://thesportjournal.org/article/analysis-of-goal-scoring-patterns-in-the-
2012-european-football-championship/ (cit. on pp. 1, 8).
Murtagh, Fionn & Pedro Contreras (2017). Algorithms for Hierarchical Clustering: An Overview,
II. Tech. rep. (cit. on pp. 2, 4).
Njororai, W. W.S. (2013). “Analysis of goals scored in the 2010 world cup soccer tournament held
in South Africa”. In: Journal of Physical Education and Sport. issn: 22478051. doi: 10.7752/
jpes . 2013 . 01002. url: https : / / scholarworks . uttyler . edu / cgi / viewcontent . cgi ?
referer=https://scholar.google.de/scholar?hl=de&as_sdt=0%2C5&q=Analysis+of+
goals+scored+in+the+2010+world+cup+soccer+tournament+held+in+South+Africa&
btnG=&httpsredir=1&article=1008&context=hkdept_fac (cit. on pp. 1, 2, 8).
Pappalardo, Luca (2019). “Explainable Injury Forecasting in Soccer via Multivariate Time Series
and Convolutional Neural Networks”. In: Barça sports analytics summit, Barcelona (Spain),
pp. 1–15. url: https://static.capabiliaserver.com/frontend/clients/barca/wp_prod/
wp-content/uploads/2020/01/c6658839-paper-format-luca-pappalardo-1.pdf (cit. on
p. 10).
Plummer, B T (2013). “Analysis of Attacking Possessions Leading to a Goal Attempt, and Goal
Scoring Patterns within Men’s Elite Soccer”. In: Journal of Sports Science (cit. on p. 1).
Pollard, Richard & Charles Reep (1997). “Measuring the effectiveness of playing strategies at
soccer”. In: Journal of the Royal Statistical Society Series D: The Statistician. issn: 00390526.
doi: 10.1111/1467-9884.00108 (cit. on p. 2).
Power, Paul et al. (2017). “Not all passes are created equal: Objectively measuring the riskt and
reward of passes in soccer from tracking data”. In: Proceedings of the 23rd ACM SIGKDD
International Conference on Knowledge Discovery and Data Mining, pp. 1605–1613. doi: 10.
1145/3097983.3098051. url: http://doi.acm.org/10.1145/3097983.3098051 (cit. on
pp. 1, 9).
Pulling, Craig (2015). “Long corner kicks in the English premier league: Deliveries into the goal
area and critical area”. In: Kinesiology 47.2, pp. 193–201. issn: 13311441 (cit. on p. 1).

15
Pulling, Craig, Matthew Robins, & Thomas Rixon (2013). “Defending corner kicks: Analysis from
the English premier league”. In: International Journal of Performance Analysis in Sport. issn:
14748185. doi: 10.1080/24748668.2013.11868637 (cit. on p. 1).
Rathke, Alex (2017). “An examination of expected goals and shot efficiency in soccer”. In: Journal
of Human Sport and Exercise 12.Proc2. issn: 1988-5202. doi: 10 . 14198 / jhse . 2017 . 12 .
proc2.05. url: http://www.redalyc.org/articulo.oa?id=301052437005 (cit. on p. 2).
Reep, Charles & B. Benjamin (1968). “Skill and Chance in Association Football”. In: Journal of the
Royal Statistical Society. Series A (General). issn: 00359238. doi: 10.2307/2343726 (cit. on
pp. 1, 8, 9).
Reich, Brian J. et al. (2006). “A spatial analysis of basketball shot chart data”. In: American
Statistician 60.1, pp. 3–12. issn: 00031305. doi: 10.1198/000313006X90305 (cit. on p. 2).
Rein, Robert & Daniel Memmert (2016). “Big data and tactical analysis in elite soccer: future
challenges and opportunities for sports science”. In: SpringerPlus 5.1. issn: 21931801. doi:
10.1186/s40064-016-3108-2 (cit. on pp. 2, 11).
Robberechts, Pieter & Jesse Davis (2020). “How data availability affects the ability to learn good
xG models”. In: Communications in Computer and Information Science, Springer, Cham 1324,
pp. 17–27. issn: 18650937. doi: 10.1007/978-3-030-64912-8 (cit. on p. 2).
Rousseeuw, Peter J. (1987). “Silhouettes: A graphical aid to the interpretation and validation of
cluster analysis”. In: Journal of Computational and Applied Mathematics 20.C, pp. 53–65. issn:
03770427. doi: 10.1016/0377-0427(87)90125-7 (cit. on p. 4).
Ruiz, H. et al. (2015). “Measuring scoring efficiency through goal expectancy estimation”. In: 23rd
European Symposium on Artificial Neural Networks, Computational Intelligence and Machine
Learning, ESANN 2015 - Proceedings April, pp. 149–154 (cit. on p. 2).
Ruiz, Hector et al. (2017). “"The Leicester City Fairytale?": Utilizing New Soccer Analytics Tools
to Compare Performance in the 15/16 &#38; 16/17 EPL Seasons”. In: Proceedings of the 23rd
ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pp. 1991–
2000. doi: 10 . 1145 / 3097983 . 3098121. url: http : / / doi . acm . org / 10 . 1145 / 3097983 .
3098121%0Ahttp://dl.acm.org/citation.cfm?doid=3097983.3098121 (cit. on p. 2).
Salmon, Paul M. & Scott McLean (2020). “Complexity in the beautiful game: implications for
football research and practice”. In: Science and Medicine in Football 4.2, pp. 162–167. issn:
24734446. doi: 10.1080/24733938.2019.1699247 (cit. on p. 2).
Santos, Alejandro Benito et al. (2018). “Data-driven visual performance analysis in soccer: An
exploratory prototype”. In: Frontiers in Psychology 9.DEC. issn: 16641078. doi: 10 . 3389 /
fpsyg.2018.02416 (cit. on pp. 3, 18).
Sarmento, Hugo et al. (2014). “Patterns of play in the counterattack of elite football teams - A
mixed method approach”. In: International Journal of Performance Analysis in Sport. issn:
14748185. doi: 10.1080/24748668.2014.11868731 (cit. on pp. 2, 9).
Schmicker, Robert H. (2013). “An application of satscan to evaluate the spatial distribution of
corner kick goals in major league soccer”. In: International Journal of Computer Science in
Sport. issn: 16844769. url: https://pdfs.semanticscholar.org/a36c/694c79c3d38d19baf
9d01a3677834289b340.pdf (cit. on p. 1).
Schulze, Emiel et al. (2018). “Effects of positional variables on shooting outcome in elite football”.
In: Science and Medicine in Football 2.2, pp. 93–100. issn: 24734446. doi: 10.1080/24733938.
2017.1383628. url: https://doi.org/10.1080/24733938.2017.1383628 (cit. on p. 2).
Sibson, R. (1973). “SLINK: An optimally efficient algorithm for the single-link cluster method”.
In: The Computer Journal 16.1, pp. 30–34. issn: 0010-4620. doi: 10.1093/comjnl/16.1.30
(cit. on p. 4).
Siegle, Malte & Martin Lames (2013). “Modeling soccer by means of relative phase”. In: Journal of
Systems Science and Complexity 26.1, pp. 14–20. issn: 15597067. doi: 10.1007/s11424-013-
2283-2 (cit. on p. 2).
Sokal, C.D. Michener (1958). A statistical method for evaluating systematic relationships. url:
https : / / archive . org / details / cbarchive _ 33927 _ astatisticalmethodforevaluatin
1902 / page / n2 / mode / 2uphttp : / / archive . org / details / cbarchive _ 33927 _ astatistic
almethodforevaluatin1902 (cit. on p. 4).

16
Spearman, William (2018). “Beyond Expected Goals”. In: MIT Sloan Sports Analytics Conference,
Boston (USA), pp. 1–17. url: https://www.researchgate.net/publication/327139841
(cit. on p. 10).
Steiner, Silvan et al. (2019). “Outplaying opponents—a differential perspective on passes using
position data”. In: German Journal of Exercise and Sport Research February. issn: 2509-3142.
doi: 10.1007/s12662-019-00579-0 (cit. on p. 3).
Szwarc, Andrzej (2007). “Efficacy of successful and unsuccessful soccer teams taking part in finals
of Champions League”. In: Research Yearbook 13.2, pp. 221–225. url: http : / / journals .
indexcopernicus.com/abstracted.php?icid=838944 (cit. on p. 1).
Taberner, Matt et al. (2020). “Interchangeability of position tracking technologies; can we merge
the data?” In: Science and Medicine in Football 4.1, pp. 76–81. issn: 24734446. doi: 10.1080/
24733938.2019.1634279. url: https://doi.org/10.1080/24733938.2019.1634279 (cit. on
p. 3).
Taylor, Joseph B, Nic James, & Stephen D Mellalieu (2005). “Notational Analysis of Corner Kicks
in English Premier League Soccer”. In: Science and Football V. url: https://books.google.
de/books?hl=de&lr=&id=nyFr-2uwPGoC&oi=fnd&pg=PA229&dq=Notational+Analysis+of+
Corner+Kicks+in+English+Premier+League+Soccer&ots=DYs2PvFfk_&sig=W- i76h0Zxw_
omd2Ty18Nu-giF6w#v=onepage&q=NotationalAnalysisofCornerKicksinEnglishPremi (cit.
on p. 1).
Tenga, Albin et al. (2010). “Effect of playing tactics on goal scoring in norwegian professional
soccer”. In: Journal of Sports Sciences 28.3, pp. 237–244. issn: 1466447X. doi: 10 . 1080 /
02640410903502774 (cit. on p. 2).
Vogelbein, Martin, Stephan Nopp, & Anita Hökelmann (2014). “Defensive transition in soccer
- are prompt possession regains a measure of success? A quantitative analysis of German
Fußball-Bundesliga 2010/2011”. In: Journal of Sports Sciences. issn: 1466447X. doi: 10.1080/
02640414.2013.879671 (cit. on p. 9).
Witts, James (2019). “Training secrets of the world’s greatest footballers : how science is transform-
ing the modern game”. In: Bloomsbury Publishing PLC. url: https://www.bookdepository.
com / Training - Secrets - Worlds - Greatest - Footballers - James - Witts / 9781472948458
(cit. on p. 1).
Wold, Svante, Kim Esbensen, & Paul Geladi (1987). “Principal component analysis”. In: Chemomet-
rics and Intelligent Laboratory Systems 2.1-3, pp. 37–52. issn: 01697439. doi: 10.1016/0169-
7439(87)80084-9 (cit. on p. 4).

Appendix
Appendix A (Tables 1, 2) detail the features that are extracted from positional and event data for the
clustering in detail. Appendix B (Tables 3, 4) provide details on goal scoring and receiving patterns on
a club level and may be of interest to analysts of the respective teams. Similarly, Appendix C (Table 5)
shows the conversion rates of every cluster. Finally, Appendix D (Table 6) contains representative goals
for selected clusters. Interested analysts may use these goals to evaluate the clustering on their own.

17
A Appendix

Table 1: Features describing the ball possession phase prior to a goal.

Feature Value Description

Start Action Categorical Describes the start of the ball possession phase in pre-defined abstraction
levels (Own Half, Offensive Ball Gain, Throw in, corner kick, Free kick,
Penalty).
Build-up Categorical Describing the build-upleading up to the goal (crossOpenPlay, pass open
play, free kick, Penalty, corner kick, throw in, loss Of Possession).
Location of ball Numeric x- and y-coordinate of a shot. The synchronized location from positional
possession start and event data as described in Anzer et al., 2021 is used.
Set-up Origin Categorical Describes where the build-up play for the shot at goal starts (inside, out-
side).
Duration ball pos- Numeric Length of the ball possession phase measured in [s]. The start of a ball
session phase possession phase is either a dead ball, or an open-play turnover. Interrup-
tions where the opposing team gains possession of the ball for less than
six consecutive seconds do not end an possession phase.
Number of passes Numeric Number of completed passes during ball possession phase.
Number of oppos- Numeric Number of opposing touches during ball possession phase.
ing touches
Bypassed players Numeric Bypassed players is defined as the positive difference between the number
of players that are closer to their own goal than the ball at the time of the
shot and when the ball possession started.
Meters dribbled Numeric Meters dribbled during ball possession phase. This feature is calculated as
the sum of all the euclidean distances between starting and end location
of each player’s possessions.
Meters passed Numeric Meters passed during ball possession phase.
Average passing Numeric Average amount of pressure passing players received during the ball pos-
pressure session phase at the moment they played a pass according Andrienko et al.,
2017.
Average receiving Numeric Average amount of pressure pass receiving players received during the ball
pressure possession phase. at the moment they received a pass according Andrienko
et al., 2017.
Counterattack Categorical Describes whether the build-up was a counterattack. Counterattacks are
defined in the official manually collected event data as attacks during which
a team gains ball control in its own half, immediately starts a quick coun-
terattack and takes a shot within at maximum 14 seconds.
Number of oppos- Numerical Counts the amount of uncontrolled touches the opposing team had during
ing touches a possession.
Maximum vertical Numeric The longest vertical distance of any pass within the possession.
pass length
Maximum horizon- Numeric The longest horizontal distance of any pass within the possession.
tal pass length
Maximum pass Numeric The distance of the longest pass within the possession.
length
Compactness Ball- Numeric Compactness of the attacking team (Santos et al., 2018) at the beginning
gain of the ball possession phase.
Compactness Shot Numeric Compactness of the attacking team (Santos et al., 2018) at the time of the
shot

18
Table 2: Features describing the assist and shot setup.

Feature Value Description

Shot location Numeric X and Y coordinate of a shot
Type of shot Categorical Describing the body part used for the shot (head or leg)
Chance evaluation Categorical Classifying the quality of a chance (chance, sitter)
Taker Ball-control Categorical Ball-control type describes the type of control the shot taker had prior to
scoring. It includes the following categories:
• “Direct” a shot with the first touch, unless the shot is considered a
volley.
• “Volley” a shot with the first touch and the ball did not touch the
ground previously.
• “Control - shot” a shot followed after a single touch to control the
ball.

• “Distance covered < 10m” a shot following a short dribble (less than
10 meters).
• “Distance covered > 10m” a shot following a longer dribble (more
than 10 meters).

• “Set piece taker” a direct set-piece shot.

Setup Categorical Describes how the person taking the shot was set up (header, long pass
from open play, other pass from open play, one two, cross from open play,
shot, free kick; corner kick, throw in, teammate action, rebound wood-
work).
xG Numeric The “Expected Goal” (xG) value of a shot according to Anzer et al., 2021.
Distance to goal Numeric Distance in meters between the location of the shot and the center of the
opposing goal.
Goal angle Numeric Angle in radians between the location of the shot and the two posts of the
opposing goal.
Speed of player tak- Numeric The speed in [km/h] the player attempting the shot was travelling at the
ing the shot time of the shot.
Pressure on the Numeric The amount of pressure the player attempting the shot was under at the
player taking the time of the shot according to Andrienko et al. (2017).
shot
Defenders in the Numeric The number of defenders in the line of the shot, defined as the triangle
line of the shot between the shot location and the two goal posts.
Distance of the Numeric The distance in meters the goalkeeper between the goalkeeper and the
goalkeeper to the center of the goal at the time of the shot.
goal
Goalkeeper in the Numeric Describes whether the goalkeeper is in the line of the shot or not.
line of the shot
Solo Categorical Solo indicates that a remarkable individual contribution (=solo) by the
goalscorer lead to the successful shot.
Assist location Numeric X and Y coordinate of the assist
After free kick Categorical Indicates whether the goal followed a freekick.
Assist type Categorical Describing whether it was a direct, indirect assist or not assisted
Assist action Categorical Describing the assist action (e.g. “long pass”)

19
B Appendix

Table 3: Scored goals per team.

20
Table 4: Received goals per team.

21
C Appendix

Table 5: Conversion rate of goals per shot. Values above (> 20%) and below (< 5%) average are indicated
by green and red arrows, respectively.

D Appendix

22
Table 6: Exemplary goals for selected clusters.

Cluster Season(League) Pairing Scoring Team Goal Scorer Assist Minute

S2 2018/2019(1) FC Augsburg:Hannover 96 Augsburg Schmid NaN 0
S2 2018/2019(2) SpVgg Greuther Fürth:FC Erzgebirge Aue Aue Hochscheidt Krüger 0
S2 2017/2018(2) Fortuna Düsseldorf:SpVgg Greuther Fürth Fürth Wittek Narey 0
S2 2018/2019(1) Sport-Club Freiburg:FC Augsburg Freiburg Grifo Grifo 0
BE 2018/2019(1) Sport-Club Freiburg:Borussia Mönchengladbach Freiburg Höler Sommer 90
BE 2018/2019(1) Eintracht Frankfurt:Sport-Club Freiburg Frankfurt Jovic Jovic 45
LP1 2017/2018(1) TSG 1899 Hoffenheim:Borussia Dortmund Dortmund Reus Guerreiro 58
LP1 2017/2018(1) FC Schalke 04:FC Bayern München Bayern Vidal Pardo Rodríguez Rubio 75
LP2 2017/2018(1) Borussia Mönchengladbach:Hamburger SV M’gladbach Hazard Caetano de Araújo 9
LP5 2018/2019(1) Sport-Club Freiburg:Borussia Mönchengladbach Freiburg Waldschmidt Haberer 57
LP5 2017/2018(1) TSG 1899 Hoffenheim:Bayer 04 Leverkusen Leverkusen Alario Bailey Butler 70
SP1 2017/2018(2) MSV Duisburg:Fortuna Düsseldorf Düsseldorf Hennings Fink 40
SP1 2018/2019(2) SpVgg Greuther Fürth:Holstein Kiel Fürth Green Dona Atanga 90
SP5 2017/2018(1) 1. FSV Mainz 05:Sport-Club Freiburg Mainz De Blasis Quaison 79
SP5 2017/2018(1) TSG 1899 Hoffenheim:Hannover 96 Hoffenheim Kramaric Gnabry 16
HV1 2018/2019(2) FC St. Pauli:SSV Jahn Regensburg St. Pauli Flum Carstens 52
HV1 2017/2018(1) RB Leipzig:Hannover 96 Leipzig Werner Forsberg 85
HV2 2017/2018(2) MSV Duisburg:SSV Jahn Regensburg Duisburg Nauber Tashchy 52
HV2 2018/2019(1) Fortuna Düsseldorf 1895 e.V.:FC Augsburg Augsburg Hahn Richter 76
HV2 2018/2019(2) 1. FC Union Berlin:1. FC Heidenheim 1846 Union Berlin Gikiewicz Andersson 90
HV2 2017/2018(1) FC Augsburg:TSG 1899 Hoffenheim Hoffenheim Kramaric Hübner 30
HV2 2018/2019(1) Borussia Mönchengladbach:SV Werder Bremen Bremen Klaassen Osako 79
HV2 2017/2018(1) Hertha BSC:FC Bayern München Bayern Hummels Boateng 10
HV2 2017/2018(2) FC Erzgebirge Aue:1. FC Nürnberg Aue Köpke Tiffert 77
HV2 2017/2018(2) Fortuna Düsseldorf:1. FC Heidenheim 1846 Heidenheim Verhoek Schnatterer 83
HV3 2017/2018(1) FC Bayern München:1. FSV Mainz 05 Bayern Lewandowski Kimmich 77
HV3 2018/2019(1) Borussia Dortmund:FC Bayern München Bayern Lewandowski Kimmich 52
HV3 2017/2018(1) RB Leipzig:FC Bayern München Bayern Wagner Rodríguez Rubio 12
SA1 2018/2019(1) FC Bayern München:Eintracht Frankfurt Bayern Ribéry Kimmich 72
SA1 2017/2018(1) TSG 1899 Hoffenheim:1. FC Köln Hoffenheim Gnabry Grillitsch 47
SA1 2017/2018(1) TSG 1899 Hoffenheim:RB Leipzig Hoffenheim Gnabry Amiri 62
K&R 2017/2018(2) 1. FC Nürnberg:FC St. Pauli St. Pauli Sobota Himmelmann 63
K&R 2017/2018(1) Sport-Club Freiburg:1. FSV Mainz 05 Mainz Berggreen Brosinski 90
OT 2018/2019(1) Eintracht Frankfurt:FC Bayern München Bayern Ribéry Kimmich 79
OT 2018/2019(1) 1. FC Nürnberg:Hertha BSC Berlin Ibisevic Selke 15
FC1 2017/2018(1) Borussia Dortmund:Eintracht Frankfurt Frankfurt Jovic de Guzmán 75
FC1 2017/2018(1) Eintracht Frankfurt:1. FC Köln Köln Terodde Risse 74
FC2 2017/2018(1) FC Augsburg:Eintracht Frankfurt Augsburg Koo Baier 19
FC2 2017/2018(1) TSG 1899 Hoffenheim:1. FSV Mainz 05 Hoffenheim Kramaric Uth 67
C1 2017/2018(2) SSV Jahn Regensburg:1. FC Heidenheim 1846 Regensburg George Lais 34
C1 2018/2019(1) FC Augsburg:Eintracht Frankfurt Augsburg Córdova Lezama da Silva 90
C3 2017/2018(1) Hamburger SV:Eintracht Frankfurt Hamburg Papadopoulos Hunt 9
C3 2018/2019(1) Hertha BSC:Eintracht Frankfurt Berlin Grujic Plattenhardt 40
C4 2017/2018(1) Bayer 04 Leverkusen:VfL Wolfsburg Leverkusen Bender Retsos 29
C4 2018/2019(1) FC Bayern München:Borussia Mönchengladbach M’gladbach Herrmann Kramer 88
C5 2017/2018(2) DSC Arminia Bielefeld:VfL Bochum 1848 Bielefeld Kerschbaumer Staude 35
C5 2018/2019(2) DSC Arminia Bielefeld:1. FC Heidenheim 1846 Bielefeld Schütz Hartherz 33
C5 2018/2019(1) Bayer 04 Leverkusen:TSG 1899 Hoffenheim Hoffenheim Nelson Grifo 19
SA 2018/2019(1) Bayer 04 Leverkusen:Eintracht Frankfurt Leverkusen Brandt Aránguiz Sandoval 13
HC 2018/2019(2) 1. FC Heidenheim 1846:SV Sandhausen Sandhausen Wooten Diekmeier 69
HC 2018/2019(1) Fortuna Düsseldorf 1895 e.V.:Eintracht Frankfurt Frankfurt Mendes Paciencia de Guzmán 48
TI 2017/2018(1) Hamburger SV:FC Schalke 04 Hamburg Kostic dos Santos Justino De Melo 17
TI 2017/2018(1) Bayer 04 Leverkusen:1. FC Köln Köln Guirassy Sørensen 23
TI 2018/2019(2) SG Dynamo Dresden:MSV Duisburg Dresden Röser Heise 39
UA1 2017/2018(2) SSV Jahn Regensburg:FC Erzgebirge Aue Aue Köpke Riese 57
UA1 2017/2018(1) Eintracht Frankfurt:SV Werder Bremen Frankfurt Rebic Willems 17
UA1 2017/2018(2) MSV Duisburg:Fortuna Düsseldorf Duisburg Tashchy Stoppelkamp 90
UA3 2018/2019(2) SSV Jahn Regensburg:SG Dynamo Dresden Dresden Dumic Koné 52
UA3 2017/2018(1) FC Bayern München:FC Augsburg Bayern Vidal Pardo Süle 31
LO1 2017/2018(1) VfB Stuttgart:Eintracht Frankfurt Stuttgart Thommy Ginczek 13
LO1 2018/2019(1) 1. FSV Mainz 05:Borussia Dortmund Mainz Quaison Hack 70
LO2 2017/2018(1) Borussia Dortmund:FC Augsburg Dortmund Reus Schürrle 16
LO2 2018/2019(2) FC Ingolstadt 04:Holstein Kiel Ingolstadt Lezcano Farina Kutschke 13
LO3 2017/2018(1) FC Bayern München:Eintracht Frankfurt Frankfurt Haller Vieira da Costa 78
LO3 2017/2018(1) FC Bayern München:Hannover 96 Bayern Coman Müller 67
LO4 2017/2018(2) 1. FC Kaiserslautern:SV Sandhausen Sandhausen Förster Linsmayer 78
LO4 2017/2018(1) Eintracht Frankfurt:FC Schalke 04 Schalke Aparecido Rodrigues Embolo 90
LO5 2017/2018(2) FC St. Pauli:FC Ingolstadt 04 Ingolstadt Träsch Pledl 33
LO5 2018/2019(2) Holstein Kiel:FC Erzgebirge Aue Aue Hochscheidt Iyoha 26
SO1 2017/2018(2) MSV Duisburg:VfL Bochum 1848 Duisburg Tashchy Bomheuer 7
SO1 2017/2018(1) VfL Wolfsburg:Borussia Mönchengladbach Wolfsburg Akoi Fara Guilavogui Gómez García 71
SO1 2017/2018(2) 1. FC Heidenheim 1846:FC St. Pauli Heidenheim Thiel Schnatterer 16
SO1 2018/2019(2) 1. FC Union Berlin:MSV Duisburg Duisburg Oliveira Souza Iljutcenko 77
SO2 2017/2018(2) SV Darmstadt 98:SG Dynamo Dresden Dresden Konrad Berko 80
SO2 2017/2018(2) MSV Duisburg:DSC Arminia Bielefeld Duisburg Wolze Stoppelkamp 72
SO2 2018/2019(2) MSV Duisburg:SC Paderborn 07 Duisburg Tashchy Wolze 63
SO2 2017/2018(2) Holstein Kiel:SV Sandhausen Sandhausen Klingmann Höler 35
SO2 2018/2019(1) Hertha BSC:TSG 1899 Hoffenheim Berlin Lazaro Plattenhardt 87
SO2 2017/2018(1) Hertha BSC:Borussia Mönchengladbach M’gladbach Caetano de Araújo Wendt 20
SO2 2017/2018(2) SG Dynamo Dresden:SV Sandhausen Sandhausen Paqarada Daghfous 25
SO2 2018/2019(2) SC Paderborn 07:1. FC Köln Paderborn Pröger Michel 86
SO2 2017/2018(2) FC Erzgebirge Aue:MSV Duisburg Aue Nazarov Tiffert 83
SO2 2017/2018(1) Hannover 96:FC Augsburg Hannover Sané Klaus 37
SO3 2018/2019(1) FC Augsburg:Bayer 04 Leverkusen Leverkusen Tah Brandt 60
SO3 2017/2018(1) FC Bayern München:Sport-Club Freiburg Bayern Coman Robben 42
SO5 2017/2018(2) Fortuna Düsseldorf:1. FC Heidenheim 1846 Düsseldorf Raman Hennings 90
SO5 2018/2019(1) FC Augsburg:Hannover 96 Hannover Weydandt Maina 8
SO5 2018/2019(1) Sport-Club Freiburg:1. FSV Mainz 05 Mainz Onisiwo Niakhaté 75
SO5 2017/2018(2) 1. FC Nürnberg:1. FC Heidenheim 1846 1. FC Nürnberg Stefaniak Ishak 38

23
C Appendix—Study III: Data-Driven Detection of
Counterpressing in Professional Football

In the following, Bauer and Anzer (2021) is reproduced with

permission from Springer.

120
Data Mining and Knowledge Discovery
https://doi.org/10.1007/s10618-021-00763-7

Data-driven detection of counterpressing in professional

football
A supervised machine learning task based on synchronized positional
and event data with expert-based feature extraction

Pascal Bauer1,2 · Gabriel Anzer1,3

Received: 17 August 2020 / Accepted: 4 May 2021

© The Author(s) 2021

Abstract
Detecting counterpressing is an important task for any professional match-analyst in
football (soccer), but is being done exclusively manually by observing video footage.
The purpose of this paper is not only to automatically identify this strategy, but also
to derive metrics that support coaches with the analysis of transition situations. Addi-
tionally, we want to infer objective influence factors for its success and assess the
validity of peer-created rules of thumb established in by practitioners. Based on a
combination of positional and event data we detect counterpressing situations as a
supervised machine learning task. Together, with professional match-analysis experts
we discussed and consolidated a consistent definition, extracted 134 features and man-
ually labeled more than 20, 000 defensive transition situations from 97 professional
football matches. The extreme gradient boosting model—with an area under the curve
of 87.4% on the labeled test data—enabled us to judge how quickly teams can win
the ball back with counterpressing strategies, how many shots they create or allow
immediately afterwards and to determine what the most important success drivers
are. We applied this automatic detection on all matches from six full seasons of the
German Bundesliga and quantified the defensive and offensive consequences when
applying counterpressing for each team. Automating the task saves analysts a tremen-

Responsible editor: Albrecht Zimmermann.

Pascal Bauer
pascal.bauer@dfb.de
Gabriel Anzer
gabriel.anzer@sportec-solutions.de

1 Department of Sport Psychology and Research Methods, Institute of Sports Science, University
of Tübingen, Tübingen, Germany
2 DFB-Akademie, Deutscher Fußball-Bund e.V. (DFB), Frankfurt, Germany
3 Sportec Solutions AG, subsidiary of the Deutsche Fußball Liga (DFL), Munich, Germany

123
 P. Bauer, G. Anzer

dous amount of time, standardizes the otherwise subjective task, and allows to identify
trends within larger data-sets. We present an effective way of how the detection and
the lessons learned from this investigation are integrated effectively into common
match-analysis processes.

Keywords Sports analytics · Football (Soccer) · Tactical performance analysis ·

Applied machine learning · Positional and event data

1 Introduction

Acquiring accurate and high frequency positional and event data is common in most
of the world’s top professional football (soccer) leagues. Manually annotated event
data provides information about the one player carrying the ball at the time of a game
relevant action only, whereas so called positional data can capture highly accurate
positions of all 22 players up to 25 times a second.
Every professional football team spends a substantial amount of time analyzing and
monitoring strategies such as counterpressing.—a complex team strategy for transi-
tion situations—of their own and opposing teams. Navarro and Javier (2018) defines
counterpressing as simple as “[..] pressure after losing the ball”. Related to this, Pep
Guardiola made the ’five second rule’ for counterpressing famous.1 Another coach
to experience tremendous success in recent seasons is Liverpool FC manager Jürgen
Klopp. He is generally accepted as the originator of the term ’Gegenpressing’, which
is well-known in both its German version and English translation.2 It is apparent to
football experts that Klopp’s counterpressing concept is closely related to Guardiola’s
strategy of regaining the ball within the first five seconds.
There are significant differences in team’s defensive and offensive tactical line-ups
(Bialkowski et al. 2014, 2015; Andrienko 2019; Shaw and Mark 2019). The transi-
tion phase describes the period following a win or loss of possession in which the
team transitions between its offensive and defensive tactical line-ups and vice versa.
When a team is in possession for at least a certain amount of time, we can assume that
generally its tactical formation is optimized for offensive play and, consequently, sub-
optimal in terms of defending its own goal (Andrienko 2019; Shaw and Mark 2019).
Therefore, the first seconds after losing the ball are critical from a defensive perspec-
tive. Several studies proved that transition phases are a substantial factor for a team’s
overall performance: As early as in 1968, Reep and Benjamin (1968) demonstrated in
the first known football analytics study, that 30% of all regained possessions lead to
shots on goal and 25% of all goals came from regained possessions in the attacking
quarter. Grant et al. (1999) confirmed these findings for the 1998 World Cup. Both
outcomes align perfectly with Jürgen Klopp’s statement that regaining the ball imme-
diately after loosing it, potentially through successful counterpressing, "...is the best
1 “[..] after losing the ball, the team has five seconds to retrieve the ball, or, if unsuccessful, tactically
foul their opponent and fallback”, Pep Guardiola; https://www.theblizzard.co.uk/article/peps-four-golden-
rules, accessed 06/20/2020.
2 https://www.sueddeutsche.de/sport/premier-league-bei-klopps-liverpoolern-klemmt-das-gaspedal-1.
2695408-2, accessed 06/20/2020.

123
Data-driven detection of counterpressing in professional football

playmaker".3 Klopp hereby claims that counterpressing can also be seen as an offen-
sive strategy. Recent studies show that regaining the ball in open play likelier leads to
a goal than a save build-up from a team’s own half (Vogelbein et al. 2014; Hobbs et al.
2018). Based on tracking data from the English Premier League, Hobbs et al. (2018)
detected possession regains close to the opponent’s goal—potential counterpressing
situations—highlighting their relevance once more. Even though many coaches and
clubs affected the development of this sophisticated strategy, neither an objective proof
of its efficiency, nor an analysis on its usage in top leagues is presented in the literature.
Hughes and Ian (2015) point out that team sports performance analysis tends to be
operationalized on the basis of notation systems, described as a replicable and consis-
tent method of recording sport performance. Recent literature explained a framework,
where coaches‘ decisions are supported by several performance analysis reports from
games, teams and players (Travassos et al. 2013) and pointed out that team tactics
in football refer to both a priori decisions made before the match, and also real-time
adaptions during the game (Rein and Daniel 2016). In accordance to that it is described
as a complex process resulting from a network of inter-dependent parameters (Kempe
et al. 2014). These processes are conducted in a time-critical set-up, especially when
it comes to the world’s top leagues and competitions where teams need to encounter
different opponents several times a week. Although many clubs extended their match-
analysis departments considerably within the past years, the limited amount of time
and resources during matches forced teams to seek ways to automate processes and
gain insights faster in order to obtain a competitive edge.
These recent developments—the availability of accurate performance data and the
need for a quick detailed tactical analysis—signifies a huge potential for the application
of sophisticated machine learning techniques to football data and requires an efficient
collaboration of computer-science and domain experts (Herold et al. 2019; Goes et al.
2020; Rein and Daniel 2016). Many recent scientific investigations aimed to estab-
lish new key performance indicator (KPI)—metrics quantifying certain aspects of the
game: pass evaluation metrics were examined (Steiner et al. 2019; Goes et al. 2019),
metrics to quantify controlled space were defined (Kim 2004; Fernandez and Bornn
2018; Brefeld et al. 2019) and several studies evaluated shot metrics (Lucey et al.
2014; Rathke 2017; Fairchild et al. 2018; Anzer and Bauer 2021)4 and goal scoring
opportunities through possession values (Link et al. 2016; Spearman 2018; Fernan-
dez and Bornn 2018; Decroos et al. 2020). Additionally, there are many approaches
for measuring the defensive behavior of teams (Santos et al. 2018; Andrienko 2019;
Goes et al. 2019), and even approaches aiming to quantify pressing (Bojinov and Luke
2016; Andrienko 2017; Robberechts 2019). Although pressing and counterpressing
are closely related, they are two different phenomena. An interesting conference pro-
ceeding describes how specific counterpressing situations can be derived from detected
general pressing scenes.5 Several approaches also showed, that analyzing these KPI’s
or even aggregating simple statistics (e.g. the pass completion rate) over one or several
3 https://tactical-times.com/the-history-and-evolution-of-jurgen-klopp/, accessed 06/20/2020.
4 Often referred to as expected goals (xG) values.
5 Will Gürpinar-Morgen (2018). "How StatsBomb data helps measure counterpressing", Statsbomb Inno-
vation in Football Conference 2018; https://statsbomb.com/2018/05/how-statsbomb-data-helps-measure-
counter-pressing/, accessed 11/11/2020.

123
 P. Bauer, G. Anzer

seasons provides a helpful indication to practitioners (Power et al. 2018; Pappalardo

et al. 2019). The primary goal of all these approaches is to derive new insights by pro-
cessing vast amounts of information. Decroos et al. (2018) presented a first approach to
detect interesting match-phases based on event data. To the best of our knowledge no
peer-reviewed study focused on automating parts of the performance analysts every-
day life by detecting complex tactical patterns based on positional and event data.
However, a noteworthy approach aiming to detect counterattacks was presented in an
established football analytics conference.6
With this practical need for process optimization in mind, it is the pivotal issue
of this study to detect counterpressing situations without human-support and provide
several ad-hoc reports for match analysts in near real-time. The outcome is optimized
to fulfill their practical requirements and fit seamlessly into their tool-ecosystem.
Additionally, the automated detection allows us to analyze large amounts of data that
would exceed manual processing capacities. Consequently, our approach enables us
to perform impartial long-term analysis of the German Bundesliga’s latest seasons
investigating the following research questions:
– Can we differentiate between varying regaining strategies and determine reasons
for a short defensive reaction time (definition in Sect. 2.1.2), i.e. to which extent
is a fast ball regain actually caused by counterpressing (RQ1)?
– Can we set objective benchmarks to quantify counterpressing strategies (amount
and effectiveness) on a match- and season-level and point out their correlation with
a team’s overall success (RQ2)?
– Do the established rules of thumb agree with the data (i.e. counterpressing is more
effective close to the sideline) (RQ3)?
– To what extent do team’s counterpressing strategies differ in the German Bun-
desliga (RQ4)?
All together, answering these research questions helps us to define the baseline for a
qualitative discussion with experts, and thus allows them to formulate requirements
for the practical application (PA) set-up.
The remainder of this paper is structured as follows: Sect. 2 provides a detailed
description of the used data, the underlying rules and definitions, the labelling process
and the extracted features. The outcomes in Sect. 3 are split into three parts: First
in Sect. 3.1, we describe a statistical evaluation of the detection models. Section 3.2
presents a subject-specific evaluation by interpreting our results on six seasons of
German Bundesliga. Lastly, in Sect. 3.3, we demonstrate how this approach can be
operationalized in the performance analysis process. This application is based on two
matches of the German national teams.7,8 All parts of this study were developed in
close cooperation with the professional match-analysts and coaches (see Acknowl-
edgements).

6 Karun Singh (2020). "Learning to watch football: self-supervised representations for tracking
data" In OptaPro Analytics Forum, London; https://www.youtube.com/watch?v=H1iho17lnoI, accessed
11/11/2020.
7 Germany against Northern Ireland; 19th of November 2019, Commerzbankarena Frankfurt.
8 Germany U21 against Belgium U21; 17th of November 2019, Schwarzbaldstadion Freiburg.

123
Data-driven detection of counterpressing in professional football

2 Methods

2.1 Data and definitions

2.1.1 Data collection

The present study uses positional and event data collected in more than six seasons
(4118 matches) of the German Bundesliga and 2nd Bundesliga, as well as the above
mentioned two matches of the German national teams. Positional data is captured by
optical tracking systems9 and event data consists of manual annotations based on a
dedicated official match data catalogue,10 defining around 30 events with more than
100 attributes. The event data can be seen as a log of the ball relevant actions (e.g.
passes, shots, tacklings or fouls), however, it does not cover complex team-tactical
behaviors such as counterpressing.
Since the two data sources are collected independently of each other, they need
to be synchronized before they can be processed together. Even though several steps
of quality management from independent institutions are conducted on the manually
collected event data, the assigned timestamp of a given event can differ significantly
to one in the positional data. The synchronization of positional and event data is
conducted by dedicated rules (per event) that extracts the exact timestamp and the
exact location on the pitch from tracking data given a manually tagged event. For
example, when synchronizing a pass, the sudden increase in the distance between the
passing player and the ball, captured by the optical tracking, can be used to align both
location and timestamp of the pass. The positional data is collected at a frequency
of 25 Hz and includes the longitudinal, latitudinal, and in case of the ball, also the
altitudinal positions of the players, ball and referees related to the pitch markings.
The information about which team is currently in possession of the ball (hereafter
referred to as ball possession) and whether the game is running or currently stopped
(hereafter referred to as ball status) are crucial for our survey. Both values are col-
lected live in the stadium for every frame of the match by a skilled operator focused
exclusively on this task.

2.1.2 Definitions

Since there are conflicting definitions of ball possession in the literature (Kempe
et al. 2014), we decided to adopt published definitions with expert feedback. The
above mentioned operators, dedicated to acquire information about ball possession
and status, are briefed to mark ball possession for one team, if and from that time
point a player of that team touches the ball with ball control, until the ball is out of
play, or an opponent player touches the ball with ball control. Ball control is defined
in this context, as the ability to conduct a contrived action with the ball. Whenever a
pass is played between two players of one team, the ball possession belongs to that
team as long as no opposing player intercepted that pass or won the ball within an

9 https://chyronhego.com/products/sports-tracking/tracab-optical-tracking/, accessed 06/20/2020.

10 https://www.sportec-solutions.de/en/index.html, accessed 06/20/2020.

123
 P. Bauer, G. Anzer

individual duel. According to the definition from Link and Hoernig (2017) we also
compute ball possession on a player level (individual ball possession). In the case of
an interception, the ball possession change is detected exactly at the time of the first
ball touch of the intercepting player. We use the term defensive reaction time—the
time it takes to regain ball possession after losing it—as defined in Vogelbein et al.
(2014). All situations where either the ball is beyond the pitch markings or the play
is stopped by the referee (e.g. because of a foul) are labeled as out-of-play. Hence,
if the ball goes out of bounds there must typically be a change in ball possession.
Situations in which the touch of the player carrying the ball outside the markings is
not declared as a ball possession due to missing control (e.g. a deflected shot), or
when the individual possession model disagrees with the team possession flag are
excluded. For all further investigations only the effective playing time (also referred
to as net playing time)—defined as all the situations while the game is running—are
considered. Shots, for example, always represent the end of a ball possession phase
per definition. Ball possession phases that end with the halftime-, or final-whistle or a
referee ball are excluded from our analysis.
In addition to these general rules, we developed the following transition-related
definitions in consultation with match-analysis experts: A defensive transition phase
is defined as the time-window when a team loses ball possession, but is not yet into
their ideal defensive formation. Within these defensive transition phases,
a team conducts counterpressing if at least one player exerts (spatio and/or temporal)
pressure on the ball carrier, or on the opponents close to the ball.
Note that there exist many different definitions for pressing: StatsBomb11 defines
pressing as a defensive player being within a five-yard radius of the ball-carrying
opponent.12 Very similarly, a more granular and non-binary definition, aggregating the
pressure of several defensive players, is presented by Andrienko (2017). Based on these
pressing definitions, counterpressing could be defined as situations where pressing is
exerted immediately after a ball possession change (Navarro and Javier 2018). Both
of these rule-based definitions are used as a baseline model for our investigation.
However, according to the match-analysts involved in this project, being close to the
player in ball possession is not the only way to exert pressure. Attacking or blocking
the easiest pass options could, for instance, also be seen as applying pressure.
To quantify the success of counterpressing, we consider it as successful if the ball
is regained within five seconds and shots and goals, scored or received, are accredited
to the previous counterpressing phase if they occur within the following 20 seconds.
From hereon the game is split into ball possession phases which could start and end
either with an in-play ball possession change or a stoppage such as a set-piece. Note
that the set-up of the in-play ball possession change, such as the defensive transition,
might not be the only influence factor on the defensive reaction time—it can also
occur due to short, uncontrolled ball possessions or risky passes of the opponent. Any
ball possession phases that either start with a set-piece or end with a stoppage in play
will not be considered further. Fig. 1 shows a heatmap displaying the occurrences of
11 Statsbomb is a football event data provider based in the UK, https://statsbomb.com/, accessed
12/17/2020.
12 StatsBomb event data, including the pressing tag, are can be accessed for many professional leagues
https://statsbomb.com/data/.

123
Data-driven detection of counterpressing in professional football

Fig. 1 Overview of where on the pitch turnovers happen most frequently (from the perspective of a team
playing from left to right)

transition situations related to the pitch. It indicates, that most turnovers happen in
the opposing half, especially near the sidelines. Ball possession changes due to a ball
going out of bounds are added to the area touching the sideline. Easily identified is
the high proportion of turnovers in the opponent’s six-yard box. This is likely because
both saved shots and shots missing the goal wide are counted as a change in possession
as soon as the goalkeeper receives the ball.

2.2 Supervised machine learning set-up

2.2.1 Hand-crafted labeling of defensive transition situations

Since the rule-based approaches to detect counterpressing we investigated lead us to

an insufficient accuracy (see Sect. 3.1), we conducted a manual tagging procedure
with trained student-analysts. It was their task to label situations with a detectable
counterpressing strategy. In total, out of 11, 108 relevant defensive turnovers, 3, 196
situations were labeled as counterpressing. The labeling was conducted for the first
eleven Bundesliga-matchdays of the 2018/2019 season from the perspective of the
home team. The percentage of counterpressings detected per transitions differs signif-
icantly per team. Borussia Mönchengladbach presented the highest value (40.07%),
whereas only 21.80% of Hannover 96’s transitions have been labeled as counter-
pressing. The aggregated outcome of the labeling process per team of the German
Bundesliga is displayed in Table 7 in the Appendix A.
To quantify the inter-labeler reliability, 20 matches were labeled by three different
students. To compute the pairwise accuracy for each defensive turnover, we checked if
both students had identified countpressing in the following two seconds. This yielded
a pairwise accuracy of 82.01%, i.e. in 82.01% of defensive turnovers both students
agreed on the nature of the actions following a turnover.
As additional information, the experts tried to detect the exact start and end-frame
of the respective transition situation. The average duration of all transitions phases is
9.34 s, 9.89 s for counterpressing, whereas all non-counterpressing turnovers took in
average 9.11 s.

123
 P. Bauer, G. Anzer

2.2.2 Expert-based feature extraction

We defined a list of 134 features that aim to characterize the transition. The features
describe the location of a ball possession change, several relevant factors describing
both teams’ exact positioning at the time of turnover and their movements in the first
two seconds immediately after the ball loss. A time-window longer than two seconds
was problematic, because it would cut off too many situations where the ball possession
changed within that time.
A teams’ decision to conduct counterpressing is heavily influenced by the situation
of the ball possession itself. To take this into consideration, all features are also cal-
culated at the moment of the ball possession change. According to football experts,
turnovers without the chance to counterpress are often characterized by immediate
clearances or aerial duels. Therefore, we included the ball position, the ball height,
and the individual ball possession time (Link and Hoernig 2017)—describing the time
a player of the ball possessing team was in direct control over the ball. The involved
football experts suggested, that counterpressing is often characterized by achieving
a local compactness close to the ball. We aimed to cover this with several metrics
measuring the regaining team’s positioning around the ball. For instance, we use the
team’s covered area, global and local stretch indices (Bourbousson and Carole Sève
2010; Santos et al. 2018) as features in our model. A team primarily aiming to defend
their own goal after losing possession does this usually with high-speed towards their
own goal, whereas counterpressing requires often only players close to the ball to
attack their opponents with a high speed towards the ball carrier. This is addressed by
calculating several speed-values and considering each team’s average position, the so-
called team-center (Bourbousson and Carole Sève 2010; Andrienko 2017). In contrast
to a more conservative transition strategy, counterpressing’s primary objective is not
to place many players in a compact unit behind the ball quickly, but rather to defend
more aggressively up the pitch. Therefore, we calculate both the number of players in
front and behind the ball, as well as their respective compactness. Although the press-
ing definition from Andrienko (2017) was not sufficient as a stand-alone rule-based
counterpressing detection criteria (see Sect. 3.1), it is incorporated in various features
of our model.
All features were discussed, consolidated and steadily improved within workshops
and based on several steps of evaluation of the detection. A detailed list and description
of the features is presented in Table 1, a video describing some of the features can be
accessed here.

2.3 Model training

2.3.1 Detection of counterpressing as a supervised machine learning task

We trained several classification algorithms based on the 11, 108 labeled defensive
turnover situations from 97 matches fulfilling our inclusion criteria (see Fig. 1).

123
Data-driven detection of counterpressing in professional football

Table 1 The extracted features that are used for counterpressing detection. Features used in both dimensions
of pitch coordinates (horizontally and vertically) and for different time points after the initial ball possession
change are listed only once

Feature Definition

Turnover position Location (x,y-coordinate) of the ball at the timepoint of the ball
possession change (BPC), hereafter 0 s.
Ball height z-coordinate of the ball tracked at several timepoints after
(BPC) (0 s, 1 s, 2 s)
Distance team Distance closest player to the ball calculated at different
timepoints after BPC (0 s, 1 s, 2 s) and for each team
(opposing, regaining).
Number players close to ball Number of players in several circles (10 m, 20 m, 30 m) around
the ball counted at different timepoints after BPC (0 s, 1 s,
2 s); calculated for both teams separately.
Team center Average position of all players of each team (goalkeeper
excluded) calculated for both dimensions (x,y) and at several
time points after BPC (0 s, 1 s, 2 s); calculated for both teams
separately.
Covered area team Team’s covered area measured as the widest distance between
two players in two dimensions (x,y) and at several time points
after BPC (0 s, 1 s, 2 s); calculated for both teams separately.
Players closer to ball Number of players from a team closer to the ball than next
player from the other team calculated at different time points
after BPC (0 s, 1 s, 2 s); calculated for each team separately.
Speed closest player Speed of player closest to the ball calculated at different time
points after BPC (0 s, 1 s, 2 s) and for each team separately.
Duration previous ball possession Duration of the previous ball possession phase; sequences
where the ball was out-of-play (ball status) have been
excluded.
Speed team Average speed of each team at different time points after BPC
(0 s, 1 s, 2 s); calculated for both teams.
Individual ball possession (absolute) Total individual ball possession time of any player from the
regaining team within up to the first 3 seconds after BPC as
defined in Link and Hoernig (2017).
Individual ball possession (relative) Total individual ball possession divided by the length of the
considered time window, so either divided by 3 seconds or by
the duration of the ball possession change if it is shorter.
Players in front of the ball Number of players for each team that are further away from the
own goal center than the ball at different time points after
BPC (0 s, 1 s, 2 s); calculated for both teams separately.
Players behind the ball Number of players for each team that are closer to the own goal
center than the ball at different time points after BPC (0 s, 1 s,
2 s); calculated for both teams separately.
Global compactness team Normalized stretch index of each team team (excluding
goalkeeper) based on definition in Bourbousson and Carole
Sève (2010) calculated at several time points after BPC (0 s,
1 s, 2 s); calculated for both teams separately.

123
 P. Bauer, G. Anzer

Table 1 continued

Feature Definition

Local compactness Normalized local stretch index of different player groups (3,4
and 5 player closest to the ball) at different time points after
BPC (0 s, 1 s, 2 s) as shown in Santos et al. (2018); calculated
for both teams separately.
Compactness in front of the ball Normalized local stretch index of all players of each team that
are further away from the own goal center than the ball at
different time points after BPC (0 s, 1 s, 2 s); calculated for
both teams separately.
Compactness behind the ball Normalized local stretch index of all players of those players of
each team that are closer to the own goal center than the ball
at different time points after BPC (0 s, 1 s, 2 s); calculated for
both teams separately.
Pressure regaining team on ball Normalized pressure (definition Andrienko (2017)) the
regaining team conducts on the ball at different time points
after BPC (0 s, 1 s, 2 s).
Pressure on ball possessing player Normalized pressure (definition Andrienko (2017)) the
regaining team conducts on the ball possessing player at
different time points after BPC (0 s, 1 s, 2 s).
Effective playing space Defines the space covered by the convex hull of all players of a
team (excluding goalkeeper) calculated at different time
points after the BPC (0 s, 1 s, 2 s) as shown in Santos et al.
(2018); calculated for each team separately.

We split the labeled data-set (75% training data, 25% test data) by taking randomly
25% of all transitions out of every match to avoid over-representing teams, scores, or
results.
We used the above defined features (section 2.2.2 or Table 1) and evaluated the best
performing models on our set of test data.
Among different basic-models, we applied extreme gradient boosting (hereafter
referred to as XGBoost), a scalable tree boosting system, introduced by Chen and Car-
los (2016), which outperformed traditional machine learning algorithms in numerous
applications (Li et al. 2019; Liu et al. 2020; Zhang et al. 2020). For our inves-
tigation, we want to point out three major advantages of XGBoost: (a) To make
use of our wide set of features without taking the risk of overfitting, an addi-
tional regularization term is added to the loss function. Additionally, (b) XGboost
is a scalable machine learning model, which can be extended seamlessly with
more data or more features being available. Furthermore, (c) no normalization is
required.
Before training our model, a set of hyperparameters (shown in Table 2) has to
be defined. As presented in Wang (2019), we applied Bayesian tree-structured Parzen
Estimator hyperparameter optimization approaches to obtain the highest possible accu-
racy and avoid overfitting. By using tree-structured Parzen Estimators (Bergstra 2011)

123
 Table 2 Hyperparameter-selection of the XGBoost models

Hyperparameter Description Range XGB-M1 XGB-M2

1 Learning rate Controls the step size used per update [0, 1] 0.045 0.05
2 Max depth Limits the depth of the tree [0, ∞) 5 7
3 Subsample Controls number samples applied to the tree (0, 1] 0.419 0.279
4 Min child weight Controls instance weight of a node [0, ∞) 0.855 1
Data-driven detection of counterpressing in professional football

5 nrounds Limits number of iterations 1–400 400 400

6 Class balancer Controls the balance of negative and positive weights (0, ∞) 1 2.5
(Number of negative cases / Number of positive cases)

123
 P. Bauer, G. Anzer

as the surrogate model in Bayesian optimization (Dewnacker et al. 2016), we reduce

the running time of hyperparameter tuning and achieve better scores on the testing set.
Note that the hyperparameter nrounds was set to a maximum of 400 iterations.
To further guarantee the stability of our model and avoid overfitting, we applied
five-fold cross-validation on the training data.
As described in Sect. 2.1.2, we also implemented two rule-based baseline models
that serve as a benchmark for our detection. A naive approach defines counterpressing
as follows: whenever one or more players are within a five-yard radius around the
ball carrier during the first individual ball possession phase following a turnover, it
is classified as a counterpress (hereafter referred to as naive rule-based approach).13
The second approach (hereafter referred to as Andrienko-approach) defines counter-
pressing as all turnovers whenever the first player in ball possession receives pressure
exceeding a certain threshold according the pressure-definition in Andrienko (2017),
whereby the final threshold of 0.74 was obtained by maximizing the F1 -score on the
training set.

2.3.2 Effectiveness for counterpressing and fast possession regains

In order to define some success metric of a transition phase, the low scoring nature of
football causes us to examine the following actions more granularly, rather than just
checking whether they are followed by a goal. For both cases—successful and unsuc-
cessful ball recoveries through counterpressing—we extracted taken shots, expected
goals14 and actual goals following a transition phase. To investigate this issue, two
definitions had to be made: Which ball recovery latency of a counterpressing strategy
should be considered as successful, and for how long a defensive and an offensive
action would be accredited to the previous ball recovery (strategy). As a starting point
for a potential threshold for successful counterpressing, a first indicator is given by
Pep Guardiola’s five second rule. We queried relevant video scenes with possession
regains after 3, 4, 5, 6, 7 and 8 seconds and discussed them with a group of profes-
sional match-analysts. The same procedure was conducted to investigate the follow-up
goal-scoring opportunities. Here scenes with shots 10, 15, 17, 20 and 25 seconds after
the initial ball loss were discussed. Through this procedure, we finally agreed on the
definitions described in Sect. 2.1.2.

13 This approach has been suggested at the Statsbomb Innovation in Football Conference
by Will Gürpinar-Morgen (https://statsbomb.com/2018/05/how-statsbomb-data-helps-measure-counter-
pressing/), accessed 11/11/2020.
14 The “Expected Goal” (xG) value of a shot denotes the a priori probability of a shot being converted to
a goal. Hence its value ranges from [0, 1]. The probability is estimated using both tracking and event data
and applying a machine learning model, that was trained on more than 100, 000 shots. Details regarding
the xG-model used can be found in Anzer and Bauer (2021).

123
Data-driven detection of counterpressing in professional football

Table 3 Statistical evaluation of the counterpressing outcome

Model Precision Recall F1 -score AUC

1 XGBoost (Model 1) 0.72 0.63 0.67 0.874

2 Logistic regression 0.69 0.51 0.59 0.841
3 Random forest 0.74 0.55 0.63 0.867
4 XGBoost with class balancer (Model 2) 0.60 0.80 0.69 0.865
5 Naive rule-based approach 0.31 0.87 0.46 0.602
6 Andrienko-approach 0.37 0.37 0.37 0.568

3 Results

3.1 Statistical evaluation

3.1.1 Detection of counterpressing

With the above described supervised machine learning set-up we are able to detect
counterpressing situations with sufficient accuracy for practical applications (see also
Sect. 3.3). Table 3 shows a statistical evaluation of the different models, from which
XGBoost performed the best. Per team and per match, we detect around 20 to 30
counterpressing situations, out of around 90 to 200 transition situations.
With the highest overall area under the curve (AU C) the above presented optimiza-
tion (Table 3, row 1) is best suited for the long term analysis of several seasons with
the goal to identify trends and underpin practitioner rules (RQ1-RQ4, Sect. 3.2). The
XGBoost model with a class balancer (Table 3, row 4) has a higher recall of 80% with
still an acceptable false positive rate. Thus, it can be applied for specific performance
analysis of either the own match or several matches of the next opponent (PA, Sect.
3.3)—where match-analysts spend a lot of time analyzing video footage either way.
The optimal hyperparameters used for both models can be found in Table 2. When
examining the results of the two baseline approaches, they exhibit a very low overall
accuracy. The naive rule-based approach (Table 3, row 5) classifies 72.41% of all
turnovers as counterpressing, which leads to a high recall but also a large number of
false positives. For the Andrienko-approach, selecting the threshold by optimizing the
F1 -score lead to a more realistic percentage of predicted counterpressing situations
in the test set (25.65%), but are, nevertheless, significantly outperformed by either
machine learning approach.
Another advantage of the XGboost approach is that the individual influences of our
rich feature set can be somewhat quantified and interpreted by analyzing the respective
SHAP-values.15 The naming was both coined by their originator Lord Shapley, who
introduced them in the context of cooperative game theory (Roth and Thomson 1988),
but also by Lundberg and Su (2017), who used the concept to interpret the features
for machine learning models. In comparison to traditional feature importance models
(e.g. gain or Saabas method), SHAP-values present a consistent and locally accurate
15 The abbreviation SHAP stands for SHapley Additive exPlanation.

123
 P. Bauer, G. Anzer

Fig. 2 Feature influence to the counterpressing prediction based on SHAP-values

method to identify the individualized feature contribution to machine learning models.

This method has been effectively used in different applications (Antipov et al. 2020;
Meng et al. 2020; Ibrahim et al. 2020; Anzer and Bauer 2021).
Figure 2 displays the most influential features according to the SHAP values in
two different representations. In the left Fig., each dot represents the contribution of
the feature to the model, whereas the color-coding describes the value of that feature.
Both, the absolute individual ball possession time (IndividualBallPossession (abs))
and the speed of the regaining team two seconds after the change in ball possession
(SpeedRegainingTeamPlayer (2s)), have a very strong and linear impact on the pre-
dictions. Besides the fact, that both features have the highest overall influence on the
prediction (widest dispersion of the dots on the left part of Fig. 2, the interpretation of
the SHAP-values can be expressed as follows: the higher the absolute amount of indi-
vidual ball possession time within the first three seconds and the higher the speed of
the regaining team two seconds after the turnover, the more likely a defensive turnover
is classified as counterpressing. The number of opposing players behind the ball after
one second (PlayersOpposingTeamBehindBall (1s)) influences the prediction in a dif-
ferent way: A high number of players behind the ball increases the chances for a
classification of counterpressing, but the relation is clearly non-linear. To get a better
idea of the influence, we will have a look at the right part of Fig. 2. The value per feature
is now displayed by the x-axis, whereas the model influence is shown on the y-axis. If
less than four players are behind the ball, this feature on average decreases the chance
of a counterpressing classification. The number of four defenders—almost half of the
team—seems to be a decisive threshold. If four or more players are behind the ball,
this feature has a positive contribution to the prediction. This not only aligns with the
expectation of the practitioners, but also led to a very valuable discussion among the
professional analysts. A more complex relation is shown by the local stretch index
of the five closest player to the ball of the opposing team two seconds after the ball
possession change (OpposingTeamlocal5 stretch (2s)). After a steep increase starting
at 600 cm, the influence of that feature reaches its maximum at roughly 1, 000 cm (see

123
Data-driven detection of counterpressing in professional football

Table 4 For all 4118 considered matches of the German Bundesliga and 2nd Bundesliga, this table shows
the probability of shots and goals per team following counterpressing situations

Strategy Goals scored % Goals conceded % Shots for % Shots against %

Counterpressing (all) 0.35 0.78 3.22 5.15

Unsuccessfull counterpressing 0.16 1.02 1.83 6.7
Successfull counterpressing 0.75 0.25 6.27 1.76

right plot in Fig. 2) but decreases afterwards. This indicates that a higher stretch index
(lower compactness) of the opposing team after two seconds increases the chances for
counterpressing.
Excluding features with little to no influence according to the SHAP-values, did
not improve neither the F1 -score nor the AUC of our prediction on the test data-set.

3.1.2 Effects of counterpressing

Table 4 shows the outcome of counterpressing regarding goals and shots scored or
conceded within 20 seconds. If one is successful, i.e. wins back the ball within five
seconds the chance of scoring increases tremendously, but if unsuccessful one is far
likelier to concede a goal.
It is no surprise, that the chances to either shoot or score are significantly higher
when counterpressing was applied successfully, since it implies the crucial attacking
advantage of having gained the possession of the ball.
While using goals scored versus conceded would theoretically be ideal to measure
success, the low scoring nature often prevents us from doing so. Therefore, we use shots
to compare teams and coaches. Nevertheless, according to the experts, looking at both
shot- and goal- (or even expected goal-)balance is a very valuable key-performance-
indicator for counterpressing.

3.2 Subject-specific evaluation of six seasons of German Bundesliga data

In the following section, we use our quantitative results as a baseline for a qualitative,
subject-specific evaluation and interpretation with the involved experts.

3.2.1 Lessons learned about defensive transitions (RQ1)

A common procedure when analyzing a team’s transition strategy is looking at the eas-
ily acquirable defensive reaction time (Vogelbein et al. 2014). This, however, comes
with the drawback that it is not able to distinguish between situations with inten-
tional counterpressing behavior and noise. Note that in non-trivial defensive turnover
situations typically a team can choose between falling back or conduct counterpress-
ing. But for the purpose of this study, the distinction between fallback and other
non-counterpressing situations was removed for the sake of simplicity. An additional
analysis of the expert-based labeling (see Appendix A) showed, that around 62.5%

123
 P. Bauer, G. Anzer

of all ball losses fulfilling the inclusion criteria cannot be assigned to either defensive
strategy (counterpressing or fallback). The sheer number of these situations—with
a very short defensive reaction time (on average 7.83 s) and without any defensive
tactical choices being detectable—significantly influences the defensive reaction time
when applied to all transitions. Further analysis (see Appendix A) shows that specif-
ically turnovers with very short individual ball possession times fall in this category.
Their exclusion presents a crucial step for a better understanding of transition situa-
tions.
In general, counterpessing is not always advantageous and needs to be executed
well. Although, the expectation of some practitioners (invigorated by Jürgen Klopp’s
statement "counterpressing is the best playmaker") may be different, it is still intuitive
that the team in possession of the ball has a higher chance to perform a successful
offensive action (e.g. a shot or goal; see Table 4, row 1). However, this is an average
over all teams independent of their skill level and as we will later see, there are some
teams/coaches that were able to apply counterpressing so successfully that they even
ended up with a positive shot balance by creating more shots after counterpressing than
conceding. In order to properly assess the risk versus reward nature of counterpressing,
one would ideally compare it to its strategic counterpart falling back. But even then,
one would need to carefully address potentially confounding variables describing
the original situations, since, as it seems from the feature importance discussion, the
situation typically dictates the strategic response. This goes, however, beyond the
scope of this study, but could be the ground for interesting future work. Additionally,
since all non-counterpressing situations consist of myriad of different circumstances,
they do not serve as reasonable baseline to effectiveness of counterpressing.

3.2.2 Define and statistically underpin objective Benchmarks (RQ2)

Based on the above explained definitions and trained prediction models, several quo-
tients and ratios were discussed with the experts. Aggregated on a season level, we
analyzed the correlation with a team’s final ranking. Our detection provides several
different metrics with a significant correlation to success. These performance indi-
cators can be calculated per match, per match phase (e.g. one halftime) or even per
turnover, which allows practitioners to objectively compare their teams performance
with pre-defined benchmarks. With a negative Pearson correlation, the ratio of suc-
cessful counterpressing-situations to the total number of transitions predicts a team’s
final ranking the best (r = −0.44). Another metric, correlating with the final ranking
that was very valuable to the experts due to its direct interpretability, was shot and
goal balance (r = −0.36 shots, r = −0.42 goals)—describing, whether more shots
or goals are taken after successful counterpressing, than are conceded after failed
attempts. However, these metrics should be used carefully since they are based on
small sample size and could contain confounding effects with the overall offensive or
defensive qualities of a team. Loosing the ball, increases the probability of an oppo-
nent conducting an offensive action. Thus we present an effective strategy to monitor
the outcome of counterpressing strategies, such as several performance indicators that
enables coaches to objectively benchmark a team’s defensive transition behavior.

123
Data-driven detection of counterpressing in professional football

Fig. 3 Heatmap of ball possession change locations before unsuccessful counterpressing situations (left)
and sucsessfull situations (right)

3.2.3 Approve established rules of thumb (RQ3)

A widely spread rule of thumb is that counterpressing is ideal after ball losses close
to the sideline or close to the corners in the opponent’s half. Fig. 3 presents two
heatmaps that underpin this statement. Secondly, we want to examine, whether a
numerical superiority of players close to the ball increases the chance of a successful
counterpress as assumed by many experts. For that we examine the 109, 852 detected
counterpressing situations satisfying the inclusion criteria. Whenever the team out
of possession has a numerical superiority within a 10 m radius around the ball, at
the time of the turnover, they regain ball possession within 5 seconds 36.2% of the
time, compared to only 30.2%, when the other team has more players in that area. This
indicates that the rule of thumb has some truth to it, but is far from the only influencing
factor, deciding whether a counterpress will be successful.

3.2.4 Compare Team’s and Coach’s Counterpressing in German Bundesliga (RQ4)

Further investigations highlight to which extent teams use completely different defen-
sive transition strategies. We investigated the coaches that were expected to have a
pronounced counterpressing-behavior by the experts. The respective ranking among
all coaches finishing a full season can be found in the Appendix C.
Table 5 also gives a first indication of which further aspects could be considered.
Teams ending up in the top 5 of a season perform above average in almost all defined
metrics. No consistent tendency over the considered seasons is detectable: games do
not get more intensive in terms of total in-play transitions per match, nor are there
significant changes in teams average counterpressing behavior. FC Bayern München
has the best goal balance after counterpressing, which might be heavily influenced
by their offensive efficiency. Jürgen Klopp and Ralf Ragnick performed extraordinary
well in terms of ending up with more created than received shots after attempting
counterpressing. Given that Jürgen Klopp ended only seventh place in one of his two
considered seasons, this should be seen as an outstanding performance. Over the course
of a whole season only nine coaches achieved a non-negative shot balance within 20 s
after their own counterpressing—the average final ranking of the respective teams was
four. Note that teams playing at home tend to conduct counterpressing slightly more
often than away teams. Considering only home-teams, our model classifies 27.24%

123
 P. Bauer, G. Anzer

Table 5 Comparison of counterpressing-related performance indicators in Bundesliga. The colored arrows

display the rankings within the respective groups compared to the average (teams by ranking; all teams by
seasons; teams and coaches). For the columns presenting the shot and goal balance we used colored dots
for the teams and coaches, showing whether the total outcome per match is positive, neutral or negative

of all included defensive transitions as counterpressing, which is roughly in line with

the labeled training data, that was conducted only on home teams (28.77%).
Figure 4 shows a shortlist of teams’ counterpressing outcomes. Note that, since on
average more shots/goals are conceded than created when counterpressing, all four
axis are scaled differently. For both sub-figures, teams on the upper left, including, for
example, 1. FC Nürnberg16 , Hannover 9617 and SV Darmstadt 9818 in the left figure,
perform worse. Teams on the bottom right tend to generate more shots/goals and
allow fewer while counterpressing. Teams with high values in the top right quadrant
like Borussia Mönchengladbach19 , FC Augsburg20 or TSG Hoffenheim21 seem to
employ risky defensive transition strategy, by both creating and also receiving many
shots after their own counterpressing.
As a general recap of the Bundesliga analysis, we would like to point out that
teams use significantly different transition strategies (RQ4). The experts’ expecta-
tions of which coaches use counterpressing more often and/or more efficiently were
underpinned by the results.

3.3 Proof of concept

The central objective of this study is to automate the detection of counterpressing

situations. This helps match-analysts in their daily processes by saving them time, but
also by providing objective and comparable benchmarks. First, we describe the general

16 Highest in the left figure.

17 Green circle with the inscription 96.
18 Blue circle around a white lily.
19 Black and white hatched diamond logo roughly in the center of both plots.
20 Fifth highest in the left figure.
21 Second from the right in the right figure.

123
Data-driven detection of counterpressing in professional football

Fig. 4 For both shots (left figure) and goals (right figure) teams are displayed depending on their percent-
age of counterpressings leading to either shots or goals for (x-axis) or against (y-axis). Team values are
aggregated across all considered seasons

set-up, whereas a prototypical application for two exemplary matches of German

national teams is conducted.22,23 Based on the results described above, we are now
able to provide match-analysts with two files fully automatically in virtually real-time:
First, they receive a list of all detected counterpressing situations. To integrate this
efficiently into their ecosystem, the files are produced in different file-formats, which
can be imported into their video-analysis tool of choice (e.g. Hudl Sportscode,24 Stats
Edge Viewer25 ). Such tools basically help to handle tags or labels in combination with
the video footage. Figure 5 shows how this eliminates the usual process of the match-
analysts labeling the videos manually in an exemplary tool (here Hudl’s Sportscode).
Usually, match-analysts use these tools to tag important situations live during the
match but also in detail post-match for opponent analysis. Once a match or parts of it
are tagged, the tool allows the analyst to output the tags either as a video-playlist or
as an xml-file, containing the category and the time-frame of each tag. Depending on
the coaches needs, the outcome is either presented as a video-playlist or a quantitative
report giving an aggregated overview—which also is typically produced manually.
An automatically generated counterpressing-playlist for the U21 match based on our
prediction can be viewed here.
Second, we automatically provide coaches and analysts a counterpressing match-
report with visualization after the respective match or entire season reports. For the U21
match an excerpt of the automated match report is presented in Fig. 6. Only two shots
occurred within 20 seconds after either team counterpressed. However, these two shots
after unsuccessful counterpressing attempts by the German team lead to two goals,
which were decisive for the total match-outcome (Germany vs. Belgium, final score

22 Germany against Northern Ireland; 19th of November 2019, Commerzbankarena Frankfurt.

23 Germany U21 against Belgium U21; 17th of November 2019, Schwarzbaldstadion Freiburg.
24 https://www.hudl.com/products/sportscode, accessed 06/20/2020.
25 https://www.statsperform.com/team-performance/football/stats-edge/, accessed 06/20/2020.

123
 P. Bauer, G. Anzer

Fig. 5 Integration of the counterpressing-analysis into the match-analysts and coaches daily business.
The yellow colored part shows the traditional process, the grey part displays how the automated analysis
assimilates seamlessly

Fig. 6 Automated match report for the match Germany U21 against Belgium U21 in Freiburg. The pitch
visualizations shows the positions of ball losses leading to a successful (green) or (unsuccessful) coun-
terpressing. Whereas the absolute values are presented in the middle block, a benchmark of the central
performance indicators against Bundesliga average is shown in the lower grey box

2–3). Whereas situations leading to goals are analyzed on a highly detailed level by
coaches and analysts either way, they arrived at the same conclusion. Nonetheless,
this report helps coaching staffs to evaluate whether the German team had a general
bias in their defensive transition behavior or whether the two goals happened as a
consequence of extraordinary opposing actions or defensive errors from the German
team. In this case, bad counterpressing behavior had a significant stake in the origin
of both goals. Another outcome was that many successful counterpressing situations
(1.23% above Bundesliga average) did not end up in a single shot.
The practical implementation of this study was prototyped by the example of two
recent matches of the German national teams. After each match, both the Sportscode-
xml file and the automated match-report sheets were produced and shared with the
experts for their post-match-analysis process. Additionally, to validate the report, the

123
Data-driven detection of counterpressing in professional football

Table 6 Statistical expert-evaluation of the counterpressing outcome. For the defensive transition column,
the number in brackets displays the number of scenes excluded by our criteria (see Fig. 1)

Match Team Defensive Counterpressing Manually Additionally

transitions detected excluded detected

GER-NIR Germany 153 (85) 25 6 0

GER-NIR Northern Ireland 155 (95) 24 3 2
GER-BEL Germany 164 (123) 33 5 0
GER-BEL Belgium 176 (87) 25 2 1

matches were manually analyzed and provided a ground truth to compare our results
with. The overall results are shown in Table 6.
For the first game 153 relevant defensive transition situations of the German national
team were queried by the above defined rules. Out of these, 25 scenes were detected
as counterpressing from which 6 were manually excluded from the final counterpress-
ing playlist. For 164 defensive transitions from the second match 33 situations were
detected as counterpressing and analysts ruled out 5 manually. The manually excluded
scenes were discussed with all experts and it turned out that different definitions and
interpretations lead to different labels. In this case study, ten of the eleven manually
excluded situations consisted of only one player exerting pressure. Although this ful-
fills our definition, some of the experts would classify situations with only one player
defending actively towards the ball as fallback (definition in Appendix A). All scenes
that were additionally labeled by the respective match-analysts (in total five) contained
a clearance closely after the initial ball loss and where thus not clearly related to the
counterpressing strategy.
Both the Sportscode-xml and the automated match report turned out to be valuable
for the coaching staff. Due to the interpretability of the inaccuracies, the experts trust
the outcome for further applications. They deemed the results to be sufficient in terms
of a practical usage and noted that it saved vast amounts of resources in the pre- and
post-match-analysis. The automated match-report (see Fig. 6) allows us to provide
an objective comparison, with a flexible benchmark (here Bundesliga average) and
therefore provide a new way to approach complex tactical strategies. The shot- and
goal-balance are very intuitive and present a direct monitoring of the efficiency of
conducted strategies.

4 Discussion

This paper shows that complex tactical strategies, such as counterpressing, can be
detected automatically based on synchronized positional and event data. Comparing
team’s counterpressing strategies objectively and on longer periods of time creates
insights that could not have been achieved with traditional methods.
The interdisciplinary cooperation turned out to be a very beneficial factor for this
study. In our opinion, such a set-up of competencies is necessary to obtain relevant
results. Machine learning techniques are required to detect complex strategies from
spatio-temporal data, but also tactical football expertise are inevitable to determine

123
 P. Bauer, G. Anzer

definitions, extract features and evaluate and interpret the resulting outcomes. A key
lesson we learned through this study is that both definitions of complex strategies
and their reading vary between football experts—this became apparent during an
intensive process of expert-supported evaluations. One of our most meaningful key-
performance-indicator is the shot-balance after counterpressing. Here, shots are used
as a proxy for a successful attack. This is a common procedure in football analytics,
however, the approach could be extended by using expected goals (e.g. Anzer and
Bauer (2021)) or expected possession values (e.g. Spearman (2018)).
Since there does not exist a comparable approach for detecting counterpressing
in the literature, we implemented two naive baseline approaches to benchmark our
model against (see Table 3). While the approach based on Andrienko (2017) was
originally not designed to quantify counterpressing but rather pure pressure, we build
this rule based approach on Fernandez and Bornn (2018), who defined counterpressing
as immediate pressure after losing the ball. Hence, it may not be the ideal approach,
but due to complete the lack of alternatives in the literature, we use it as a benchmark
model. Even though our model outperforms different rule-based baseline models (see
Table 3) and the prediction accuracy is sufficient for practical application (see section
3.3), the basic limitation to achieve further accuracy is the inter-labeler reliability of
82.01%. After discussing the definitions, no further steps of consolidation between the
labelers were conducted—but we would highly recommend such a step including the
strict monitoring of the inter-labeler reliability for similar investigations. However, data
labeling is a time-consuming process which cannot be conducted for each occurring
philosophy and definition. Furthermore, methodologies that reduce labeling efforts,
such as weak supervision, should be implemented on top of general detections as
the one presented here to adjust definitions to the specific needs and to improve both
the accuracy and the degree of individualization (Ratner et al. 2016, 2017). With an
even larger amount of labeled data, one could consider using continuous features, or
even the raw positional data of all players instead of features at discrete time points.
The application of labeling-support methods could lead to more individualized and
accordant labels and thus to a better prediction.
With an even more accurate model using one of the above described approaches,
an improved and team-individual success prediction model for counterpressing could
support the reflection of teams’ decision making processes significantly. Also, the
adaption of the counterpressing detection itself to team-specific definitions, provide a
huge potential for further investigations.
Vogelbein et al. (2014) evaluated 306 matches of the 2010/2011 Bundesliga season
and showed the time it takes to regain the ball also depends of the score at the time.
They pointed out that teams with a lead tend to regain the ball slower than the ones that
are trailing, and that teams finishing their season in the top third of the table regain ball
possession significantly faster than the other teams—especially in drawn and loosing
match states. We found that the defensive reaction time, which serves as a baseline for
our success definition, typically includes many noisy situations, where no clear strat-
egy is observable. Further lessons learned regarding defensive transitions are described
in section 3.2.1 and extended in Appendix A. The high influence of individual ball
possession times on the predictions (see Fig. 2) can be attributed to uncontrolled situ-
ations without the possibility for either defensive strategy (counterpressing, fallback).

123
Data-driven detection of counterpressing in professional football

This also shows a limitation of our work, that some of the model’s most important
features focus on the situation itself and fewer on the strategy conducted thereafter.
A possible explanation for this is that besides filtering out noisy situations, the model
found that most often the situation dictates the defensive response.
Nevertheless, a tendency of the opponent to play counterattacks, or especially risky
passes could lead to many fast ball recoveries independent from the defensive transition
strategy. This issue could be considered by either combining this approach with an
equivalent offensive transition strategy detection as shown for example in Hobbs et al.
(2018) or by including more features of the opposing team or even the raw data of
all 22 players and the ball. Not only the question in which situations counterpressing
induces a high chance for a possession regain, but furthermore—given a situation
where counterpressing is conducted—how likely is it to take/receive a shot when
conducting that strategy are of high interest for practitioners. Regaining the ball fast
might not be the only objective of counterpressing. Consequently, future investigations
should also consider quantifying alternative success definitions, e.g. slowing down the
opposing attack, forcing back-passes etc.
Another missing piece which should be investigated further is an accurate selection
of fallback situations. Comparing their risk-reward structure to counterpressing situ-
ations, could lead to crucial insights by evaluating a teams’ decision to counterpress
versus falling back objectively. Since different teams may have their own club-specific
definitions, our experimental set-up could be applied to arbitrary counterpressing-
definitions or even other tactical patterns, as long as they gather sufficient labeled
data. An interesting follow-up study, could investigate how many labeled matches
would be necessary to achieve a sufficient accuracy depending on the definition. In
our case we found 100 labeled matches to be sufficient, but also stress that a high
inter-labeler reliability is necessary.
Note that counterpressing is only one example of a complex tactical pattern, that is
of interest to match-analysts, but not covered in typical event level data. The needs of
match-analysis departments combined with the growing availability and accuracy of
positional and event data present a huge potential for task automating approaches.

5 Conclusions

Based on expert-evaluated definitions and hand-crafted labels, we are able to detect

counterpressing strategies automatically with a sufficient accuracy in a supervised
machine learning set-up. By producing both an understandable match-report and
tagging-files suitable for conventional video-analysis software, the integration of the
process into a match-analyst’s daily business saves a significant amount of time. The
outcome helps to analyze the own team’s performance and provides helpful informa-
tion about the next opponent’s defensive transition behavior (PA).
We can differentiate between intended counterpressing strategies and the many
uncontrolled transition situations with short defensive reaction times. This provides
not only a better understanding of transitions but also several more granular perfor-
mance indicators describing defensive transitions (RQ1). The respective performance
indicators, consolidated by statistical influences and expert opinions, derived inter-

123
 P. Bauer, G. Anzer

pretable and intuitive metrics (RQ2), such as the goal- or shot balance presenting
an effective efficiency quantification for the counterpressing strategy—that were not
used before but seem to have a huge potential according to the experts. Two of the
proven rules of thumb are that counterpressing is more likely to succeed closer to
the sidelines and a numerical superiority close to the ball increases the chance of
winning it back (RQ3). Through analyzing different facets over several seasons we
are also able to quantify trends over a large period of time: teams within the German
Bundesliga follow appreciably different transition strategies (RQ4). Furthermore, suc-
cessful teams—measured against their final ranking—tend to use the counterpressing
strategy more efficiently, giving credence to the notion of declaring it as an offensive
strategy (RQ2).
Supplementary Information The online version contains supplementary material available at https://doi.
org/10.1007/s10618-021-00763-7.

Acknowledgements This work would not have been possible without the perspective of professional match-
analysts from world class teams who helped us to define relevant features and spend much time evaluating
(intermediate) results. We would cordially like to thank Dr. Stephan Nopp and Christofer Clemens (head
match-analysts of the German mens National team), Jannis Scheibe (head match-analyst of the German U21
mens national team) as well as Sebastian Geißler (former match-analyst of Borussia Mönchengladbach).
Additionally, the authors would like to thank Dr. Hendrik Weber and Deutsche Fußball Liga (DFL) / Sportec
Solutions AG for providing the positional and event data.

Funding Open Access funding enabled and organized by Projekt DEAL.

Declarations
Ethical Approval By informing all participating players, all tracking is compliant to the general data pro-
tection regulation (GDPR) https://gdpr-info.eu/, accessed 07/20/20. An ethics approval for wider research
program using the respective data is authorized by the ethics committee of the Faculty of Economics and
Social Sciences at the University of Tübingen. The data are property of the DFL e.V. / DFB e.V. and cannot
be shared public. However, interested researchers can request samples of data under non-disclosure agree-
ment constraints at the respective institutions. With the description of the respective tracking vendors and
systems, peers working in the football industry can reproduce the results by using any kind of professional
football data.
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which
permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give
appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence,
and indicate if changes were made. The images or other third party material in this article are included
in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If
material is not included in the article’s Creative Commons licence and your intended use is not permitted
by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the
copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.

Appendix

Appendix A gives additional information regarding the labeling conducted for this
study including further tactical explanation of defensive transitions. Appendix B, C,
and D show the outcome of the defined counterpressing performance indicator per
team, per coach, and per season on the full Bundesliga data-set.

123
Data-driven detection of counterpressing in professional football

For all columns in Appendix B, C and D, T stands for turnovers, M for matches,
CP for counterpressing, S for Shots and G for Goals. Whereas + indicates successful
or positive offensive actions and − the other way around, +/− points out the offen-
sive/defensive balance as described above. The first table in Appendix B compares
all teams playing in Bundesliga between the 2013/2014 and 2019/2020 (until match-
day 26) seasons. The teams are ordered by their average final ranking, the number
of matches considered are shown in brackets. In a second table in Appendix C all
coaches playing in Bundesliga between the 2013/2014 and 2018/2019 seasons are
presented, whereas a third table shows all teams per season ordered first by their final
ranking and secondly by the column CP+/T. Only full seasons with 34 matches one
and the same coach where considered. The succession is made based on the highest
correlating feature with teams final ranking, number of successful counterpressings
per transition. Appendix D compares teams on a season-level and is sorted by the
respective final ranking which is shown in brackets.

Appendix A

As discussed in Sect. 3.2, the actual labeling was simplified for our analysis. During
the expert labeling process (see Sect. 2.2.1), not only counterpressing, but also fallback
was labeled by the experts. This strategy is said to be the alternative to counterpressing
and was defined as a defensive transition phase, where all players’ intention is to
either react inactively, or move backwards to their defensive line-up without exerting
pressure on the ball. The terms forward- and backward-defending are often used in
this context interchangeably to counterpressing and fallback. Figure 7 shows two
exemplary situations for both strategies and lists some typical observations for each
strategy.

Fig. 7 This figure shows an exemplary ball loss of the blue team, who conducts counterpressing in the
right figure and a fallback in the left picture. The pass leading to the turnover is displayed as a solid arrow
and the dotted arrows show the expected player movements in the seconds after the ball loss. The metrics
stretch-index and team-center are defined in Table 1

123
 P. Bauer, G. Anzer

Table 7 Outcome of the manual expert labeling. The first number per entry shows the number of scenes
that fulfill the criteria defined in Sect. 2.1.2 and are used for the model building. The total number of tagged
scenes is displayed in brackets

Team (Matches) Defensive Turnovers Counterpressing Fallback Undefined % CP

All (97) 11,108 (20,928) 3,196 (4,523) 970 (1,268) 6,942 (15,137) 28.77
1. FC Nürnberg (6) 666 (1,167) 178 (228) 94 (111) 394 (828) 26.72
1. FSV Mainz 05 (6) 689 (1,253) 214 (276) 76 (88) 399 (889) 31.06
Bayer 04 Leverkusen (5) 622 (1,092) 175 (262) 57 (63) 390 (767) 28.14
Borussia Dortmund (6) 656 (1,227) 210 (340) 38 (53) 408 (834) 32.01
Borussia Möchengladbach (5) 534 (963) 214 (289) 46 (50) 274 (624) 40.07
Eintracht Frankfurt (5) 656 (1,092) 206 (273) 61 (75) 389 (744) 31.40
FC Augsburg (5) 583 (1,075) 130 (188) 49 (59) 404 (828) 22.30
FC Bayern München (4) 443 (684) 133 (165) 46 (53) 264 (466) 30.02
FC Schalke 04 (5) 623 (1,054) 149 (190) 47 (56) 427 (808) 23.91
Fortuna Düsseldorf (6) 660 (1,233) 186 (233) 74 (96) 400 (904) 28.18
Hannover 96 (5) 610 (1,079) 133 (156) 44 (55) 433 (868) 21.80
Hertha BSC (5) 548 (1,010) 178 (258) 35 (41) 335 (711) 32.48
RB Leipzig (6) 752 (1,298) 196 (246) 43 (47) 513 (1005) 26.06
Sport-Club Freiburg (5) 556 (1,044) 128 (167) 34 (45) 394 (832) 23.02
SV Werder Bremen (6) 667 (1,176) 224 (304) 60 (66) 383 (806) 33.59
TSG Hoffenheim (6) 677 (1,187) 226 (313) 52 (59) 399 (815) 33.38
VfB Stuttgart (5) 488 (946) 143 (192) 50 (63) 295 (691) 29.30
VfL Wolfsburg (6) 678 (1,181) 173 (215) 64 (77) 441 (889) 25.52

Turnovers in which neither of the two strategies can be identified are labeled as
undefined or uncontrolled transition situations. As pointed out in this study, this usually
occurs due to short or very uncontrolled ball possession phases (e.g. headers after a
corner situation). These situations typically exhibit a very low (relative) individual ball
possession time in the three seconds after the turnover. Therefore, it has a high impact
on the resulting XGBoost model (see Fig. 2). In total, 20, 928 defensive transition
situations from the first eleven matchdays of the 2018/2019 Bundesliga season were
labeled based on the definitions formulated above (and in Sect. 2.1.2). The task was
to label situations with a detectable strategy with either counterpressing or fallback
and not to label any defensive transitions where no strategy was noticeable. Through
this procedure the students with a background in football tactics excluded on average
62.50% of all turnover situations by implicitly labeling them as undefined. In total, out
of 11, 108 relevant defensive turnovers (after the inclusion criteria), 3, 196 situations
were labeled as counterpressing, 970 as fallback, and 6, 942 were explicitly dropped
as uncontrolled. Table 7 shows the outcome of the full labeling.
Further evaluation of the labeling outcome showed that more than 95% of all as
fallback detected situations start with a goalkeeper catching the ball, what can also be
queried solely rule-based on the event data. Due to this fact and because it was our
focus to investigate counterpressing situations we decided to exclude the distinction

123
Data-driven detection of counterpressing in professional football

between fallbacks and undefined ball possession changes for all further investigations.
Note that fallback situations last on average 18.30 s, whereas undefined situations were
the shortest with 7.83 s on average, which again highlights the tremendous influence
on the defensive reaction time (counterpressing: 9.89 s; all turnovers average: 9.34 s).

123
 Appendix B

123
Table 8 Counterpressing per team for Bundesliga seasons 2013/2014 to 2019/2020

Team (Games) T/M CP+/T (%) %CP S+/ − G+/ − % CP+ %CP − % S+ % G+ % S− % G−

FC Bayern München (229) 112.21 8.02 24.63 − 0.24 0.03 32.56 67.44 3.90 0.66 7.10 0.80
Borussia Dortmund (229) 119.57 8.51 24.87 − 0.16 − 0.04 34.20 65.80 3.61 0.44 6.29 0.87
RB Leipzig (127) 126.69 7.84 22.13 − 0.02 − 0.06 35.41 64.59 3.79 0.51 6.00 1.13
FC Bayer 04 Leverkusen (229) 125.59 8.13 22.91 − 0.24 − 0.04 35.48 64.52 3.66 0.46 6.96 0.92
Borussia Mönchengladbach (229) 104.44 7.27 24.60 − 0.69 − 0.12 29.54 70.4 3.18 0.41 8.32 1.23
Schalke 04 (229) 113.48 7.06 23.27 − 0.59 − 0.11 30.33 69.67 3.11 0.31 7.64 1.07
TSG 1899 Hoffenheim (228) 115.36 7.48 23.27 − 0.43 − 0.04 32.16 67.84 3.82 0.47 8.02 0.94
VfL Wolfsburg (229) 111.14 6.96 23.88 − 0.49 − 0.15 29.14 70.86 2.93 0.20 6.76 1.07
AVERAGE (4118) 115.60 7.25 23.08 − 0.51 − 0.11 31.42% 68.58% 3.22% 0.35% 7.51% 1.13%
Hertha BSC (229) 113.14 7.36 23.67 − 0.50 − 0.10 31.09 68.91 2.58 0.31 6.46 1.02%
1. FC Köln (161) 115.12 6.46 22.45 − 0.54 − 0.16 28.79 71.21 2.91 0.24 7.02% 1.18%
1. FC Augsburg (229) 111.49 6.75 22.53 − 0.62 − 0.14 29.95 70.05 3.25 0.28% 8.14% 1.19%
Eintracht Frankfurt (228) 119.78 7.44 23.50 − 0.68 − 0.10 31.66 68.34 2.77 0.37 7.61 1.07%
SV Werder Bremen (228) 115.01 6.87 23.06 − 0.58 − 0.19 29.79 70.21 3.31 0.20 7.84 1.32%
1. FSV Mainz 05 (229) 120.28 7.48 22.95 − 0.45 − 0.13 32.58 67.42 3.40 0.32 7.46 1.15%
Sport-Club Freiburg (195) 111.33 6.14 20.87 − 0.83 − 0.11 29.40 70.60 2.85 0.26 9.07 1.06%
Fortuna Düsseldorf (59) 103.44 6.83 24.56 − 0.78 − 0.37 27.82 72.18 2.94 0.33 8.32 2.50%
1. FC Union Berlin (25) 114.96 5.50 18.65 − 0.48 − 0.24 29.48 70.52 2.99 0.00 7.41 1.59%
VfB Stuttgart (170) 113.80 7.44 23.41 − 0.50 − 0.14 31.80 68.20 3.33 0.33 7.64 1.23%
ING (68) 147.91 7.55 21.10 − 0.35 − 0.06 35.77 64.23 2.59 0.38 5.80 0.88
Hannover 96 (170) 114.62 6.43 22.05 − 1.02 − 0.25 29.14 70.86 2.61 0.14 9.40 1.61%
Hamburger SV (170) 120.76 7.10 22.05 − 0.42 − 0.21 32.21 67.79 3.42 0.35 7.40 1.66%
SV Darmstadt 98 (68) 118.96 6.21 18.90 − 0.88 − 0.04 32.83 67.17 2.55 0.33 9.64 0.78%
1. FC Nürnberg (68) 102.16 6.36 24.37 − 1.06 − 0.19 26.11 73.89 2.78 0.24 9.51 1.36%
Eintracht Braunschweig (33) 104.42 6.99 22.20 − 0.42 − 0.09 31.50 68.50 3.01 0.13 7.06 0.76%
SC Paderborn 07 (59) 115.51 7.31 23.15 − 0.12 − 0.12 31.56 68.44 3.36 0.19 5.56 0.93%
P. Bauer, G. Anzer
Data-driven detection of counterpressing in professional football

Appendix C

Table 9 Counterpressing per coach for Bundesliga seasons 2013/2014 to 2019/2020. Only full seasons
with in total 34 matches per coach are taken into consideration

Coach T/M CP+/T %CP S+/− G+/− % CP+ %CP− % S+ % G+ % S− % G−

(Games) (%)

Pep Guardiola 109.47 8.61 25.24 − 0.21 0.02 34.10 65.90 3.87 0.50 7.00 0.65
(102)
Thomas 113.00 8.60 24.61 − 0.27 − 0.06 34.93 65.07 3.49 0.53 6.88 1.14
Tuchel (102)
Jürgen Klopp 129.00 8.54 23.51 0.26 - 0.04 36.32 63.68 4.22 0.29 5.26 0.69
(68)
Roger 156.29 8.49 18.84 0.91 0.09 45.05 54.95 5.29 0.80 4.00 0.91
Schmidt (34)
Jos Luhukay 104.97 8.27 24.99 − 0.12 − 0.09 33.07 66.93 2.80 0.11 4.86 0.67
(34)
Adi Hütter 128.94 8.05 23.27 − 0.59 − 0.09 34.61 65.3 3.63 0.39 8.55 1.05
(34)
Markus Gisdol 141.03 8.03 22.34 0.24 0.12 35.95 64.05 4.30 0.56 5.54 0.29
(34)
Ralph 136.35 8.00 21.99 − 0.27 − 0.03 36.38 63.62 2.91 0.56 6.01 1.03
Hasenhttül
(102)
Sandro 120.76 7.93 24.05 − 0.26 − 0.18 32.96 67.0 3.70 0.25 6.87 1.28
Schwarz
(68)
Carlo 120.03 7.69 23.52 − 0.56 0.15 32.71 67.29 2.60 0.83 6.81 0.46
Ancelotti
(34)
Heiko Herrlich 115.71 7.68 23.64 − 0.44 − 0.03 32.47 67.53 3.76 0.54 7.96 0.96
(34)
Martin 128.29 7.62 21.92 − 0.46 − 0.06 34.78 65.22 2.98 0.37 7.06 0.88
Schmidt (68)
Domenico 113.94 7.56 23.88 − 0.12 − 0.09 31.68 68.32 3.35 0.22 5.54 0.79
Tedesco (34)
Niko Kovac 118.18 7.46 23.92 − 0.28 − 0.03 31.18 68.82 3.09 0.55 5.95 0.96
(102)
AVERAGE 116.22 7.40 23.06 - 0.39 - 0.08 32.07 67.93 3.28% 0.38% 6.99% 1.01%
(2176)
Julian 110.10 7.36 23.70 − 0.38 0.03 31.03 68.97 4.43 0.64 8.55 0.76
Nagelsmann
(102)
Thomas 127.44 7.29 22.62 − 0.24 − 0.12 32.24 67.76 2.96 0.31 5.57 1.05
Schaaf (34)

123
 P. Bauer, G. Anzer

Table 9 continued

Coach T/M CP+/T %CP S+/− G+/− % CP+ %CP− % S+ % G+ % S− % G−

(Games) (%)

Dieter 106.00 7.27 24.13 − 0.41 − 0.14 30.13 69.87 3.40 0.39 7.18 1.32
Hecking
(170)
Florian 111.24 7.22 23.43 − 0.71 − 0.26 30.81 69.19 3.61 0.11 9.14 1.63
Kohfeldt
(34)
Markus 109.07 7.21 23.53 − 0.46 − 0.10 30.66 69.34 3.30 0.32 7.31 0.99
Weinzierl
(136)
Pal Dardai 113.18 7.15 23.36 − 0.69 − 0.08 30.62 69.38 2.50 0.33 7.37 0.92
(136)
Bruno 114.91 7.06 24.75 − 0.63 − 0.19 28.54 71.46 2.95 0.21 7.24 1.23
Labbadia
(68)
Robin Dutt 106.94 7.04 23.93 − 0.56 − 0.24 29.43 70.57 3.45 0.23 7.98 1.63
(34)
Ralf Rangnick 131.47 7.00 21.72 0.56 − 0.03 32.23 67.77 4.63 0.41 3.95 0.76
(34)
Lucien Favre 94.10 6.95 25.38 − 0.87 − 0.03 27.40 72.60 2.65 0.43 8.65 0.76
(68)
Armin Veh 93.38 6.93 25.98 − 0.79 − 0.24 26.67 73.33 2.91 0.00 8.43 1.32
(34)
Andr 117.86 6.82 22.20 − 0.67 − 0.18 30.72 69.28 3.07 0.26 8.11 1.35
Breitenreiter
(102)
Viktor 125.35 6.69 21.73 − 0.32 − 0.18 30.78 69.22 3.02 0.32 6.08 1.40
Skripnik
(34)
Friedhelm 100.44 6.62 24.77 − 0.59 − 0.47 26.71 73.29 3.43 0.35 7.90 3.06
Funkel (34)
Peter Stöger 116.09 6.60 21.81 − 0.35 −0.12 30.25 69.75 2.48 0.19 5.55 0.94
(102)
Manuel Baum 121.62 6.41 21.45 − 0.65 0.00 29.88 70.12 2.71 0.34 7.40 0.48
(34)
Christian 114.29 6.26 20.34 − 0.68 − 0.07 30.77 69.23 2.66 0.25 8.04 0.78
Streich (136)
Dirk Schuster 125.82 5.82 16.88 − 0.65 − 0.03 34.49 65.51 2.35 0.28 8.25 0.63
(34)

123
 Appendix D
Table 10 Counterpressing of Bundesliga teams on a season level (1/2). The number in brackets shows the respective position in the final ranking

Team (Ranking) T/M CP+/T %CP S+/− G+/− % CP+ %CP− % S+ % G+ % S− % G−

FCB - 2013/2014 (1) 97.29 8.98% 26.18% − 0.71 0.03 34.30% 65.70% 3.58% 0.46% 9.67% 0.53%
FCB - 2014/2015 (1) 116.65 8.50% 24.79% 0.00 0.06 34.28% 65.72% 3.87% 0.51% 5.88% 0.46%
FCB - 2015/2016 (1) 114.47 8.40% 24.90% 0.09 − 0.03 33.75% 66.25% 4.13% 0.52% 5.76% 0.93%
FCB - 2019/2020 (1) 116.68 8.40% 26.02% − 0.32 − 0.04 32.28% 67.72% 4.87% 0.66% 8.75% 1.17%
FCB - 2016/2017 (1) 120.03 7.69% 23.52% − 0.56 0.15 32.71% 67.29% 2.60% 0.83% 6.81% 0.46%
FCB - 2018/2019 (1) 111.15 7.33% 24.21% 0.32 0.12 30.27% 69.73% 5.14% 1.20% 5.64% 1.10%
FCB - 2017/2018 (1) 110.41 7.03% 23.39% − 0.56 − 0.06 30.07% 69.93% 3.30% 0.46% 7.82% 0.98%
BVB - 2015/2016 (2) 120.79 9.40% 24.71% − 0.09 0.03 38.03% 61.97% 3.15% 0.59% 5.56% 0.79%
Data-driven detection of counterpressing in professional football

BVB - 2013/2014 (2) 113.12 8.71% 24.60% − 0.12 − 0.09 35.41% 64.59% 3.59% 0.21% 6.22% 0.82%
BVB - 2018/2019 (2) 105.82 8.59% 26.57% − 0.32 − 0.12 32.32% 67.68% 2.41% 0.21% 5.26% 0.93%
RBL - 2016/2017 (2) 134.97 8.35% 21.16% − 0.24 − 0.06 39.44% 60.56% 2.68% 0.72% 5.78% 1.53%
BVB - 2019/2020 (2) 104.00 8.19% 27.65% − 0.32 − 0.08 29.62% 70.38% 4.17% 0.70% 7.51% 1.38%
S04 - 2017/2018 (2) 113.94 7.56% 23.88% − 0.12 − 0.09 31.68% 68.32% 3.35% 0.22% 5.54% 0.79%
WOB - 2014/2015 (2) 118.44 7.18% 22.00% 0.15 0.00 32.62% 67.38% 4.63% 0.45% 6.03% 0.67%
LEV - 2015/2016 (3) 143.85 8.79% 22.33% − 0.24 0.15 39.38% 60.62% 2.66% 0.55% 5.59% 0.15%
BVB - 2016/2017 (3) 122.38 8.29% 24.75% − 0.24 − 0.03 33.50% 66.50% 3.69% 0.58% 6.72% 1.02%

123
123
Table 10 continued
Team (Ranking) T/M CP+/T %CP S+/− G+/− % CP+ %CP− % S+ % G+ % S− % G−

RBL - 2019/2020 (3) 110.60 7.88% 22.21% − 0.20 − 0.12 35.50% 64.50% 4.07% 0.33% 7.58% 1.26%
HOF - 2017/2018 (3) 109.15 7.49% 23.31% − 0.18 0.12 32.14% 67.86% 4.74% 0.81% 8.01% 0.51%
RBL - 2018/2019 (3) 131.47 7.00% 21.72% 0.56 − 0.03 32.23% 67.77% 4.63% 0.41% 3.95% 0.76%
S04 - 2013/2014 (3) 96.74 6.87% 23.44% − 0.03 0.03 29.31% 70.69% 4.02% 0.65% 5.87% 0.73%
BMG - 2014/2015 (3) 103.12 6.85% 24.02% − 0.62 − 0.03 28.50% 71.50% 2.85% 0.36% 7.48% 0.66%
LEV - 2014/2015 (4) 156.29 8.49% 18.84% 0.91 0.09 45.05% 54.95% 5.29% 0.80% 4.00% 0.91%
LEV - 2019/2020 (4) 114.76 8.16% 23.60% − 0.64 − 0.12 34.56% 65.44% 3.10% 0.00% 8.35% 0.68%
BVB - 2017/2018 (4) 121.85 7.89% 24.79% − 0.71 0.00 31.84% 68.16% 3.51% 0.49% 8.57% 0.71%
BMG - 2015/2016 (4) 117.32 7.60% 24.34% − 0.47 − 0.15 31.20% 68.80% 3.30% 0.41% 7.19% 1.35%
HOF - 2016/2017 (4) 107.94 7.52% 23.98% − 0.35 − 0.06 31.36% 68.64% 3.64% 0.34% 7.28% 0.83%
LEV - 2013/2014 (4) 101.26 7.44% 26.08% − 0.38 − 0.06 28.51% 71.49% 3.01% 0.56% 6.23% 1.09%
LEV - 2018/2019 (4) 114.68 7.28% 24.26% − 0.29 − 0.15 30.02% 69.98% 4.65% 0.21% 8.16% 1.06%
LEV - 2017/2018 (5) 115.71 7.68% 23.64% − 0.44 − 0.03 32.47% 67.53% 3.76% 0.54% 7.96% 0.96%
AUG - 2014/2015 (5) 121.35 7.54% 22.95% − 0.85 − 0.21 32.84% 67.16% 2.75% 0.21% 8.65% 1.42%
WOB - 2013/2014 (5) 94.82 7.44% 24.88% − 0.41 − 0.12 29.93% 70.07% 2.24% 0.12% 5.69% 0.89%
BMG - 2018/2019 (5) 99.29 7.35% 26.04% − 0.85 − 0.29 28.21% 71.79% 3.98% 0.34% 10.14% 2.06%
KOE - 2016/2017 (5) 113.41 6.95% 22.15% − 0.26 − 0.12 31.38% 68.62% 2.34% 0.00% 4.95% 0.68%
S04 - 2015/2016 (5) 115.44 6.80% 22.39% − 0.76 − 0.12 30.38% 69.62% 3.30% 0.34% 8.99% 1.14%
P. Bauer, G. Anzer
 Table 10 continued
Team (Ranking) T/M CP+/T %CP S+/− G+/− % CP+ %CP− % S+ % G+ % S− % G−

BMG - 2019/2020 (5) 110.48 6.59% 22.95% − 0.56 − 0.24 28.71% 71.29% 2.84% 0.32% 7.08% 1.77%
RBL - 2017/2018 (6) 125.44 8.14% 23.56% − 0.26 − 0.06 34.53% 65.47% 3.88% 0.50% 7.29% 1.06%
M05 - 2015/2016 (6) 123.71 7.75% 22.92% − 0.50 0.03 33.82% 66.18% 3.01% 0.31% 7.21% 0.31%
HBER - 2016/2017 (6) 115.29 7.73% 22.65% − 0.47 − 0.03 34.12% 65.88% 2.48% 0.34% 6.50% 0.68%
WOB - 2018/2019 (6) 110.38 7.49% 27.15% − 0.79 − 0.21 27.58% 72.42% 2.85% 0.29% 7.59% 1.36%
BMG - 2013/2014 (6) 85.09 7.09% 27.03% − 1.12 − 0.03 26.21% 73.79% 2.43% 0.51% 9.88% 0.87%
WOB - 2019/2020 (6) 123.36 6.74% 23.09% − 0.40 − 0.12 29.21% 70.79% 3.23% 0.14% 6.55% 0.79%
S04 - 2014/2015 (6) 123.56 6.19% 20.97% − 1.03 0.00 29.51% 70.49% 2.50% 0.34% 9.18% 0.48%
BVB - 2014/2015 (7) 144.88 8.40% 22.66% 0.65 0.00 37.10% 62.90% 4.75% 0.36% 4.42% 0.57%
FRA - 2018/2019 (7) 128.94 8.05% 23.27% − 0.59 − 0.09 34.61% 65.39% 3.63% 0.39% 8.55% 1.05%
Data-driven detection of counterpressing in professional football

M05 - 2013/2014 (7) 95.82 7.98% 24.31% − 0.50 − 0.18 32.83% 67.17% 3.66% 0.38% 8.65% 1.69%
STU - 2017/2018 (7) 114.26 7.54% 24.53% − 0.71 − 0.06 30.75% 69.25% 2.31% 0.31% 6.97% 0.76%
HBER - 2015/2016 (7) 111.53 7.30% 22.97% − 0.56 − 0.09 31.80% 68.20% 2.18% 0.34% 6.40% 1.01%
SCF - 2016/2017 (7) 122.44 5.79% 19.10% − 0.76 − 0.03 30.31% 69.69% 2.26% 0.25% 7.94% 0.54%
SCF - 2019/2020 (7) 106.48 5.48% 22.46% − 1.84 − 0.44 24.41% 75.59% 3.18% 0.17% 14.38% 2.65%
HOF - 2014/2015 (8) 141.03 8.03% 22.34% 0.24 0.12 35.95% 64.05% 4.30% 0.56% 5.54% 0.29%
BRE - 2018/2019 (8) 111.24 7.22% 23.43% − 0.71 − 0.26 30.81% 69.19% 3.61% 0.11% 9.14% 1.63%
S04 - 2019/2020 (8) 113.68 7.14% 24.77% − 0.88 − 0.24 28.84% 71.16% 2.84% 0.14% 8.38% 1.40%

123
123
Table 10 continued
Team (Ranking) T/M CP+/T %CP S+/− G+/− % CP+ %CP− % S+ % G+ % S− % G−

AUG - 2013/2014 (8) 93.50 7.14% 25.01% − 0.12 0.00 28.55% 71.45% 3.65% 0.63% 5.81% 0.88%
FRA - 2017/2018 (8) 120.18 7.12% 25.01% − 0.97 − 0.12 28.47% 71.53% 2.25% 0.39% 7.66% 1.09%
BRE - 2016/2017 (8) 110.15 6.92% 23.68% − 0.76 − 0.24 29.20% 70.80% 3.16% 0.34% 8.60% 1.75%
WOB - 2015/2016 (8) 109.85 6.53% 23.51% − 0.35 − 0.26 27.79% 72.21% 2.62% 0.34% 5.52% 1.89%
BMG - 2017/2018 (9) 107.59 7.90% 24.69% − 0.59 0.00 32.00% 68.00% 3.43% 0.66% 8.31% 0.98%
HOF - 2013/2014 (9) 108.24 7.61% 23.91% − 0.33 − 0.18 31.85% 68.15% 3.28% 0.23% 6.70% 1.37%
FRA - 2014/2015 (9) 127.44 7.29% 22.62% − 0.24 − 0.12 32.24% 67.76% 2.96% 0.31% 5.57% 1.05%
BMG - 2016/2017 (9) 109.79 7.26% 23.36% − 0.59 − 0.12 31.08% 68.92% 3.21% 0.23% 7.99% 1.00%
AVERAGE (9) 115.60 7.25% 23.08% − 0.51 − 0.11 31.42% 68.58% 3.22% 0.35% 7.51% 1.13%
HOF - 2018/2019 (9) 113.21 7.07% 23.82% − 0.62 0.03 29.66% 70.34% 4.91% 0.76% 10.23% 0.93%
KOE - 2015/2016 (9) 115.47 6.11% 21.88% − 0.59 − 0.12 27.94% 72.06% 2.10% 0.12% 6.14% 0.81%
HOF - 2019/2020 (9) 102.44 5.66% 22.73% − 1.28 − 0.12 24.91% 75.09% 2.75% 0.69% 10.98% 1.60%
S04 - 2016/2017 (10) 114.38 7.17% 23.68% − 0.47 − 0.09 30.29% 69.71% 3.04% 0.43% 6.85% 1.09%
BRE - 2014/2015 (10) 127.21 6.68% 21.39% − 0.44 − 0.12 31.24% 68.76% 3.57% 0.00% 7.55% 0.63%
HSV - 2015/2016 (10) 119.44 6.67% 22.53% − 0.47 − 0.18 29.62% 70.38% 3.06% 0.11% 6.83% 1.09%
DUE - 2018/2019 (10) 100.44 6.62% 24.77% − 0.59 − 0.47 26.71% 73.29% 3.43% 0.35% 7.90% 3.06%
H96 - 2013/2014 (10) 98.32 6.61% 23.24% − 0.38 − 0.15 28.44% 71.56% 4.12% 0.26% 8.09% 1.26%
P. Bauer, G. Anzer
 Table 10 continued
Team (Ranking) T/M CP+/T %CP S+/− G+/− % CP+ %CP− % S+ % G+ % S− % G−

KOE - 2019/2020 (10) 106.52 6.57% 23.73% − 0.96 − 0.28 27.69% 72.31% 3.16% 0.00% 9.63% 1.53%
HBER - 2017/2018 (10) 112.94 6.25% 23.41% − 0.62 − 0.09 26.70% 73.30% 2.89% 0.33% 7.13% 0.91%
HBER - 2013/2014 (11) 104.97 8.27% 24.99% − 0.12 − 0.09 33.07% 66.93% 2.80% 0.11% 4.86% 0.67%
FRA - 2016/2017 (11) 123.21 7.90% 22.58% − 0.21 − 0.09 34.99% 65.01% 2.01% 0.11% 4.23% 0.65%
ING - 2015/2016 (11) 148.65 7.58% 21.43% − 0.32 0.03 35.36% 64.64% 2.22% 0.46% 5.00% 0.57%
HBER - 2019/2020 (11) 104.04 7.38% 25.03% 29.49% 70.51% 2.76% 0.31% 6.75% 1.74%
Data-driven detection of counterpressing in professional football

− 0.52 − 0.24
HBER - 2018/2019 (11) 112.94 7.32% 24.43% − 1.12 − 0.12 29.96% 70.04% 2.45% 0.32% 9.28% 1.07%
BRE - 2017/2018 (11) 116.06 6.79% 24.00% − 0.68 − 0.12 28.30% 71.70% 2.75% 0.11% 7.22% 0.74%
M05 - 2014/2015 (11) 129.59 6.13% 21.56% − 0.59 − 0.09 28.42% 71.58% 3.47% 0.21% 7.79% 0.74%

123
 Table 11 Counterpressing of Bundesliga teams on a season level (2/2). The number in brackets shows the respective position in the final ranking

Team (Ranking) T/M CP+/T %CP S+/− G+/- % CP+ %CP- % S+ % G+ % S- % G-

123
LEV - 2016/2017 (12) 129.71 8.64% 23.70% − 0.71 − 0.18 36.46% 63.54% 3.06% 0.38% 8.43% 1.51%
M05 - 2018/2019 (12) 121.79 7.49% 24.70% − 0.62 − 0.26 30.30% 69.70% 3.42% 0.10% 7.85% 1.40%
BRE - 2013/2014 (12) 106.94 7.04% 23.93% − 0.56 − 0.24 29.43% 70.57% 3.45% 0.23% 7.98% 1.63%
AUG - 2015/2016 (12) 107.03 6.95% 22.73% − 0.38 -0.09 30.59% 69.41% 3.87% 0.00% 7.84% 0.52%
AUG - 2019/2020 (12) 104.40 6.82% 22.68% − 0.52 -0.36 30.07% 69.93% 3.55% 0.68% 8.21% 3.14%
KOE - 2014/2015 (12) 119.38 6.73% 21.41% − 0.21 -0.12 31.42% 68.58% 2.99% 0.46% 5.54% 1.34%
AUG - 2017/2018 (12) 121.62 6.41% 21.45% − 0.65 0.00 29.88% 70.12% 2.71% 0.34% 7.40% 0.48%
FRA - 2013/2014 (13) 93.38 6.93% 25.98% − 0.79 − 0.24 26.67% 73.33% 2.91% 0.00% 8.43% 1.32%
BRE - 2015/2016 (13) 125.35 6.69% 21.73% − 0.32 − 0.18 30.78% 69.22% 3.02% 0.32% 6.08% 1.40%
H96 - 2017/2018 (13) 115.74 6.45% 22.21% − 1.06 -0.35 29.06% 70.94% 2.52% 0.11% 9.35% 2.10%
AUG - 2016/2017 (13) 118.97 6.18% 21.63% − 0.71 -0.09 28.57% 71.43% 3.20% 0.23% 8.32% 0.80%
H96 - 2014/2015 (13) 130.62 6.10% 19.88% − 1.06 -0.21 30.69% 69.31% 1.81% 0.00% 8.50% 1.14%
SCF - 2018/2019 (13) 103.03 6.08% 22.01% − 0.68 − 0.06 27.63% 72.37% 3.37% 0.39% 8.78% 0.90%
UBER - 2019/2020 (13) 114.96 5.50% 18.65% − 0.48 − 0.24 29.48% 70.52% 2.99% 0.00% 7.41% 1.59%
M05 - 2017/2018 (14) 119.74 8.38% 23.38% 0.09 − 0.09 35.82% 64.18% 3.99% 0.42% 5.73% 1.15%
S04 - 2018/2019 (14) 116.68 7.71% 24.33% − 0.88 − 0.32 31.71% 68.29% 2.80% 0.10% 8.65% 1.82%
HSV - 2016/2017 (14) 139.88 7.59% 20.90% 0.24 0.03 36.32% 63.68% 3.52% 0.50% 4.27% 0.63%
STU - 2014/2015 (14) 120.24 7.46% 22.06% − 0.65 − 0.26 33.81% 66.19% 2.88% 0.22% 8.04% 1.84%
FRA - 2019/2020 (14) 123.00 7.35% 21.71% − 1.12 − 0.12 33.85% 66.15% 3.28% 0.62% 11.32% 1.65%
SCF - 2013/2014 (14) 96.24 6.57% 22.31% − 0.79 − 0.12 29.45% 70.55% 2.33% 0.14% 8.54% 0.97%
P. Bauer, G. Anzer
 Table 11 continued

Team (Ranking) T/M CP+/T %CP S+/− G+/- % CP+ %CP- % S+ % G+ % S- % G-

DAR - 2015/2016 (14) 125.82 5.82% 16.88% − 0.65 − 0.03 34.49% 65.51% 2.35% 0.28% 8.25% 0.63%
STU - 2013/2014 (15) 92.38 8.28% 25.66% 0.03 − 0.03 32.26% 67.74% 4.22% 0.50% 6.04% 0.92%
HOF - 2015/2016 (15) 121.85 8.21% 22.95% − 0.74 − 0.24 35.75% 64.25% 2.73% 0.00% 8.35% 1.31%
M05 - 2016/2017 (15) 132.88 7.50% 20.98% − 0.41 − 0.15 35.76% 64.24% 2.95% 0.42% 6.90% 1.48%
HBER - 2014/2015 (15) 127.85 7.34% 22.87% − 0.12 − 0.12 32.09% 67.91% 2.52% 0.40% 4.30% 1.19%
M05 - 2019/2020 (15) 117.72 7.24% 23.48% − 0.68 − 0.16 30.82% 69.18% 3.33% 0.43% 8.37% 1.46%
SCF - 2017/2018 (15) 114.32 6.51% 20.63% − 0.68 − 0.18 31.55% 68.45% 2.87% 0.12% 8.38% 1.28%
AUG - 2018/2019 (15) 111.68 6.29% 21.83% − 1.06 − 0.29 28.83% 71.17% 3.26% 0.00% 10.68% 1.69%
HSV - 2013/2014 (16) 95.32 7.44% 22.96% − 0.59 − 0.44 32.39% 67.61% 4.17% 0.40% 10.14% 3.58%
FRA - 2015/2016 (16) 123.26 7.25% 23.50% − 1.00 0.06 30.86% 69.14% 2.54% 0.81% 8.66% 0.88%
DUE - 2019/2020 (16) 107.52 7.11% 24.29% − 1.04 − 0.24 29.25% 70.75% 2.30% 0.31% 8.87% 1.73%
WOB - 2016/2017 (16) 117.50 6.93% 22.38% − 0.74 − 0.18 30.98% 69.02% 2.46% 0.00% 7.62% 0.97%
STU - 2018/2019 (16) 116.62 6.86% 22.65% − 0.59 − 0.18 30.29% 69.71% 3.79% 0.11% 8.63% 1.12%
Data-driven detection of counterpressing in professional football

HSV - 2014/2015 (16) 119.79 6.58% 21.46% − 0.50 − 0.21 30.66% 69.34% 3.20% 0.46% 7.43% 1.82%
WOB - 2017/2018 (16) 106.88 6.38% 24.38% − 0.88 − 0.15 26.19% 73.81% 2.48% 0.00% 7.95% 0.76%
ING - 2016/2017 (17) 147.18 7.51% 20.76% − 0.38 − 0.15 36.19% 63.81% 2.98% 0.29% 6.64% 1.21%
STU - 2015/2016 (17) 125.50 7.27% 22.73% − 0.59 − 0.15 31.96% 68.04% 3.61% 0.52% 8.33% 1.52%
HSV - 2017/2018 (17) 129.35 7.21% 22.71% − 0.79 − 0.24 31.73% 68.27% 3.30% 0.30% 8.80% 1.61%

123
 Table 11 continued

Team (Ranking) T/M CP+/T %CP S+/− G+/- % CP+ %CP- % S+ % G+ % S- % G-

123
NUR - 2013/2014 (17) 99.47 6.86% 24.78% − 1.09 − 0.12 27.68% 72.32% 3.58% 0.36% 11.06% 1.16%
BRE - 2019/2020 (17) 105.25 6.81% 24.03% − 0.62 − 0.21 28.34% 71.66% 3.79% 0.33% 8.74% 1.61%
H96 - 2018/2019 (17) 109.35 6.70% 23.02% − 1.59 − 0.32 29.09% 70.91% 2.22% 0.12% 12.03% 1.98%
SCF - 2014/2015 (17) 124.18 6.25% 19.78% − 0.47 0.06 31.62% 68.38% 3.11% 0.48% 7.36% 0.35%
PAD - 2019/2020 (18) 106.12 7.50% 24.95% − 0.04 − 0.20 30.06% 69.94% 3.32% 0.00% 4.97% 1.08%
PAD - 2014/2015 (18) 122.41 7.18% 22.01% − 0.18 − 0.06 32.64% 67.36% 3.38% 0.33% 6.00% 0.81%
BRA - 2013/2014 (18) 104.42 6.99% 22.20% − 0.42 − 0.09 31.50% 68.50% 3.01% 0.13% 7.06% 0.76%
DAR - 2016/2017 (18) 112.09 6.64% 21.18% − 1.12 − 0.06 31.35% 68.65% 2.73% 0.37% 10.83% 0.90%
H96 - 2015/2016 (18) 119.06 6.35% 22.38% − 1.03 − 0.24 28.37% 71.63% 2.54% 0.22% 8.94% 1.54%
KOE - 2017/2018 (18) 118.53 6.00% 23.50% − 0.79 − 0.18 25.55% 74.45% 3.91% 0.53% 9.08% 1.56%
NUR - 2018/2019 (18) 104.85 5.89% 23.98% − 1.03 − 0.26 24.56% 75.44% 1.99% 0.12% 8.06% 1.55%
P. Bauer, G. Anzer
Data-driven detection of counterpressing in professional football

References
Andrienko G et al (2017) Visual analysis of pressure in football. Data Mining Knowl Discov 31(6):1793–
1839. https://doi.org/10.1007/s10618-017-0513-2
Andrienko G et al (2019) Constructing Spaces and Times for Tactical Analysis in Football. IEEE Trans Vis
Comput Graph. https://doi.org/10.1109/tvcg.2019.2952129
Antipov EA, Pokryshevskaya EB (2020) Interpretable machine learning for demand modeling with high-
dimensional data using Gradient Boosting Machines and Shapley values. J Revenue Pricing Manag
19(5):355–364. https://doi.org/10.1057/s41272-020-00236-4
Anzer G, Bauer P (2021) “A Goal Scoring Probability Model based on Synchronized Positional and Event
Data”. Frontiers in Sports and Active Learning (in print) Using Artificial Intelligence to Enhance Sport
Performance, pp. 1-18. https://doi.org/10.3389/fspor.2021.624475.
Bergstra J et al (2011) “Algorithms for hyper-parameter optimization”. In: Advances in Neural Information
Processing Systems 24: 25th Annual Conference on Neural Information Processing Systems 2011,
NIPS 2011, pp. 1–9
Bialkowski A et al (2014) “Large-Scale Analysis of Soccer Matches Using Spatiotemporal Tracking Data”.
In: Proceedings - IEEE International Conference on Data Mining, ICDM 2015-January, pp. 725–730.
issn:15504786. https://doi.org/10.1109/ICDM.2014.133
Bialkowski A et al (2015) “Identifying team style in soccer using formations learned from spatiotemporal
tracking data”. In: IEEE International Conference on Data Mining Workshops, ICDMW 2015.Jan-
uary, pp. 9-14. issn: 23759259. https://doi.org/10.1109/ICDMW.2014.167.29 https://gdpr-info.eu/,
accessed 07/20/20.Data-Driven Detection of Counterpressing in Professional Football 15
Bojinov I, Bornn L (2016) “The Pressing Game: Optimal Defensive Disruption in Soccer”. In: MIT Sloan
Sports Analytics Conference, pp. 1–8
Bourbousson Jérôme, Carole Sève Tim McGarry (2010) “Space-time coordination dynamics in basketball:
Part 2. the interaction between the two teams”. In: Journal of Sports Sciences 28.3, pp. 349–358. issn:
02640414. https://doi.org/10.1080/02640410903503640
Brefeld U, Lasek J, Mair S (2019) Probabilistic movement models and zones of control. Mach Learn
108(1):127–147. https://doi.org/10.1007/s10994-018-5725-1
Chen T, Guestrin C (2016) “XGBoost: A scalable tree boosting system”. In: Proceedings of the ACM
SIGKDD International Conference on Knowledge Discovery and Data Mining 13-17-Augu, pp. 785–
794. https://doi.org/10.1145/2939672.2939785.
Decroos T, Van Haaren J, Davis J (2018) “Automatic discovery of tactics in spatio-temporal soccer match
data”. In: Proceedings of the ACM SIGKDD International Conference on Knowledge Discovery and
Data Mining. isbn: 9781450355520. https://doi.org/10.1145/3219819.3219832. url: https://people.cs.
kuleuven.be/~jesse.davis/decroos-kdd18.pdf
Decroos T et al (2020) “VAEP: An Objective Approach to Valuing On-the-Ball Actions in Soccer (Extended
Abstract)”. In: Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelli-
gence (IJCAI-20), pp. 4696–4700. issn: 10450823. https://doi.org/10.24963/ijcai.2020/648.
Dewnacker Ian, Michael McCourt, Scott Clark (2016) “Bayesian Optimization for Machine Learning. A
Practical Guidebook”. arXiv:1612.04858
Fairchild A, Pelechrinis K, Kokkodis M (2018) “Spatial analysis of shots in MLS: A model for expected
goals and fractal dimensionality”. In: Journal of Sports Analytics 4.3, pp. 165–174. issn: 2215020X.
https://doi.org/10.3233/jsa-170207.
Navarro Férnandez, J (2018) Analysis of Styles of Play in Soccer and Their Effectiveness. isbn:
9788413060576
Fernandez J, Bornn L (2018) “Wide Open Spaces : A statistical technique for measuring space creation in
professional soccer”. In: MIT Sloan Sports Analytics Conference, pp. 1–19
Goes F, Kempe M, Koen L (2019). “Predicting match outcome in professional Dutch football using tactical
performance metrics computed from position tracking data”. In: MathSport International Conference
June, pp. 4-5
Goes F et al (2020) Interpretable machine learning for demand modeling with high-dimensional data using
Gradient Boosting Machines and Shapley values. Euro J Sport Sci. https://doi.org/10.1080/17461391.
2020.1747552
Grant AG et al (1999) Analysis of the goals scored in the 1998 World Cup. J Sports Sci 17(10):826–827

123
 P. Bauer, G. Anzer

Herold M et al (2019) Machine learning in men’s professional football: current applications and
future directions for improving attacking play. Int J Sports Sci Coaching. https://doi.org/10.1177/
1747954119879350
Hobbs J et al. (2018). “Quantifying the Value of Transitions in Soccer via Spatiotemporal Trajectory
Clustering”. In: MIT Sloan Sports Analytics Conferencece, pp. 1–11
Hughes M, Ian M Franks (2015) Essentials of performance analysis in sport. Vol. 53. 04, pp. 53-1831. https://
doi.org/10.5860/choice.193440. url:https://books.google.de/books?hl=de&lr=&id=KorCDwAAQB
AJ&oi=fnd&pg=PT14&dq=the+essentials+of+performance+analysis&ots=ZhJd6413Fq&sig=F
MnUcL21bpzACnWTzNhxESEgUY4\#v=onepage&q=theessentialsofperformanceanalysis&f=false
Ibrahim L et al (2020) Explainable prediction of acute myocardial infarction using machine learning and
shapley values. IEEE Access. https://doi.org/10.1109/access.2020.3040166
Kempe M et al (2014) Possession vs. direct play: evaluating tactical behavior in elite soccer. Int J Sports
Sci 4(6A):35–41. https://doi.org/10.5923/s.sports.201401.05
Kim S (2004) Voronoi Analysis of a Soccer Game. Nonlinear Anal Model Control 9(3):233–240. https://
doi.org/10.15388/na.2004.9.3.15154
Li TR et al (2019). Sentiment-based prediction of alternative cryptocurrency price fluctuations using gradient
boosting tree model. https://doi.org/10.3389/fphy.2019.00098
Link D, Lang S, Seidenschwarz P (2016) Real time quantification of dangerousity in football using spa-
tiotemporal tracking data. PLoS ONE. https://doi.org/10.1371/journal.pone.0168768
Link D, Hoernig M (2017) Individual ball possession in soccer. PLoS ONE 12(7):1–15. https://doi.org/10.
1371/journal.pone.0179953
Liu K, Chen W, Lin H (2020) XG-PseU: an eXtreme Gradient Boosting based method for identifying
pseudouridine sites. Mole Genet Genom 295(1):13–21. https://doi.org/10.1007/s00438-019-01600-9
Patrick L et al (2014) “Quality vs Quantity”: Improved Shot Prediction in Soccer using Strategic Features
from Spatiotemporal Data”. In: MIT Sloan Sports Analytics Conference, pp. 1–9. url: http://www.
sloansportsconference.com/?p=15790
Lundberg SM, SI Lee (2017) “Consistent feature attribution for tree ensembles”. In: Proceedings of the
34th International Conference on Machine Learning, pp. 1–9
Yuan M et al (2020) What Makes an Online Review More Helpful: An Interpretation Framework Using
XGBoost and SHAP Values. J Theor Appl Electron Commerce Res 16(3):466–490. https://doi.org/
10.3390/jtaer16030029
Pappalardo L et al (2019) A public data set of spatio-temporal match events in soccer competitions. Scientific
Data 6(1):236. https://doi.org/10.1038/s41597-019-0247-7
Paul P et al (2018) “Mythbusting Set-Pieces in Soccer”. In: MIT Sloan Sports Analytics Conference, pp.
1-12
Rathke A (2017) “An examination of expected goals and shot efficiency in soccer”. In: Journal of Human
Sport and Exercise 12.Proc2. issn: 1988-5202. https://doi.org/10.14198/jhse.2017.12.proc2.05. url:
http://www.redalyc.org/articulo.oa?id=301052437005
Ratner Alexander J et al (2017) “Learning to compose domain-specific transformations for data augmenta-
tion”. In: Advances in Neural Information Processing Systems Nips, pp. 3237–3247. issn: 10495258
Ratner A et al (2016) “Data programming: Creating large training sets, quickly”. In: Advances in Neural
Information Processing Systems Nips, pp. 3574–3582. issn: 10495258
Reep C, Benjamin B (1968) Skill and Chance in Association Football. J Royal Stat Soc Series A (General).
https://doi.org/10.2307/2343726
Robert R, Memmert D (2016) “Big data and tactical analysis in elite soccer: future challenges and opportu-
nities for sports science”. In: SpringerPlus 5.1. issn: 21931801. https://doi.org/10.1186/s40064-016-
3108-2.
Robberechts P (2019) “Valuing the Art of Pressing”. In: StatsBomb Innovation in Football Conference
2019, p. 11. url: http://statsbomb.com/wp-content/uploads/2019/10/Pieter-Robberechts-Valuing-the-
Art-of-Pressing.pdf
Roth Alvin E, Thomson W (1988) The Shapley Value: Essays in Honor of Lloyd S. Shapley. isbn:
052136177X. https://doi.org/10.2307/2554979.
Santos A. Benito et al (2018) “Data-driven visual performance analysis in soccer: An exploratory prototype”.
In: Frontiers in Psychology 9. issn: 16641078. https://doi.org/10.3389/fpsyg.2018.02416.

123
Data-driven detection of counterpressing in professional football

Shaw L, Glickman M (2019) “Dynamic analysis of team strategy in pro-

fessional football”. In: Barça sports analytics summit. Retrieved from
https://static.capabiliaserver.com/frontend/clients/barca/wp_prod/wp-content/uploads/2020/01/
56ce723e-barca-conferencepaper-laurie-shaw.pdf
Spearman W (2018) “Beyond Expected Goals”. In: MIT Sloan Sports Analytics Conference, pp. 1–17
Steiner S et al (2019) Outplaying opponents-a differential perspective on passes using position data. German
J Exerc Sport Res. https://doi.org/10.1007/s12662-019-00579-0
Travassos B et al (2013) Performance analysis in team sports: advances from an ecological dynamics
approach. Int J Perform Anal Sport 13(1):83–95. https://doi.org/10.1080/24748668.2013.11868633
Vogelbein M, Nopp S, Hökelmann A (2014) “Defensive transition in soccer - are prompt possession regains
a measure of success? A quantitative analysis of German Fußball-Bundesliga 2010/2011”. In: Journal
of Sports Sciences 32.11, pp. 1076-1083. issn: 1466447X. https://doi.org/10.1080/02640414.2013.
879671.url: http://dx.doi.org/10.1080/02640414.2013.879671
Wang Y (2019) “A Xgboost Risk Model Via Feature Selection and Bayesian Hyper -Parameter Optimiza-
tion”. arXiv:1901.08433
Zhang W et al (2020)“Prediction of undrained shear strength using extreme gradient boosting and random
forest based on Bayesian optimization”. Geoscience Frontiers 12(1): 469–477. https://doi.org/10.
1016/j.gsf.2020.03.007

Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps
and institutional affiliations.

123
D Appendix—Study IV: Putting Team Formations
in Association Football into Context

162
Noname manuscript No.
(will be inserted by the editor)

Putting Team Formations in Association Football into Context

ID ID
Gabriel Anzer1,2 , Pascal Bauer 2,3
, Laurie Shaw4
1
Sportec Solutions AG, subsidiary of the Deutsche Fußball
Liga (DFL), Munich, Germany
2
Institute of Sports Science, University of Tübingen,
Tübingen, Germany
3
DFB-Akademie, Deutscher Fußball-Bund e.V. (DFB),
Frankfurt, Germany
4
Department of Statistics, Harvard University, Boston,
USA.

Received: date / Accepted: date

Abstract Choosing the right formation is one of the coach’s most important decisions in football. Teams
change formation dynamically throughout matches to achieve their immediate objective: to retain possession,
progress the ball up-field and create (or prevent) goal-scoring opportunities. In this work we identify the unique
formations used by teams in distinct phases of play in a large sample of tracking data. This we achieve in two
steps: first, we trained a convolutional neural network to decompose each game into non-overlapping segments
and classify these segments into phases with an average F1 -score of 0.76. We then measure and contextualize
unique formations used in each distinct phase of play. While conventional discussion tends to reduce team
formations over an entire match to a single three-digit code (e.g. 4-4-2; 4 defender, 4 midfielder, 2 striker), we
provide an objective representation of teams formations per phase of play. Using the most frequently occurring
phases of play, mid-block, we identify and contextualise six unique formations. A long-term analysis in the
German Bundesliga allows us to quantify the efficiency of each formation, and also to present a helpful scouting
tool to identify how well a coach’s preferred playing style is suited to a potential club.
Keywords Football, sports analytics, human-in-the-loop machine learning.

1 Introduction

The great Dutch football player Johan Cruyff famously observed that, on average, each player is in possession
of the ball for only 3 of the 90 minutes during a football match.1 He expanded on this observation by stating “...
so, the most important thing is: what do you do during those 87 minutes when you do not have the ball? That
is what determines whether you are a good player or not.” 2 The implication is that a player can significantly
influence the game through their positioning and movement on the field, even when they do not directly interact
with the ball (Brefeld et al. 2019; Fernandez et al. 2018).
The movement of players in a football match represent a high-dimensional spatio-temporal configuration.
Various approaches aimed to embed teams’ behavior in higher-level problems. Balague et al. (2013) focuses
on coordination of motion within a team by modelling a team’s movement as collective behavior in a complex
system. Indeed, synchronicity of movements is investigated in football in specific situations (Goes et al. 2020b;
Sarmento et al. 2018). Several studies described football matches more concrete as a multi-agent systems (Beetz
et al. 2006; Fujii 2021) highlighting the intelligence of interactions between the agents (players). Analysing
movement patterns in spatio-temporal data, especially the detection of repeating, collective patterns is not
only researched in invasion sports (Gudmundsson et al. 2017a), but also in traffic management, surveillance
and security or in the military and battlefield domain (Gudmundsson et al. 2017b). Key challenges in spatio-
temporal pattern detection are: (a) Using the interaction of movement for dimensional reduction (Balague et al.

P. Bauer
E-mail: pascal.bauer@dfb.de
Gabriel Anzer
E-mail: gabriel.anzer@herthabsc.de
1 Link et al. (2017) showed that it is even less with large differences between playing positions: central forwards (0:49 ± 0:43

min), central defenders (1:38 ± 1:09 min), central midfielders (1:27 ± 1:08 min) and, surprisingly, the longest for goalkeepers (1:38
± 0:58 min).
2 https://wheecorea.com/johan-cruyff-football-my-philosophy/25-johan-cruyff-quotes/, accessed 02/07/2021
2 G. Anzer, P. Bauer, L. Shaw

2013), (b) finding appropriate similarity metrics for related, but never identical trajectories of multiple entities
(Vilar et al. 2013), and (c) project multi-agents in a permutation-invariant space (Yeh et al. 2019).
The literature differentiates between tactics (decisions made during a match as a consequence of the dynamic
interaction in a match) and strategy (decisions made before the match) (Gréhaigne et al. 1999). However, these
concepts are often hard to distinguish (Rein et al. 2016). Coming from a more general understanding of team
formations (Wang et al. 2015), Budak et al. (2019) highlighted the problem of optimizing the team composition
(i.e. which players should be on the pitch) before the season, before the match and during the match stage as
a relevant problem in team sports. According to this definition, several approaches presented evidence-based
strategies to optimize this composition of players (Boon et al. 2003). However, this neglects the players actual
interaction on the pitch (i.e. tactics), what is in the focus of our investigation and will further be declared as
the (playing) formation.
One potential reason for this high-level consideration is the lack of available data quantifying what happens
on the pitch. For the longest time one could not objectively measure a team’s playing formation, since the only
available data describing football matches was so-called event data. Dating as far back as 1968 when Charles
Reep started manually collecting events such as shots or passes (Reep et al. 1968), this event data, which is
still being manually collected today, describes all ball actions and the players involved (Pappalardo et al. 2019a;
Stein et al. 2017). Although event data allowed for ground-breaking discoveries in football tactics (Xu 2021;
Pantzalis et al. 2020; Decroos et al. 2019; Danisik et al. 2018; Decroos et al. 2018; Pappalardo et al. 2019b;
Cintia et al. 2015; Haaren et al. 2013), it does not include any information about the positioning of all other
players. Now, with recent developments in computer vision technologies (Thinh et al. 2019; Baysal et al. 2016;
Teoldo et al. 2009) it has become possible to capture exactly that: optical tracking systems are able to record
centimeter-accurate positions of all players at every moment of a match (hereafter referred to as positional or
tracking data). This development unlocked huge potentials for professional football (Anzer et al. 2022; Anzer
et al. 2021a; Araújo et al. 2021; Wang et al. 2020; Goes et al. 2020a; Andrienko et al. 2019; Rein et al. 2016;
Herold et al. 2019).
The first approaches in football analysed formations assuming that teams play with a fixed formation across
the whole match, describing them simply as playing with a 4-4-2 (4 defenders, 4 midfielders and 2 forwards), 5-3-
2, 4-3-3, or one of approximately ten other formations that are commonly referenced (Wilson 2009). Differences
in physical requirements for similar player-roles in different formation (e.g. a central defender in a 4-4-2 versus a
5-4-1) were analysed (Vilamitjana et al. 2021; Tierney et al. 2016; Carling 2011; Bradley et al. 2011). However,
breaking a team’s formation down to three digits in a complex sport like football is a gross over-simplification
(Müller-Budack et al. 2019).
Driven by the increasing availability of tracking data, analysing team formations has been a research issue
in several sports (Gudmundsson et al. 2017b). Initiated by a pioneering work in 1999 (Intille et al. 1999), unique
formations were derived at the moment a play starts using positional data in American football (Atmosukarto
et al. 2013). Hochstedler et al. (2017) build on the static formation detection in American football by classifying
the routes of chosen player during the plays. In basketball, event data has been used to investigate established
player roles (Bianchi et al. 2017). Lucey et al. (2013) published a quantitative analyses of team formations in
field-hockey using tracking data, which was transferred to football (Wei et al. 2013) and incrementally extended
Bialkowski et al. (2014a), Bialkowski et al. (2015), and Bialkowski et al. (2016). They describe formations as
a "a coarse spatial structure which the players maintain over the course of the match" and which assigns each
player at every time of the match a unique role. Bialkowski et al. (2015) further define a role as a players
position relative to the other roles. They describe a role-identification methodology for measuring formations,
iteratively refining estimates of the average spatial positions (and deviations from those positions) of ten unique
outfield roles throughout a match. Applying a clustering algorithm on tracking data for a season of a 20-team
professional league, Bialkowski et al. (2014a) identified six unique formation types: 4-4-2, 3-4-3, 4-4-1-1 and
4-1-4-1 are all visible in their results. Variations in formations between game-states (i.e. offensive, defensive)
were first explored in Bialkowski et al. (2016). Using a more supervised approach, Müller-Budack et al. (2019).
annotated twelve typical formations (split between offense and defense) and addressed the formation problem as
a classification task. Narizuka et al. (2019) derived unique formations of 45 Japanese J1 league using a Delaunay
method combined with hierarchical clustering.
Ric et al. (2021) and Shaw et al. (2019) presented a data-driven technique for measuring and classifying team
formations as a function of game-state (offensive, defensive, transition), analysing the offensive and defensive
configurations of each team separately and dynamically detecting major tactical changes during the course of a
match. Defensive and offensive formations were measured separately by aggregating together consecutive periods
of possession of the ball for each team into two-minute windows of in-play data. Splitting up formations into
different game-states, i.e. excluding fuzzy transition situations, presented a major improvement of formation
analysis, however, they stated that further sub-game-states should be considered in future work to achieve even
more granularity (Ric et al. 2021).
While these pioneering studies have provided methods for measuring team formations and demonstrated
observations of the coherent structures formed by teams as they move around the field (and validated by football
Putting Team Formations in Association Football into Context 3

experts), they do not fully account for the changing objectives of a football team as a match evolves, influencing
team formations drastically (Andrienko et al. 2019; Gudmundsson et al. 2017b; Shaw et al. 2019; Lucey et al.
2013; Bialkowski et al. 2016). Several studies pointed out, that football consist of repetitive movement patterns,
that can be recognized by experts (Sampaio et al. 2012). We define a tactical pattern as a recurring, collective
behaviour conducted by a team or a sub-group of a team in a specific situation of a match, that can be clearly
identified by experts (Rein et al. 2016; Kempe et al. 2015; Wang et al. 2015; Grunz et al. 2012). Whereas the
detection of tactical patterns has been a relevant issue in basketball (Kempe et al. 2015; Chen et al. 2014; Perse
et al. 2006), handball (Pfeiffer et al. 2015), American football (Hochstedler et al. 2017; Stracuzzi et al. 2011;
Li et al. 2010; Siddiquie et al. 2009), and Australian rules football (Alexander et al. 2019), often only patterns
conducted by subgroups of players are analysed. The complexity of a football match requires so called team
tactics in which the whole team is involved (Rein et al. 2016). Some exemplary patterns like counterattacks
(Fassmeyer et al. 2021; Hobbs et al. 2018), ball regain strategies (Vogelbein et al. 2014), i.e. counterpressing
(Bauer et al. 2021) or general offensive strategies (Decroos et al. 2018; Kempe et al. 2014; Grunz et al. 2012;
Borrie et al. 2002; Montoliu et al. 2015; Fernando et al. 2015) have been addressed in literature and classified
as sub-categories of game-states (e.g. counterattacks and counterpressing as a subgroup of transitions in Bauer
et al. (2021) and Hobbs et al. (2018)). For such well established tactical patterns, which unavoidably occur
in every match, practitioners often use the term (tactical) phases of play 3 (although no scientific definition
established) or (tactical) game-phases (Lucey et al. 2014).
The consequence of this is that the results are not observations of a single distinct formation of a team,
but a mixture (or ‘superposition’) of the different formations used in different phases of play (Shaw et al. 2019;
Müller-Budack et al. 2019). This paper resolves this problem by using a convolution neural network (CNN) to
classify a football match over time into distinct phases of play, before measuring the formations used by either
team in each distinct phase. There are therefore two parts to our approach:
(1) A phases of play detection CNN, with architecture specifically designed for the purpose, was trained using
labeled tracking data from 97 matches in the German Bundesliga based on phases of play classifications
provided by professional analysts. Our classification scheme is described in Section 3.
(2) Within each match, periods of play classified to the same phases of play (from the perspective of one team)
are then aggregated to obtain precise measurements of the formations used. This is described in Section 4.
We apply the phases of play classifier and formation measurement tools to tracking data obtained for 2, 142
matches in the German Bundesliga over seven seasons, identifying the unique formations used in each phase of
play across our sample. This combination of a phase of play detection and formation detection fully automates
the process of identifying the distinct formation configurations used by teams during a game, revealing the
specific instructions that managers gave their team. This research was conducted in close collaboration with
professional match analysts from German Bundesliga clubs and the German national teams, who have provided
human validation of our methodology and results. This project therefore combines machine learning and human
experience aiming to obtain results that are insightful, meaningful and of practical use to coaches, managers
and scouts.
As a side-product of a practical relevant process automatization for match analysis departments, we outline
two clear use-cases of our work in Sec. 5. We are the first to quantify the strengths and weaknesses of a specific
formation when pitted against another, providing the foundation for evidence-based advice for managers seeking
the most effective counter to an opponent’s strategy during specific phases of the game (Sec. 5.1). Second, we
assess the tactical preferences of individual managers, highlighting how our tools can be used to find managers
that would provide continuity to a team’s existing playing style (Sec. 5.2). Style-matching is a crucial element
of managerial recruitment, helping to prevent a large turnover of players as a manager seeks to impose a new
playing style on a new team.

2 Positional Data

The German Bundesliga collects consistent positional data on a league-wide level, making this data available
to every team. Positional data, often also referred to as tracking or movement data (Stein et al. 2017), contains
measurements of the positions of all players, referees and the ball, sampled at a frequency of 25 Hz. These data
are gathered by an optical tracking system that captures high resolution video footage from different camera
perspectives.
In this paper, we make use of positional data from seven seasons of the German Bundesliga, from 2013/2014
until 2019/2020: a total of 2,142 matches and nearly half a billion frames are acquired by Chryronhego’s
TRACAB system.4 Validating the quality of such tracking data presents somehow an ill-posed problem due to
3 An exemplary explanation of the definition can be found here: https://www.statsperform.com/resource/phases-of-play-a

n-introduction/.
4 https://chyronhego.com/wp-content/uploads/2019/01/TRACAB-PI-sheet.pdf (accessed 02/05/2021).
4 G. Anzer, P. Bauer, L. Shaw

missing ground truth positions. Even though, several studies evaluated the accuracy of the underlying data used
in this study (Redwood-Brown et al. 2012; Linke et al. 2018; Linke et al. 2020; Taberner et al. 2020), and found
an average diversion of less then 10 cm for player positioning compared to an accurate measurement system.
Pettersen et al. (2014) presents a publicly available set of positional data, which can be used for reproduction.5

3 Phases of Play Classification

3.1 Defining Phases of Play

The primary goal in football is to score more goals than the respective opponent. Consequently, the two major
objectives are scoring goals and preventing the opponent from doing so (Kempe et al. 2014). However, given
specific situations those goals are often only implicitly followed, while sub-tasks (e.g. (re)gaining possession of
the ball), are predominant in certain situations. The concept of phases of play derives from the idea that any
moment of a match can be categorized based on the immediate intentions of each team, e.g. in defense, teams
always have to balance between the two most relevant objectives of regaining the ball (preferably in a good
position to perform an attack) and purely prevent the opponent from scoring. At the simplest level, a match
can be divided into the phases of offense and defense for each team (António et al. 2014), i.e., periods in and
out of possession of the ball. At a more granular level, professional analysts involved in our project classified the
progressive stages of attacking and defense into distinct phases.6 Fig. 1 provides an example of the phases of
play classification scheme developed by German Bundesliga analysts (see Acknowledgements). In this scheme,
open-play during a match revolves between periods of offense, transition to defense, defense and transition to
offense, with set-pieces providing a separate category (which could also be broken further down into offensive
and defensive set-pieces as well as different categories like corner kicks, throw-ins, freekicks, etc.).
Offensive play is divided into two phases: build-up, where the objective is to breach the opponent’s first
defensive line, and attacking-play, where the first line of defenders has been outplayed and the main objective is
to create a goal-scoring opportunity. In defense, professional analysts differentiate between aggressive attempts
to reclaim possession near the opponent’s goal (high-block ), a default defensive stance as the opponent progresses
the ball up the field (midfield-block or mid-block ) and a very compact defensive stance near to a team’s own
goal, where the sole objective is to prevent the opponent from scoring (low-block ). These defensive phases were
also explored in Anzer et al. (2021b) and Power et al. (2017).

Fig. 1: Overview of tactical phases of play considered.

Fig. 2 shows the phases of play break-down of a two-minute sequence of play during the Nations League
match between the German men’s national team and Spain in September 2020. The central plot shows the
distance between the German team centroid (the average position of the outfield players) and their own goal
from the 36th to 38th minutes of the game. The highlighted regions indicate the phases of play classifications,
5 Other (non-scientific) open-source positional data sets can be accessed from Skillcorner (https://github.com/SkillCorner/o

pendata) or Metrica sports (https://github.com/metrica-sports/sample-data).

6 See also: https://www.statsperform.com/resource/phases-of-play-an-introduction/.
Putting Team Formations in Association Football into Context 5

from the perspective of the German team, as determined by professional German match analysts. Freeze frames
from the footage are shown at four different instants.
The passage of play starts with a Spanish goalkick. Germany confronted this situation by attempting to force
a turnover near to the Spanish goal with a high-block. Over the first 30 seconds of play, the Spanish team played
through the high-block, forcing Germany to retreat, first into a mid-block and then to a low-block to defend
their own goal. Germany regained possession after a shot saved by Manuel Neuer (Germany’s goalkeeper) and
immediately initiated a build-up phase of possession. A long pass towards Leroy Sané on the right side of the
field briefly brought Germany into the attacking-play phase. However, Spain rapidly won the ball back, after
which Germany transitioned into a defensive mid-block and then a low-block as Spain advanced again.

Fig. 2: Team behaviour per phase of play by the reference of Germany against Spain (3rd of September 2020, venue: Stuttgart,
result: 1:1). The highlighted areas (red) in the video-footage mark the current ball action.

Match analysts spend a substantial proportion of their time manually breaking down and classifying matches
into tactical phases by watching video footage. There are very few methods published in the literature that
attempt to automate this process. Those that do focus on finding a single specific transition phases, such as
counterattacking (Fassmeyer et al. 2021; Decroos et al. 2018; Hobbs et al. 2018) or counterpressing (Bauer et al.
2021), but none attempt to classifying entire games. We now describe our methodology for achieving this.

3.2 Automated detection of Phases of Play

The phases of play definitions shown in Fig. 1 were established in collaboration with professional match analysts
from Bundesliga teams (see Acknowledgements). These definitions were then adopted by professional match
analysts to annotate 97 Bundesliga matches from the 2018/2019 season. Using the expert-labelled matches as
a training set, we explored two different machine learning approaches for automated classification of phases of
play using optical tracking data.
6 G. Anzer, P. Bauer, L. Shaw

Table 1: Rules for baseline model formation detection.

Phase of Play Rule

Offensive The first 6 seconds after a team gains ball possession are classified as transition to offense. The
remaining time during a ball possession are classified as the offensive phase.
Build-up Any moment during the offensive phase, when the ball is within its own third or the mid third
of the pitch is classified as build-up.
Attacking-play Any moment during the offensive phase, when the ball is within the opponents third is classified
as attacking-play.
Defensive The first 6 seconds after a team looses ball possession are classified as transition to defense.
The remaining time during a ball possession are classified as the defensive phase.
Low-Block Any moment during the defensive phase, when the defending team’s center (of the outfield
players) is at most 20 meters from its own goal-line, is classified as low-block.
Mid-block Any moment during the defensive phase, when the defending team’s center (of the outfield
players) is between 20 meters and 60 meters from its own goal-line, is classified as mid-block.
High-block Any moment during the defensive phase, when the defending team’s center (of the outfield
players) is at further than 60 meters from its own goal-line, is classified as high-block.

Table 2: Outcome of the phases of play detection CNN.

Tactical Phase of Play Low-block Mid-block High-block Build-up Attacking-play

Labeled phases 1 h 57 min 23 h 30 min 1 h 53 min 27 h 37 min 4 h 53 min
Average duration 9.1 s 19.0 s 13.3 s 18.6 s 8.1 s
F1 -score 0.37 0.80 0.29 0.83 0.54
Baseline model F1 -score 0.18 0.75 0.26 0.76 0.39
Inter-labeller reliability (avg. F1 -score) 0.38 0.78 0.24 0.79 0.45

The first approach is a rule-based baseline model, as described in Table 1; the results of the prediction of
the rule-based approach (compared to the inter-labeller accordance) are shown in Table 2.
The second approach makes use of convolutional neural networks (CNN), which enables us to model spatio-
temporal football data in a high dimensional, permutation-invariant space (see also Dick et al. (2019), Zheng
et al. (2016), and Wang et al. (2016)), using the raw positional data as input instead of requiring a costly step
of feature engineering (as conducted in Bauer et al. (2021) to detect counterpressing as another example of a
tactical pattern). For the CNN’s the positional data is mapped to 2-D images. Further details regarding the
network architecture can be found in the Appendix A.
On a frame-by-frame level, the CNN predicts the phases of play in our test set with a weighted average F1
score of 0.76, which is basically limited by the pairwise inter-labeller reliability of 85% (weighted F1 -score 0.72)
and exceeds the accuracy of the baseline model (0.69). On further examination, we found that the mis-classified
frames mainly occurred near the start and end points of each phase of play.
Table 2 shows some basic statistics for the training data, including the F1 -score—the harmonic mean of
recall and precision (see also Goutte et al. (2005))—for each phase of play. By taking both false positives and
false negatives into consideration, the F1 -score (calculated for each class individually) presents a very stable
evaluation metric for our purpose. Mid-block and build-up are clearly the dominant phases, making up 39%
and 47% of the phases shown in Table 2. They are also the phases with the longest duration, lasting an average
of 19.0 seconds (mid-block) and 18.6 seconds (build-up). As the mid-block is the standard opponent response
to the build-up phase, it is not surprising that the average durations are similar in length. These phases also
have the highest classification accuracy for our CNN, with both having F1 -scores exceeding 0.8. The next most
regular phase is attacking-play, making up 7% of the training data. Low-block (3%) and high-block (3%) are
the least frequently occurring phases.
The trained model was applied on seven full seasons of German Bundesliga (2013/2014-2019/2020). Much
of the following analysis focuses on the two most frequent phases: build-up and mid-block.

4 Formation Detection

4.1 Phase-dependent formations

Although positional data has been used in recent literature to quantify team-formations (Shaw et al. 2019;
Müller-Budack et al. 2019; Bialkowski et al. 2016; Bialkowski et al. 2014b; Bialkowski et al. 2015; Wei et al.
2013), they aggregate player positions over the entire match ignoring tactical changes during the match. In the
following we motivate the relevance of a more granular contemplation.
Fig. 3 shows the different formations employed across each of the five phases of play for one team during a
Bundesliga match.The dots indicate the average position of each player in the formation; the ellipses provide
an estimate of how far players tend to move from their average positions (the team is playing from left to
right), visualized through their 80% confidence region. The lower three images show the formations in the three
Putting Team Formations in Association Football into Context 7

defensive phases: low-block (left), mid-block (center) and high-block (right); the top images show the formation
in the two offensive phases: build-up (left) and offense (right).
The figure clearly indicates that team formations do not only depend on which team is in possession of the
ball, it is also heavily influenced by the tactical patterns teams are applying in different situations on the pitch,
e.g. whether the team is currently building up in their own half or attacking in the last third of the pitch. Also,
in defensive phases of play, Fig 3 (lower row) shows significant differences depending on the teams defending
strategy (high-/mid-/low-block).

Fig. 3: Average player positions of a team per tactical phase of play during one match. The ellipses provide an estimate of how far
the player would tend to move from their average position during each phase of play. The considered team plays from left to right.
Player’s positions are collected by optical tracking systems at 25 Hz (positional data).

4.2 Measuring Formation in Distinct Phases of Play

A major objective of this work is to identify the distinct formations used by teams during different phases of play
during their matches. We focus specifically on the three defensive phases (high-block, mid-block and low-block)
and two offensive phases (build-up and attacking-play) shown in Fig. 1. Transitions and set-pieces are ignored:
by definition, teams do not have a clear spatial structure during transitions, while positioning during set-pieces
are extremely dependent on the position of the ball (Casal et al. 2015). Furthermore, as it takes some time for
a team to change from one formation to another—for example, they cannot instantly shift from a high-block to
a mid-block—we ignore the first three seconds of any continuous sequence of play that was classified to a single
phase of play; if the duration of the entire sequence is less than three seconds, we discard it from our sample.
In our case, the range of observations encompasses all frames classified to the same phase of play. At least 60
seconds of (aggregated) data are required to obtain a precise measure of a formation: if the total amount of
time spent by a team in any given phase does not meet this criterium, we do not measure a formation for that
phase.
Our method for measuring formations proceeds as follows. For each team, we aggregate together all the
tracking data frames classified to a particular phase during the match and use them to measure the formation
of the team in that phase. This is achieved using the methodology of Shaw et al. (2019), who introduced a
geometric approach to measuring formations, calculating the vectors between each pair of teammates at a given
instant during a match and averaging these over a range of observations (frames) to gain a clear measure of
the team formation: each player’s position is calculated relative to the position of his nearest teammate. This
process starts with the player in the centre of the team (specifically, the player with the lowest average distance
to their third-nearest neighbour), stepping from player to player until the entire team formation is mapped out.
This method is founded on the intuition that players orient themselves relative to their nearest teammates to
retain the relative positioning required by the team’s formation.
A coach may, of course, make a major tactical change during a match, changing their team’s formations
across all phases of play. To avoid mixing two different formation strategies within a match, we search for major
tactical changes in formation by looking at each player’s average position relative to their teammates over a
rolling time window. If the relative positions change for more then ten meters (based on a three minute rolling
average), we start a new set of formation observations; more details are given in appendix B. At least one major
change in formation of either team is found in 43% of matches—taking this factor into consideration presents
8 G. Anzer, P. Bauer, L. Shaw

Table 3: Included formation observations from seven years of the German Bundesliga (2013/2014 until 2018/2019)

Tactical Phase Low-block Mid-block High-block Build-up Attacking-play

Formation Observations 1,212 5,200 638 4,867 3,164

a major improvement compared to prior work. In these games there are therefore two (or more) formation
measures for each phase of play.
From the 2, 142 matches, we exclude 345 matches that did not end with 22 players on the pitch (e.g. due to
injuries or expulsions) resulting in a final sample of 1, 803 matches. The final number of formation observations
in each phase of play are given in Table 3. As discussed above, there was not always sufficient data to measure
a formation in all phases of play during a match for both teams. Therefore, there are fewer observations in the
least frequent phases, the low-block and high-block (furthermore, not all teams employ a high-block for tactical
reasons). There are observations of the mid-block, build-up and attacking-play for almost all teams in every
match in our sample (and, on occasion, more if a team made a major tactical change during the match).

4.3 Formation Classification

To study how a specific team plays over multiple matches, we must reduce the size of our formation dataset
by identifying the unique formations within each phase of play over our entire sample of matches and classi-
fying individual observations into these unique formations. The pioneering football coach, Marcelo Bielsa, has
previously claimed that there are not more than ten formations7 in common use in professional football—our
methods enable us to explore this claim directly. Classifying formations allows us to quantify the strengths and
weaknesses of a given formation when pitted against another (Section 5.1), and study the preferred formations
used by individual Bundesliga coaches (Section 5.2).
To identify unique formation types, we apply agglomerative hierarchical clustering to the formation ob-
servations within each phase of play, using the Wasserstein metric to quantify formation similarity and the
Ward metric (Ward et al. 1963) as the linkage criterion, as described in Shaw et al. (2019). The square of the
Wasserstein distance is calculated according Olkin et al. (1982):

W (µ1 , µ2 )2 = km1 m2 k2
✓ ⇣p p ⌘1/2 ◆
+ trace C1 + C2 2 C2 C1 C2 ,

whereby µi = N (mi , Ci ) are bivariate normal distributions, m is the mean and Ci is the covariance matrix. To
solve the player-assignment problem of two formations the Hungarian algorithm is used (Kuhn 1955). Hierar-
chical clustering does not automatically identify the number of unique formations. Therefore, for each phase of
play, we varied the number of clusters from 3 to 15, creating a visual representation of the aggregated formations
within each cluster before consulting with professional match analysts to determine the true number of unique
formations within each phase of play. The final number of clusters was determined during several discussions
with expert video analysts, using quantitative metrics (i.e. Silhouette values) to achieve an alignment among
the involved experts. For different number of clusters, we plotted the cluster centroid formations (focusing on
regions with good Silhouette values). For clusters of interest, we inspected the full set of detected formations
to the analysts. Based on these observations, taking the Silhouette values into consideration, we decided on the
number of clusters for each playing phase liaising with the experts. Once the final number of unique formations
per phases of play was determined, the match analysts named each formation with a typical declaration (e.g.
4-4-2).
Fig. 4 shows the unique formations identified in the most frequently observed defensive phases of play: the
midfield-block. Results for all the most-frequently observed in-possession phases of play, build-up, are provided
in Appendix C. All the formations shown were familiar to the match analysts that inspected them. Indeed,
the analyst’s input was important in distinguishing the 4-2-3-1 formation from the 4-4-2: while the two appear
similar in the figure, inspection of the individual observations that comprised each cluster indicated that the
outside midfielders in the 4-2-3-1 (top-left plot) formed part of a triplet of attacking midfielders rather than two
conventional wingers, as in the case of the 4-4-2 (top-center).
Formations #1 #4 in Fig. 4 are all variants of a player configuration that uses four defenders as a foundation
and are distinguished by differences in the structure of the midfield and attacking players. Formation #3 sacrifices
a forward for a central defensive midfielder, while formation #4 is a narrow ‘Christmas tree’ formation8 (see
7 Marco Bielsa’s explanation of those ten formations can be found https://www.youtube.com/watch?v=qXt3rKnfbz8 (accessed

12/06/2020).
8 The term Christmas tree formation—associated with a 4-3-2-1—has established in the football community (see https://thef

alse9.com/2017/08/football-tactics-beginners-christmas-tree-formation.html, accessed 12/12/2020).

Putting Team Formations in Association Football into Context 9

Fig. 4: Outcome of the clustering for mid-block including the number of observations (obs.) of our sample.

also: Janetzko et al. (2015)) with three defensive midfielders, two attacking midfielders and a lone forward. The
remaining two formations show variants of player configurations with five defensive players.

5 Practical Applications

The primary aim of this paper is to describe our methodology for automating the process of formation detection
per phase of play. In this section we highlight two practical applications of our methods that are enabled by our
approach.

5.1 Formation versus Formation

A very common question in tactical discussion is: what is the most effective way to counter a particular formation
(Wilson 2009)? This is a challenging question as it requires a large sample of formation observations as well
as a contextualised formation detection per game-phase to attempt a quantitative answer. With over 13, 081
formation observations measured over a sample of 1, 803 Bundesliga games, we have a sufficient sample size to
attempt a comparison of the relative performance of different formation options.
The most frequently observed offensive phase of play is the build-up; the most frequently observed formation
in the build-up phase is the 2-4-3-1 (2 central defenders, 4 midfielders, 3 attacking midfielders and one forward),
hereafter referred to as a ‘two-defender’ build-up. As the most frequently observed defensive phase of play is
the mid-block, we attempt to quantify the performance of different mid-block formations in our data set when
defending against a team using a two-defender build-up. Since goals are rare events in football9 and not all shots
have an equal chances to score a goal, the concept of expected goals (xG) is often used as a more granular proxy
for the offensive contribution of a team (Anzer et al. 2021a).10 XG values are only taken into consideration in
periods of the match, where no formation change (see Appendix B) was detected. For such periods, xG values
created from all phases of play were taken into consideration, since our experts claim that the formation in the
basic phases of play (mid-block and build-up) has a latent influence on almost all situations.
The top row of Fig. 5 shows the strongest and weakest mid-block options. A 4-2-3-1 concedes, on average,
1.32 (SE: ±0.03; SD: ±0.81) xG11 per match against the two-defender build-up, while the 5-2-3 (a five-defender
formation) concedes 1.59±0.06 xG per match. The unconditional scoring rate of the two-defender build-up
formation is 1.41±0.02 xG per match; the 4-2-3-1 therefore appears to significantly reduce the attacking threat
of the two-defender build-up, while the 5-2-3 is the least effective counter-formation. The difference between the
two amounts to 0.27 xG per game, or nearly nine goals over a 34-game season.
9 For the given data set of seven seasons German Bundesliga, 3.1 goals were scored in average per match.
10 The xG value of a shot denotes the a priori probability of a shot being converted to a goal, hence its value ranges from [0, 1].
The probability is estimated using both tracking and event data and applying a machine learning model, that was trained on more
than 100.000 shots. A detailed description of the xG-model used can be found in Anzer et al. (2021a).
11 Errors quoted are the standard error on the mean.
10 G. Anzer, P. Bauer, L. Shaw

An ongoing discussion in the football tactics community is whether a build-up with two or three central
defenders is more effective (Wilson 2009).12 In the lower row of Fig. 5 we repeat the exercise for the 3-1-4-2
build-up formation, which utilizes three, rather than two, players at the back. The base scoring rate of the
three-defender build-up is 1.36±0.03 xG per game, slightly below the two-defender build-up formation. This
drops to just 1.17±0.08 xG per game when facing a 4-2-3-1 mid-block formation (lower-left)—the most effective
counter-formation—and increases to 1.45±0.08 xG per game against a 4-1-4-1 (lower-right, the weakest mid-
block formation against a 3-1-4-2). The conclusion is that the three-defender build-up formation appears to
be more easily countered than the two-defender formation while showing less of an up-side benefit against
other formations. Building up with two defenders is significantly more popular amongst Bundesliga teams than
building with three defenders; our results indicate that the latter does indeed appear to be a weaker option.

Fig. 5: Effectiveness of defensive formations (blue) against two (upper) and three (lower) player build-up (red).

Of course, even with a sample-size 1,803 matches, there are several potentially confounding factors, most
notable if there is a preference for stronger (or weaker) teams to use a particular formation, although an initial
inspection showed that every mid-block formation was used by at least 21 distinct teams once or more across
the seven seasons. Future work (as described in the discussion) should investigate these confounding factors in
significantly more detail.

5.2 Scouting the Tactical Preferences of Coaches

A major task that clubs must answer when seeking to fill a managerial vacancy is to ascertain the tactical
preferences of the candidates and determine whether each represents continuity in the team’s existing tactical
style or a significant departure. While some clubs may specifically seek a completely new style of play, there
are considerable risks associated with this. Most notably, a new tactical system will require different players,
creating turnover in the playing style as the new manager implements their preferred tactical systems and sells
the players that they do not require. Our methods allow a characterization of the types of formations that
coaches prefer to use, which is often a clear indication of their overall strategic preferences.
12 An exemplary blog-article can be found here https://thefalsefullback.de/2019/12/23/the-advantages-of-the-build-up-

with-a-back-three/.
Putting Team Formations in Association Football into Context 11

Individual teams demonstrate a preference for certain formations. Fig. 6 compares the frequency with which
a selection of Bundesliga clubs, have utilized different formation options in the mid-block phase (radar-charts).
Whereas Eintracht Frankfurt tends to play in a modern 5-3-2, Bayern Munich prefers the (somewhat similar)
4-2-3-1 or 4-1-4-1 systems. Another difference is that Bayern’s formation in the build-up phase is rather tradi-
tional, utilizing two central defenders, whereas Eintracht Frankfurt more regularly builds up with three central
defenders, which aligns with their significantly preferred 5-3-2 mid-block formation.

Fig. 6: Formations used by selected German Bundesliga clubs in the mid-block phase.

This visualization shows how different teams’ preferences can be over a long period of seven seasons. These
formation-profiles may often be determined by the key players of each team, some of whom may be particularly
suited to one formation type. Bayern Munich’s success in the past few seasons has been greatly influenced
by the central axis consisting of Jérôme Boateng, Robert Lewandowski and individually strong wingers like
Frank Ribéry, Arjen Robben, Kingsley Coman or Serge Gnabry. Our match analysts agreed the formations
most frequently utilized by Bayern’s coaches over the previous seven years—a 4-2-3-1 or a 4-1-4-1—are the most
suitable formations for the players that were at the club.
Fig. 7 demonstrates the tactical preferences of four Bayern-coaches in the mid-block phase over this period.
Guardiola, Heynckes and Flick all maintained a similar strategic approach, and all three had successful tenures.
Only Niko Kovac is generally perceived to have been a failure. One reason, often referenced in the media, is that
he was unwilling to part with the 5-3-2 build-up formation—with which he experienced success at his previous
club, Eintracht Frankfurt—instead of adapting his style of play to exploit the full potential of the players at
Bayern. The appointment of Niko Kovac did not represent continuity in Bayern’s playing style.
A valuable use-case of our methods is in the search for future managers with a similar playing style (at
least in terms of formations) to the existing approach at the hiring club. Fig. 8 shows a short-list of coaches
that could be touted as potential successors of Hansi Flick—head coach at FC Bayern from 2019 until 2021. By
comparing the coaches’ formation profiles (black)13 with that of FC Bayern (red) a similarity metric (top left
in Fig. 8) can be calculated. Although Julian Nagelsmann (currently head coach at FC Bayern Munich) is often
considered to be one of the biggest German coaching talents, his preferred formations diverge significantly from
Bayern’s existing style, resulting in a similarity score of only 44%. Jürgen Klopp and Thomas Tuchel represent
intermediate fits (72% and 73%), but Ralph Hasenhüttel, currently head coach of FC Southampton, is the best
fit for FC Bayern in our managerial database, with a similarity score of 81%. Again, the choice of a coach relies
on various factors, not solely on formations played in one or two phases of play (as displayed here). However,
our approach provides evidence for one key component, which can drastically help club’s management to take
informed decisions.

13 Note that only data from the respective coaches’ time in the German Bundesliga (2013/2014-2018/2019) are used for this

analysis.
12 G. Anzer, P. Bauer, L. Shaw

Fig. 7: FC Bayern Munich coaches by their formation (black) in comparison to the overall Bayern profile (red). The data from all
coaches and FC Bayern are aggregated over the seasons 2013/2014 to 2019/2020. The trophies (Bundesliga championship, DFB-Cup
and UEFA Champions-League) that each coach earned at his time at FC Bayern are displayed.

Fig. 8: Formation similarity. Who is the best fit for FC Bayern Munich? Top left the similarity of each coach compared to FC
Bayern is displayed.

6 Discussion

The availability of accurate and league-wide tracking data has motivated several research investigations into
team formations, the basis of team-tactics in football. The main objective of this paper was to detect phases of
play as a preliminary for contextualized formation analysis. Previous work has attempted to detect only single
specific phases of play, such as counterattacking (Fassmeyer et al. 2021; Hobbs et al. 2018) or counterpressing
(Bauer et al. 2021). For the first time, we present a method for classifying games into five distinct phases of
play. While the phases of plays used in our approach are well established among football experts, their exact
definitions may vary depending on a club’s playing philosophy. The definitions we used in the labeling process
were consolidated among professional match analysts of German Bundesliga clubs. In future work, a proper
qualitative study, that formalizes and extends the framework presented in Fig. 3 should be conducted in order
to have a proper scientific baseline for further investigations on phases of play—a well established theory in
professional football. In this context, our work shows, that (a) phases of play can be defined and identified by
experts with an appropriate accordance, and (b) that these phases of play influence the collective behavior of
teams (i.e. their formations) significantly.
We used this time-domain classification to measure team formations in distinct phases of play, achieving
a spatial classification. Phases of play measurement and classification of formations represents a major step
towards decrypting the complexities of strategy in football and provide a new insight into the tactical preferences
of individual managers and coaches. While the methodology for the formation classification is mostly similar
to the one introduced in Shaw et al. (2019), a crucial difference is not only that five different phases of play
are considered separately, but also how closely subject experts were involved throughout the whole project.
Selecting the final number of clusters purely on a statistical measure, would not lead to the same results
as when taking expert-knowledge into consideration as well. This interplay between data-science and domain
experts also turned out to be beneficial for the contextualisation of the clusters, as well as for the identification
of meaningful use-cases (see also Andrienko et al. (2019), Herold et al. (2019), Goes et al. (2020a), and Rein
et al. (2016)).
The benefit of our approach to practitioners is threefold: by automatically detecting phases of play of the next
opponent over an arbitrary number of their previous games we save the match analysis departments significant
amounts of time. An objective long-term analysis enables us to assess which formations are the most effective
counter to a particular reference formation, drastically supporting a coaches decision-making process of how to
Putting Team Formations in Association Football into Context 13

approach the next opponent. Last but not least, we show a unique use-case for club decision-makers on how
to quantify candidate coaches’ tactical style and identify those that represent continuity to the current playing
style of the club.
Besides these applications, the full potential of this approach is yet to be unlocked. Future studies could
analyse the interplay of different formations more thoroughly and control for confounding factors. On one hand,
quantitative tendencies should always be evaluated by qualitative analysis, i.e. by analysing video footage of
formation-pairings of interest to generate expert-based ad- and disadvantages when playing a specific formation
(against another). On the other hand, the most critical confounding factor (the strength of a team playing a
formation) should be modelled with a rating system of teams (see e.g., Baysal et al. (2016)) and used to validate
the hypothesis presented in Section 5.1 Additionally, when evaluating a coach’s tactical fingerprint, all phases
of play as well as other factors could be taken into consideration.

Acknowledgements

This work would not have been possible without the perspective of professional match-analysts from world class
teams who helped us to define relevant features and spend much time evaluating (intermediate) results. We
would cordially like to thank Dr. Stephan Nopp and Christofer Clemens (head match-analysts of the German
mens National team), Jannis Scheibe (head match-analyst of the German U21 mens national team), Leonard
Höhn (head match-analyst of the German women national team) as well as Sebastian Geißler (former match-
analyst of Borussia Mönchengladbach). Additionally, the authors would like to thank Dr. Hendrik Weber and
Deutsche Fußball Liga (DFL) / Sportec Solutions GmbH for providing the positional and event data.

References

Alexander, Jeremy P. et al. (2019). “The influence of match phase and field position on collective team behaviour
in Australian Rules football”. In: Journal of Sports Sciences 37.15, pp. 1699–1707. issn: 1466447X. doi:
10.1080/02640414.2019.1586077. url: https://doi.org/10.1080/02640414.2019.1586077 (cit. on
p. 3).
Andrienko, Gennady et al. (2017). “Visual analysis of pressure in football”. In: Data Mining and Knowledge
Discovery 31.6, pp. 1793–1839. issn: 1573756X. doi: 10.1007/s10618-017-0513-2 (cit. on p. 18).
Andrienko, Gennady et al. (2019). “Constructing Spaces and Times for Tactical Analysis in Football”. In: IEEE
Transactions on Visualization and Computer Graphics 27.4, pp. 2280–2297. doi: 10 . 1109 / TVCG . 2019 .
2952129. url: https://ieeexplore.ieee.org/document/8894420 (cit. on pp. 2, 3, 12).
António, Doutor et al. (2014). “The emergence of team synchronization during the soccer match: understanding
the influence of the level of opposition, game phase and field zone”. In: (cit. on p. 4).
Anzer, Gabriel and Pascal Bauer (2021a). “A Goal Scoring Probability Model based on Synchronized Positional
and Event Data”. In: Frontiers in Sports and Active Learning (Special Issue: Using Artificial Intelligence to
Enhance Sport Performance) 3.0, pp. 1–18. doi: 10.3389/fspor.2021.624475 (cit. on pp. 2, 9).
— (2022). “Expected Passes—Determining the Difficulty of a Pass in Football (Soccer) Using Spatio-Temporal
Data”. In: Data Mining and Knowledge Discovery, Springer US. issn: 1573-756X. doi: 10.1007/s10618-
021-00810-3 (cit. on p. 2).
Anzer, Gabriel, Pascal Bauer, and Ulf Brefeld (2021b). “The origins of goals in the German Bundesliga”. In:
Journal of Sport Science. doi: 10.1080/02640414.2021.1943981. url: https://www.tandfonline.com/
doi/full/10.1080/02640414.2021.1943981 (cit. on p. 4).
Araújo, Duarte et al. (2021). Artificial Intelligence in Sport Performance Analysis. April. isbn: 9781000380125.
doi: 10.4324/9781003163589 (cit. on p. 2).
Atmosukarto, Indriyati et al. (2013). “Automatic recognition of offensive team formation in american football
plays”. In: IEEE Computer Society Conference on Computer Vision and Pattern Recognition Workshops,
pp. 991–998. issn: 21607508. doi: 10.1109/CVPRW.2013.144 (cit. on p. 2).
Balague, Natàlia et al. (2013). “Overview of complex systems in sport”. In: Journal of Systems Science and
Complexity 26.1, pp. 4–13. issn: 15597067. doi: 10.1007/s11424-013-2285-0 (cit. on p. 1).
Bauer, Pascal and Gabriel Anzer (2021). “Data-driven detection of counterpressing in professional football—A
supervised machine learning task based on synchronized positional and event data with expert-based feature
extraction”. In: Data Mining and Knowledge Discovery. issn: 1573-756X. doi: 10.1007/s10618-021-00763-
7. url: https://doi.org/10.1007/s10618-021-00763-7 (cit. on pp. 3, 5, 6, 12).
Baysal, Sermetcan and Pinar Duygulu (2016). “Sentioscope: A Soccer Player Tracking System Using Model
Field Particles”. In: IEEE Transactions on Circuits and Systems for Video Technology 26.7, pp. 1350–1362.
issn: 10518215. doi: 10.1109/TCSVT.2015.2455713 (cit. on pp. 2, 13).
14 G. Anzer, P. Bauer, L. Shaw

Beetz, Michael et al. (2006). “Camera-based observation of football games for analyzing multi-agent activities”.
In: Proceedings of the International Conference on Autonomous Agents 2006, pp. 42–49. doi: 10.1145/
1160633.1160638 (cit. on p. 1).
Bialkowski, Alina et al. (2014a). “Large-Scale Analysis of Soccer Matches Using Spatiotemporal Tracking
Data”. In: IEEE International Conference on Data Mining, ICDM (Proceeding) January, pp. 725–730. issn:
15504786. doi: 10.1109/ICDM.2014.133 (cit. on p. 2).
Bialkowski, Alina et al. (2014b). “"Win at Home and Draw Away": Automatic Formation Analysis Highlighting
the Differences in Home and Away Team Behaviors”. In: MIT Sloan Sports Analytics Conference June 2016.
url: http://www.sloansportsconference.com/wp- content/uploads/2014/02/2014_SSAC_Win- at-
Home-Draw-Away.pdf (cit. on p. 6).
Bialkowski, Alina et al. (2015). “Identifying team style in soccer using formations learned from spatiotemporal
tracking data”. In: IEEE International Conference on Data Mining Workshops, ICDMW January, pp. 9–14.
issn: 23759259. doi: 10.1109/ICDMW.2014.167 (cit. on pp. 2, 6).
Bialkowski, Alina et al. (2016). “Discovering team structures in soccer from spatiotemporal data”. In: IEEE
Transactions on Knowledge and Data Engineering 28.10, pp. 2596–2605. issn: 10414347. doi: 10.1109/
TKDE.2016.2581158 (cit. on pp. 2, 3, 6).
Bianchi, Federico, Tullio Facchinetti, and Paola Zuccolotto (2017). “Role revolution: Towards a new meaning
of positions in basketball”. In: Electronic Journal of Applied Statistical Analysis 10.3, pp. 712–734. issn:
20705948. doi: 10.1285/i20705948v10n3p712 (cit. on p. 2).
Boon, Bart H. and Gerard Sierksma (2003). “Team formation: Matching quality supply and quality demand”.
In: European Journal of Operational Research 148.2, pp. 277–292. issn: 03772217. doi: 10.1016/S0377-
2217(02)00684-7 (cit. on p. 2).
Borrie, Andrew, Gudberg K. Jonsson, and Magnus S. Magnusson (2002). “Temporal pattern analysis and its
applicability in sport: An explanation and exemplar data”. In: Journal of Sports Sciences 20.10, pp. 845–852.
issn: 02640414. doi: 10.1080/026404102320675675 (cit. on p. 3).
Bradley, Paul S. et al. (2011). “The effect of playing formation on high-intensity running and technical profiles in
English FA premier League soccer matches”. In: Journal of Sports Sciences 29.8, pp. 821–830. issn: 02640414.
doi: 10.1080/02640414.2011.561868 (cit. on p. 2).
Brefeld, Ulf, Jan Lasek, and Sebastian Mair (2019). “Probabilistic movement models and zones of control”. In:
Machine Learning 108.1, pp. 127–147. issn: 15730565. doi: 10.1007/s10994- 018- 5725- 1. url: https:
//doi.org/10.1007/s10994-018-5725-1 (cit. on p. 1).
Budak, Gerçek et al. (2019). “New mathematical models for team formation of sports clubs before the match”. In:
Central European Journal of Operations Research 27.1, pp. 93–109. issn: 16139178. doi: 10.1007/s10100-
017-0491-x (cit. on p. 2).
Carling, Christopher (2011). “Influence of opposition team formation on physical and skill-related performance
in a professional soccer team”. In: European Journal of Sport Science 11.3, pp. 155–164. issn: 17461391. doi:
10.1080/17461391.2010.499972 (cit. on p. 2).
Casal, Claudio A. et al. (2015). “Analysis of corner kick success in elite football”. In: International Journal of
Performance Analysis in Sport 15.2, pp. 430–451. issn: 14748185. doi: 10.1080/24748668.2015.11868805
(cit. on p. 7).
Chen, Sheng et al. (2014). “Play Type Recognition in Real-World Football Video”. In: IEEE Winter Conference
on Applications of Computer Vision, pp. 652–659. doi: 10.1109/WACV.2014.6836040. (cit. on p. 3).
Cintia, Paolo et al. (Dec. 2015). “The harsh rule of the goals: Data-driven performance indicators for football
teams”. In: Proceedings of the 2015 IEEE International Conference on Data Science and Advanced Analytics,
DSAA 2015. Institute of Electrical and Electronics Engineers Inc. isbn: 9781467382731. doi: 10.1109/DSAA.
2015.7344823 (cit. on p. 2).
Danisik, Norbert, Peter Lacko, and Michal Farkas (Oct. 2018). “Football match prediction using players at-
tributes”. In: DISA 2018 - IEEE World Symposium on Digital Intelligence for Systems and Machines, Pro-
ceedings. Institute of Electrical and Electronics Engineers Inc., pp. 201–206. isbn: 9781538651025. doi:
10.1109/DISA.2018.8490613 (cit. on p. 2).
Decroos, Tom, Jan Van Haaren, and Jesse Davis (2018). “Automatic discovery of tactics in spatio-temporal
soccer match data”. In: Proceedings of the ACM SIGKDD International Conference on Knowledge Discovery
and Data Mining, pp. 223–232. doi: 10.1145/3219819.3219832 (cit. on pp. 2, 3, 5).
Decroos, Tom et al. (2019). “Actions speak louder than goals: Valuing player actions in soccer”. In: Proceedings
of the ACM SIGKDD International Conference on Knowledge Discovery and Data Mining 1, pp. 1851–1861.
doi: 10.1145/3292500.3330758 (cit. on p. 2).
Dick, Uwe and Ulf Brefeld (2019). “Learning to Rate Player Positioning in Soccer”. In: Big Data 7.1, pp. 71–82.
issn: 2167647X. doi: 10.1089/big.2018.0054 (cit. on p. 6).
Fassmeyer, Dennis et al. (2021). “Toward Automatically Labeling Situations in Soccer”. In: Frontiers in Sports
and Active Living 3.November. doi: 10.3389/fspor.2021.725431 (cit. on pp. 3, 5, 12).
Putting Team Formations in Association Football into Context 15

Fernandez, Javier and Luke Bornn (2018). “Wide Open Spaces : A statistical technique for measuring space
creation in professional soccer”. In: MIT Sloan Sports Analytics Conference, Boston (USA), pp. 1–19 (cit. on
p. 1).
Fernando, T et al. (2015). “Discovering Methods of Scoring in Soccer Using Tracking Data”. In: KDD Workshop
on Large-Scale Sports Analytics, pp. 1–4. url: https://large- scale- sports- analytics.org/Large-
Scale-Sports-Analytics/Submissions2015_files/paperID19-Tharindu.pdf (cit. on p. 3).
Fujii, Keisuke (2021). “Data-Driven Analysis for Understanding Team Sports Behaviors”. In: Journal of Robotics
and Mechatronics 33.3, pp. 505–514. issn: 0915-3942. doi: 10.20965/jrm.2021.p0505 (cit. on p. 1).
Goes, F. R. et al. (2020a). “Unlocking the potential of big data to support tactical performance analysis in
professional soccer: A systematic review”. In: European Journal of Sport Science 0.0, pp. 1–16. issn: 15367290.
doi: 10.1080/17461391.2020.1747552. url: https://doi.org/10.1080/17461391.2020.1747552 (cit.
on pp. 2, 12).
Goes, Floris R. et al. (2020b). “The tactics of successful attacks in professional association football—large-scale
spatiotemporal alanalysis of dynamic subgroups using position tracking data”. In: Journal of Sports Sciences
39.5, pp. 523–532. doi: 10.1080/02640414.2020.1834689 (cit. on p. 1).
Goutte, Cyril and Eric Gaussier (2005). “A Probabilistic Interpretation of Precision, Recall and F-Score, with
Implication for Evaluation”. In: Lecture Notes in Computer Science. Vol. 3408. Springer Verlag, pp. 345–359.
url: https://link.springer.com/chapter/10.1007/978-3-540-31865-1_25 (cit. on p. 6).
Gréhaigne, Jean Francis, Paul Godbout, and Daniel Bouthier (1999). “The foundations of tactics and strategy
in team sports”. In: Journal of Teaching in Physical Education 18.2, pp. 159–174. issn: 02735024. doi:
10.1123/jtpe.18.2.159 (cit. on p. 2).
Grunz, Andreas, Daniel Memmert, and Jürgen Perl (2012). “Tactical pattern recognition in soccer games by
means of special self-organizing maps”. In: Human Movement Science 31.2, pp. 334–343. issn: 01679457.
doi: 10.1016/j.humov.2011.02.008. url: http://dx.doi.org/10.1016/j.humov.2011.02.008 (cit. on
p. 3).
Gudmundsson, Joachim and Michael Horton (2017a). “Spatio-temporal analysis of team sports”. In: ACM Com-
puting Surveys 50.2, pp. 1–34. issn: 15577341. doi: 10.1145/3054132 (cit. on p. 1).
Gudmundsson, Joachim, Patrick Laube, and Thomas Wolle (2017b). “Movement Patterns in Spatio-Temporal
Data”. In: Shekhar S., Xiong H., Zhou X. (eds) Encyclopedia of GIS. Springer, Cham. doi: 10.1007/978-
3-319-17885-1{\_}823 (cit. on pp. 1–3).
Haaren, Jan Van et al. (2013). Machine Learning and Data Mining for Sports Analytics. September, p. 2013.
isbn: 9783030649111. doi: 10.1007/978-3-030-64912-8 (cit. on p. 2).
Herold, M. et al. (2019). “Machine learning in men’s professional football: Current applications and future
directions for improving attacking play”. In: International Journal of Sports Science and Coaching 14.6.
issn: 2048397X. doi: 10.1177/1747954119879350 (cit. on pp. 2, 12).
Hobbs, Jennifer et al. (2018). “Quantifying the Value of Transitions in Soccer via Spatiotemporal Trajectory
Clustering”. In: MIT Sloan Sports Analytics Conference, Boston (USA), pp. 1–11 (cit. on pp. 3, 5, 12).
Hochstedler, Jeremy and Paul T Gagnon (2017). “American Football Route Identification Using Supervised
Machine Learning”. In: MIT Sloan Sports Analytics Conference, Boston (USA), pp. 1–11 (cit. on pp. 2, 3).
Intille, Stephen S. and Aaron F. Bobick (1999). “Framework for recognizing multi-agent action from visual
evidence”. In: Proceedings of the National Conference on Artificial Intelligence, pp. 518–525 (cit. on p. 2).
Janetzko, Halld’Or et al. (2015). “Feature-driven visual analytics of soccer data”. In: 2014 IEEE Conference on
Visual Analytics Science and Technology, VAST 2014 - Proceedings. isbn: 9781479962273. doi: 10.1109/
VAST.2014.7042477 (cit. on p. 9).
Kempe, Matthias, Andreas Grunz, and Daniel Memmert (2015). “Detecting tactical patterns in basketball:
Comparison of merge self-organising maps and dynamic controlled neural networks”. In: European Journal
of Sport Science 15.4, pp. 249–255. issn: 15367290. doi: 10.1080/17461391.2014.933882. url: http:
//dx.doi.org/10.1080/17461391.2014.933882 (cit. on p. 3).
Kempe, Matthias et al. (2014). “Possession vs. Direct Play: Evaluating Tactical Behavior in Elite Soccer”. In:
International Journal of Sports Science 4.6A, pp. 35–41. issn: 2169-8791. doi: 10.5923/s.sports.201401.
05 (cit. on pp. 3, 4).
Kuhn, H.W. (1955). “The Hungarian method for the assignment problem”. In: Naval Research Logistics 2,
pp. 83–97. doi: 10.1002/nav.3800020109 (cit. on p. 8).
Li, Ruonan and Rama Chellappa (2010). “Group motion segmentation using a spatio-temporal driving force
model”. In: Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recog-
nition, pp. 2038–2045. issn: 10636919. doi: 10.1109/CVPR.2010.5539880 (cit. on p. 3).
Link, Daniel and Martin Hoernig (2017). “Individual ball possession in soccer”. In: PLoS ONE 12.7, pp. 1–15.
issn: 19326203. doi: 10.1371/journal.pone.0179953 (cit. on p. 1).
Linke, Daniel, Daniel Link, and Martin Lames (2018). “Validation of electronic performance and tracking systems
EPTS under field conditions”. In: PLoS ONE 13.7, pp. 1–20. issn: 19326203. doi: 10.1371/journal.pone.
0199519 (cit. on p. 4).
16 G. Anzer, P. Bauer, L. Shaw

Linke, Daniel, Daniel Link, and Martin Lames (2020). “Football-specific validity of TRACAB’s optical video
tracking systems”. In: PLoS ONE 15.3, pp. 1–17. issn: 19326203. doi: 10.1371/journal.pone.0230179
(cit. on p. 4).
Lucey, Patrick et al. (2013). “Representing and discovering adversarial team behaviors using player roles”.
In: Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition,
pp. 2706–2713. issn: 10636919. doi: 10.1109/CVPR.2013.349 (cit. on pp. 2, 3).
Lucey, Patrick et al. (2014). “"Quality vs Quantity": Improved Shot Prediction in Soccer using Strategic Features
from Spatiotemporal Data”. In: Proc. 8th Annual MIT Sloan Sports Analytics Conference, pp. 1–9. url:
http://www.sloansportsconference.com/?p=15790 (cit. on p. 3).
Montoliu, Raúl et al. (2015). “Team activity recognition in Association Football using a Bag-of-Words-based
method”. In: Human Movement Science 41, pp. 165–178. issn: 18727646. doi: 10.1016/j.humov.2015.03.
007 (cit. on p. 3).
Müller-Budack, Eric et al. (2019). “"Does 4-4-2 exist?" – An analytics approach to understand and classify
football team formations in single match situations”. In: Proceedings Proceedings of the 2nd International
Workshop on Multimedia Content Analysis in Sports (Nice, France) MMSports ’.September, pp. 25–33. doi:
10.1145/3347318.3355527 (cit. on pp. 2, 3, 6).
Narizuka, Takuma and Yoshihiro Yamazaki (2019). “Clustering algorithm for formations in football games”.
In: Scientific Reports 9.1, pp. 1–8. issn: 20452322. doi: 10 . 1038 / s41598 - 019 - 48623 - 1. url: http :
//dx.doi.org/10.1038/s41598-019-48623-1 (cit. on p. 2).
Olkin, I. and F. Pukelsheim (1982). “The distance between two random vectors with given dispersion matrices”.
In: Linear Algebra and Its Applications 48.C, pp. 257–263. issn: 00243795. doi: 10.1016/0024-3795(82)
90112-4 (cit. on p. 8).
Pantzalis, Victor Chazan and Christos Tjortjis (July 2020). “Sports Analytics for Football League Table and
Player Performance Prediction”. In: 11th International Conference on Information, Intelligence, Systems and
Applications, IISA 2020. Institute of Electrical and Electronics Engineers Inc. isbn: 9780738123462. doi:
10.1109/IISA50023.2020.9284352 (cit. on p. 2).
Pappalardo, Luca et al. (2019a). “A public data set of spatio-temporal match events in soccer competitions”.
In: Scientific Data 6.1, pp. 1–15. issn: 20524463. doi: 10.1038/s41597- 019- 0247- 7. url: http://dx.
doi.org/10.1038/s41597-019-0247-7 (cit. on p. 2).
Pappalardo, Luca et al. (2019b). “PlayeRank: Data-driven performance evaluation and player ranking in soccer
via a machine learning approach”. In: ACM Transactions on Intelligent Systems and Technology 10.5. issn:
21576912. doi: 10.1145/3343172 (cit. on p. 2).
Perse, Matej et al. (2006). “A Template-Based Multi-Player Action Recognition of the Basketball Game”. In:
CVBASE ’06 - Proceedings of ECCV Workshop on Computer Vision, pp. 71–82 (cit. on p. 3).
Pettersen, Svein Arne et al. (2014). “Soccer video and player position dataset”. In: Proceedings of the 5th ACM
Multimedia Systems Conference, MMSys 2014 (Singapore, March 2014), pp. 18–23. doi: 10.1145/2557642.
2563677 (cit. on p. 4).
Pfeiffer, Mark and Jürgen Perl (2015). “Analysis of tactical defensive behavior in team handball by means of
artificial neural networks”. In: IFAC-PapersOnLine 28.1, pp. 784–785. issn: 24058963. doi: 10.1016/j.
ifacol.2015.05.169 (cit. on p. 3).
Power, Paul et al. (2017). “"Not all passes are created equal:" Objectively measuring the risk and reward of
passes in soccer from tracking data”. In: Proceedings of the ACM SIGKDD International Conference on
Knowledge Discovery and Data Mining Part F1296, pp. 1605–1613. doi: 10.1145/3097983.3098051 (cit.
on p. 4).
Redwood-Brown, A., W. Cranton, and C. Sunderland (2012). “Validation of a real-time video analysis system
for soccer”. In: International Journal of Sports Medicine 33.8, pp. 635–640. issn: 01724622. doi: 10.1055/s-
0032-1306326 (cit. on p. 4).
Reep, C. and B. Benjamin (1968). “Skill and Chance in Association Football Author”. In: Journal of the Royal
Statistical Society 131.4, pp. 581–585. issn: 14698005. url: https://www.jstor.org/stable/2343726?
seq=1 (cit. on p. 2).
Rein, Robert and Daniel Memmert (2016). “Big data and tactical analysis in elite soccer: future challenges and
opportunities for sports science”. In: SpringerPlus 5.1. issn: 21931801. doi: 10.1186/s40064-016-3108-2
(cit. on pp. 2, 3, 12).
Ric, Angel et al. (2021). “Football Analytics 2021: The role of context in transferring analytics to the pitch”.
In: p. 158 (cit. on p. 2).
Sampaio, J. and V. MaçÃs (2012). “Measuring tactical behaviour in football”. In: International Journal of Sports
Medicine 33.5, pp. 395–401. issn: 01724622. doi: 10.1055/s-0031-1301320 (cit. on p. 3).
Sarmento, Hugo et al. (2018). “What Performance Analysts Need to Know About Research Trends in Association
Football (2012–2016): A Systematic Review”. In: Sports Medicine 48.4, pp. 799–836. issn: 11792035. doi:
10.1007/s40279-017-0836-6 (cit. on p. 1).
Putting Team Formations in Association Football into Context 17

Shaw, Laurie and Mark Glickman (2019). “Dynamic analysis of team strategy in professional football”. In: Barça
sports analytics summit, pp. 1–13 (cit. on pp. 2, 3, 6–8, 12).
Siddiquie, Behjat, Yaser Yacoob, and Larry S Davis (2009). “Recognizing Plays in American Football Videos”. In:
Technical Report, University of Maryland 1.1, pp. 1–8. url: http://www.researchgate.net/publication/
228519111_Recognizing_Plays_in_American_Football_Videos (cit. on p. 3).
Stein, Manuel et al. (2017). “How to Make Sense of Team Sport Data: From Acquisition to Data Modeling and
Research Aspects”. In: Data 2.1, p. 2. issn: 2306-5729. doi: 10.3390/data2010002 (cit. on pp. 2, 3).
Stracuzzi, David J et al. (2011). “An Application of Transfer to American Football : From Observation of
Raw Video to Control in a Simulated Environment An Application of Transfer to American Football :
From Observation of Raw Video to Control in a Simulated Environment”. In: AI Magazine 32.2. doi:
10.1609/aimag.v32i2.2336 (cit. on p. 3).
Taberner, Matt et al. (2020). “Interchangeability of position tracking technologies; can we merge the data?” In:
Science and Medicine in Football 4.1, pp. 76–81. issn: 24734446. doi: 10.1080/24733938.2019.1634279.
url: https://doi.org/10.1080/24733938.2019.1634279 (cit. on p. 4).
Teoldo, Israel, Júlio Manuel, and Pablo Juan Greco (2009). “Tactical Principles of Soccer Game: concepts and
application”. In: Motriz. Journal of Physical Education. UNESP 15.3, pp. 657–668. issn: 1980-6574. doi:
10.5016/2488 (cit. on p. 2).
Thinh, Nguyen Hong et al. (Oct. 2019). “A video-based tracking system for football player analysis using Effi-
cient Convolution Operators”. In: International Conference on Advanced Technologies for Communications.
Vol. 2019-Octob. IEEE Computer Society, pp. 149–154. isbn: 9781728123929. doi: 10.1109/ATC.2019.
8924544 (cit. on p. 2).
Tierney, Peter J. et al. (2016). “Match play demands of 11 versus 11 professional football using Global Positioning
System tracking: Variations across common playing formations”. In: Human Movement Science 49.October,
pp. 1–8. issn: 18727646. doi: 10.1016/j.humov.2016.05.007. url: http://dx.doi.org/10.1016/j.
humov.2016.05.007 (cit. on p. 2).
Vilamitjana, Javier J. et al. (2021). “High-intensity activity according to playing position with different team
formations in soccer”. In: Acta Gymnica 51.March, pp. 2–7. issn: 23364920. doi: 10.5507/AG.2021.003.
url: https://doi.org/10.5507/ag.2021.003 (cit. on p. 2).
Vilar, Luís et al. (2013). “Science of winning soccer: Emergent pattern-forming dynamics in association football”.
In: Journal of Systems Science and Complexity 26.1, pp. 73–84. issn: 15597067. doi: 10.1007/s11424-013-
2286-z (cit. on p. 2).
Vogelbein, Martin, Stephan Nopp, and Anita Hökelmann (2014). “Defensive transition in soccer - are prompt
possession regains a measure of success? A quantitative analysis of German Fußball-Bundesliga 2010/2011”.
In: Journal of Sports Sciences 32.11, pp. 1076–1083. issn: 1466447X. doi: 10.1080/02640414.2013.879671.
url: http://dx.doi.org/10.1080/02640414.2013.879671 (cit. on p. 3).
Wang, Jian and Jia Zhang (2015). “A win-win team formation problem based on the negotiation”. In: Engineering
Applications of Artificial Intelligence 44, pp. 137–152. issn: 09521976. doi: 10.1016/j.engappai.2015.
06.001. url: http://dx.doi.org/10.1016/j.engappai.2015.06.001 (cit. on pp. 2, 3).
Wang, Kuan-Chieh and Richard Zemel (2016). “Classifying NBA Offensive Plays Using Neural Networks”. In:
MIT Sloan Sports Analytics Conference, pp. 1–9 (cit. on p. 6).
Wang, Yikang, Hao Wang, and Mingyue Qiu (July 2020). “Performance Analysis of Everton Football Club
Based on Tracking Data”. In: Proceedings of 2020 IEEE International Conference on Power, Intelligent
Computing and Systems, ICPICS 2020. Institute of Electrical and Electronics Engineers Inc., pp. 49–53.
isbn: 9781728198736. doi: 10.1109/ICPICS50287.2020.9202246 (cit. on p. 2).
Ward, Tr and H Joe (1963). “Hierarchical Grouping to Optimize an Objective Function”. In: Journal of the
American Statistical Association 58.301, pp. 236–244 (cit. on p. 8).
Wei, Xinyu et al. (2013). “Large-scale analysis of formations in soccer”. In: 2013 International Conference on
Digital Image Computing: Techniques and Applications, DICTA 2013. doi: 10.1109/DICTA.2013.6691503
(cit. on pp. 2, 6).
Wilson, Jonathan (2009). Inverting the pyramid, a history of football tactics. London: Orion, p. 374. isbn:
978-1-4091-0204-5 (cit. on pp. 2, 9, 10).
Xu, Haoran (Mar. 2021). “Prediction on Bundesliga Games Based on Decision Tree Algorithm”. In: 2021
IEEE 2nd International Conference on Big Data, Artificial Intelligence and Internet of Things Engineering
(ICBAIE). IEEE, pp. 234–238. isbn: 978-1-6654-1540-8. doi: 10.1109/ICBAIE52039.2021.9389986. url:
https://ieeexplore.ieee.org/document/9389986/ (cit. on p. 2).
Yeh, Raymond A. et al. (2019). “Diverse generation for multi-agent sports games”. In: Proceedings of the IEEE
Computer Society Conference on Computer Vision and Pattern Recognition 2019-June, pp. 4605–4614. issn:
10636919. doi: 10.1109/CVPR.2019.00474 (cit. on p. 2).
Zheng, Stephan, Yisong Yue, and Patrick Lucey (2016). “Generating long-term trajectories using deep hierarchi-
cal networks”. In: Advances in Neural Information Processing Systems Nips, pp. 1551–1559. issn: 10495258
(cit. on p. 6).
18 G. Anzer, P. Bauer, L. Shaw

Appendix

A Detecting Phases of Play with a CNN

A schematic visualization of the CNN-architecture is displayed in Fig. 9. The input images are of size 105x68

Fig. 9: Schematic architecture of the CNN predicting the phases of play.

pixels—corresponding to the typical dimensions of a football pitch in meters—and consist of up to nine layers
(e.g. home-team positions, away-team positions, ball) containing information from a half-second period of the
game. To feed time-related information to the CNN, player trajectories, weighted with a linearly decreasing
function of time, were added to each image. To differentiate home team, away team and the ball, each information
is imported as a separate layer. Additional layers contain smoothed speed values, which slightly improved the
accuracy of our prediction. Finally, the CNN predicts one out of 15 possible phases of play14 for each frame,
although in this work only the phases shown in Fig. 1 in white boxes are taken into consideration. We split the
labeled data into 75% training and 25% test data. On the training data we used a Bayesian hyper-parameter
optimization and a 5-fold cross-validation. The final model has a batch size of 32 and was trained over 10 epochs.
The imbalanced dispersion of the phases of play (see Table 2) was addressed by resampling and weighted inputs
for each batch. The best performing CNN yielding the highest F1 -score on the test data consists of a base
model with three convolutional layers, one fully connected layer and one concatenation with one-dimensional
features. The additional features include for example a binary indicator whether the ball is in play or the game
is interrupted during the corresponding frame. Another feature, which is included in the positional data, is the
information which team is currently in possession of the ball. This base model is applied at 13 consecutive time
points (roughly half a second) and the outputs are combined using a 1-D convolution. It uses a drop-out of 50%
and a ReLu-activation function. To avoid noisy outcomes in the framewise prediction, the outcome is smoothed
afterwards by joining short sequences to its neighbouring sequences until each phase of play lasts at least one
second.

B Detecting Changes in Formation

As tactical changes in the team formation may occur at any point in the game, we need to identify the moment
when this may have happened. We use the following steps to approximate the moment when a change may
have occurred. Our approach is player specific; for example, if two wingers switch sides at half time, we want
to identify this as a change of formation. For simplicity we use the out of possession formations as a reference,
because they tend to be a bit more stable than while in possession. Therefore, we consider only the positional
data of a team (excluding the goalkeeper), while the ball is in play and the opposing team is in ball possession.
We define the current formation position of a player as his average centered position, i.e. his mean average
x and y coordinates relative to the team’s center (see also (Andrienko et al. 2017)), between the start of this
formation (e.g. the beginning of the match, or the latest identified formation change) and the current time, t.
His current formation position is then compared to his position during the last three minutes of eligible frames
up to time t. If the Euclidean distance between any player’s current formation position and his three-minute
rolling window position is greater than ten meters, we identify time t-minus-three minutes as the moment of
a formation change and start to compute the current team formations starting at this time. Both thresholds
were set by manually evaluating them on video footage with experts. Minor changes to these thresholds, do
not strongly affect the presented results. Substituted players are compared to the position of the players they
replaced. Using this algorithm over the past seven seasons of Bundesliga matches we identify on average 1.7
formation changes per match, which underpins the importance of this additional step to aggregating suitable
sequences in our clustering step.
14 These 15 phases of play contain further splits for the transition phases (e.g. counterattacking, counterpressing) and set-pieces.
Putting Team Formations in Association Football into Context 19

C Clustering for Build-up

Fig. 10 displays the clustering outcome of the second relevant phase—the build-up phase. As discussed in Section
5.1, a major decision that has to be made by a team is whether to build up with two central defenders (formations
#1, #2, #3, #4) or with three central defenders (formations #5 and #6).15 In Fig. 10, formation #1 displays a

Fig. 10: Outcome of the clustering for build-up including the number of observations (obs.) of our sample.

2-4-3-1 with two central defenders playing on the same line and the full-backs pushed into midfield. In formation
#4, one central midfielder clearly plays a more offensive role which allows the strikers not to participate in the
build-up and rather plays a more offensive part, which was declared as a 2-4-4 by our experts. The formations
shown in #2 (2-3-2-3) and #3 (2-1-4-3) also display similar patterns. The major difference is that the left and
right striker tend to support the wing-back moving forward in #2, whereas in formation #3 all three strikers
focus on playing in the center and leave the wings completely to the wing-backs. Formations #5 (3-4-3) and
#6 (3-1-4-2) shows what our experts expected: building up with three central defenders provides a distinct
flexibility during the build-up phase. A typical phenomenon when building up with three defenders is that the
wing-backs have to conquer the wing-territories on their own, which should lead to a superiority in the center
in both cases.

D Implementation Details

While the newly available positional data allows for novel insights, the sheer size poses a significant compu-
tational challenge for non-IT-focused organisations such as football clubs or federations. All implementations
were made in Python. We implemented the CNN (Section A) using Keras and Tensorflow and trained it on
a local GPU-Cluster. Additionally, we used sklearn to perform the training test data split. In order to enable
rapid feedback loops with match analysts, the tracking data is locally stored in Parquet files , compressing them
from 500mb to 20mb per match. This step not only saves storage in the analytics environment but also enables
us to read in an entire match in less than a second. For the computations necessary in this paper, the code is
parallelized whenever possible to speed up the analysis even further.

15 Note that for the formation versus formation contemplation in Section 5.1, the hierarchical clustering is further aggregated to

n=2, so that only three-defender versus two defender build-up is compared.

E Appendix—Study V: Individual Role Classifica-
tion for Players Defending Corners in Football
(Soccer)

182
 De Gruyter Journal YYYY; aop

Research Article

Gabriel Anzer, Pascal Bauer*, and Joshua Wyatt Smith

Individual role classification for players

defending corners in football (soccer)
Categorisation of the defensive role for each player in a corner kick using positional data

https://doi.org/10.1515/sample-YYYY-XXXX
Received Month DD, YYYY; revised Month DD, YYYY; accepted Month DD, YYYY

Abstract: Choosing the right defensive corner-strategy is a crucial task for each coach in professional football
(soccer). Although corners are repeatable and static situations, due to their low conversion rates, several
studies in literature failed to find usable insights about the efficiency of various corner strategies. Our
work aims to fill this gap. We hand-label the role of each defensive player from 213 corners in 33 matches,
where we then employ an augmentation strategy to increase the number of data points. By combining a
convolutional neural network with a long short-term memory neural network, we are able to detect the
defensive strategy of each player based on positional data. We identify which of seven well-established roles
a defensive player conducted (player-marking, zonal-marking, placed for counterattack, back-space, far-post,
near-post, and far-post). The model achieves an overall weighted accuracy of 89.3%, and in the case of
player-marking, we are able to accurately detect which offensive player the defender is marking 80.8% of
the time. The performance of the model is evaluated against a rule-based baseline model, as well as by
an inter-labeller accuracy. We show three concrete use-cases on how this approach can support a more
informed and fact-based decision making process.

Keywords: Sports analytics, Football (Soccer), Tactical performance analysis, Applied machine learning,
Positional and event data

Data science to support decision making in sport has become more predominant over the years. This
is especially the case for baseball [1, 2], American football [3], and basketball [4–6]. The application of
such methods in football (soccer) is only more recently gaining attention in literature [7–10]. Achieving a
comparable impact on roster and in-play decisions is a more sophisticated challenge due to the complexity
(i.e. low-scoring nature, invasive character) of the game. However, considering a repeating, static set-piece
scenario (e.g. corner kicks) substantially reduces the complexity. Compared to baseball or American football,
set-pieces only present a small fraction of a football match, nevertheless, their relevance is pointed out in
literature [11, 12] and addressed carefully in sport science [13–25] and in practise through the installation of
dedicated set-piece coaches by many teams.
The applied strategies for defending corner kicks in the top professional leagues are heterogeneous.1 A
pure player-marking approach (each offensive player is marked by one defender) is the most traditional

1 A video describing the most common roles can be found here: https://drive.google.com/file/d/1tqcxuV9-
WXAvoDw4CfvJZsVnHQaxNG5G/view?usp=sharing.

Gabriel Anzer, Hertha BSC Berlin, Berlin, Germany

Institute of Sports Science, University of Tübingen, Tübingen, Germany,
*Corresponding author: Pascal Bauer, DFB-Akademie, Deutscher Fußball-Bund e.V. (DFB), Frankfurt, Germany
Institute of Sports Science, University of Tübingen, Tübingen, Germany, pascal.bauer@dfb.de
Joshua Wyatt Smith, Department of Mathematics & Statistics, Concordia University, Montreal, Canada
Wyatt AI, Montreal, Canada, jwsmith@wyattai.com
2 G. Anzer et al., Individual role classification for players defending corners in football

tactic. In contrast, some teams prefer a pure zonal-marking approach (each defensive player is assigned to
a specific area close to the goal). Power et al. [26] pointed out that a hybrid strategy, combining player-
and zonal-marking is used most frequently, nowadays. Another crucial decision is how many players should
be dedicated to defend the corner, and how many should be positioned for a potential counterattack.
Traditionally, at least one player is placed close to the mid-line in case of a turnover. A controversial
strategy decision is whether to position players at the posts [21]. Proponents of post-marking will argue
that these players reduce the area of the goal that the goalkeeper has to cover by defending critical space
next to the posts. Many modern coaches argue that having an additional player in a more pro-active role
(e.g. player-/zonal-marking) prevents more shots, and therefore more goals in the long run. However, their
benefit is not as obvious compared to post-marking players clearing situations that would otherwise lead to
goals. Even if controversial practitioner discussions are ubiquitous since the existence of corners, statistical
analysis conducted so far on the respective efficiency has only touched the surface.
Recent success in the area of computer vision has enabled off-the-shelf availability of highly accurate
player and ball tracking data across all professional football leagues [27, 28]. Collection of players’ coordinates
now happens in an automated way, allowing for more advanced algorithms to gain new insights into the
data. With respect to analyzing corners, Power et al. [26] were the first to make use of this data and looked
at 12, 000 corners from the English Premier League (2016/2017 season). They trained a (supervised) neural
network to detect player- versus zonal-marking on a team level using the players’ positions and hand-collected
labels. Their finding that 80% of teams defend corners using a hybrid approach means that analyzing
which players are assigned to specific attacking players could be explored with more granularity. Thus, to
determine the role of each defensive player, Shaw et al. [29]2 hand-labelled 500 corners and trained an
extreme gradient boosting model based on 10 hand-crafted features. However, in the case of player-marking,
still missing was the ability to identify which attacking player was being marked. In basketball, the problem
of who is marking whom was solved for open play situations using hidden Markov models [31]. However,
the problem of detecting individual roles during corners in football is more complex. This is due to the fact
that there are roughly double the number of players (if goalkeepers are excluded), and that players can
have different roles other than just player-marking.
The objective of this paper is to accurately classify the role of each defensive player. We detect whether
a defender is player-marking (PM), zonal-marking (ZM), defending a short-corner (SD), marking the
near-post (NP) or far-post (FP), defending the back-space (BS), or being positioned for a counterattack
(CA). For those defenders that are classified as PM, we also detect which opposing player is marked. The
automated and accurate detection of defensive roles provides many practical use-cases to match analysts
and coaching staffs, e.g. for opponent analysis or player scouting.
The organisation of this paper is as follows: Section 1 describes the data used to train the algorithm and
the definitions used for the above classes. Section 2 details the process in-which a baseline model was created
using domain specific rules (Section 2.1), as well as the augmentation of the original dataset (Section 2.2)
increasing the number of training samples. The neural-network algorithm is described in detail in Section
2.3 (Model) and Section 2.4 (Training). Section 2.5 describes the results of the classification process and
compares a test dataset to hand-labelled data, consolidated between multiple domain experts. Section 3
discusses the diverse range of applications an algorithm such as this has, and Section 4 concludes with a
discussion as well as what we foresee as possible future studies.

2 Note that a first version of the approach was also presented at the 7th International Workshop on Machine Learning
and Data Mining for Sports Analytics [30].
 G. Anzer et al., Individual role classification for players defending corners in football 3

1 Data and Definitions

1.1 Data

This study makes use of data from professional German teams provided by the Sportec Solutions AG.3 In
total, we make use of 33 matches containing 213 corner kicks (after excluding corners that were played
short, or the ones containing data anomalies).
Positional data for each match is collected using optical tracking techniques with footage from cameras
installed in each stadium. The Chyronhego4 TRACAB system is used. All players and the ball positions
relative to the pitch boundaries are provided, as well as metrics such as speed and acceleration sampled at
25 Hz. Several studies have evaluated the quality of positional data [27, 32, 33], especially the TRACAB
system [28]. In order to locate corner kicks, we make use of so called event data, acquired by human
operators tagging a predefined set of events manually. Since event data provides us a rough estimate of
the time-frame the corner was played, we make use of positional data to get the exact frame following the
method of Anzer et al. [34, 35]. The event data is also used to filter out “short-played” corner kicks, since
we are only interested in corner crosses.
Tracking data of the players is taken from 0.5 s before the actual kick up until 0.8 s after the kick. This
reduces noise in the players’ movements and provides a more accurate representation of their trajectories
during the critical time-window of our investigation. The selected time-window was determined by football
experts observing video footage. Each corner is normalised to be taken from the bottom-right side of the
pitch.
Figure 1 shows an example of the normalised tracking data for three separate corner kicks and the
respective trajectories of the players and the ball. In each scenario, players of the defending team can be
seen to deploy various strategies to defend their goal, either marking specific players, or specific zones. The
definitions are discussed in the next section.

Fig. 1: Three separate corner kick scenarios showing the trajectories of all players and the ball 0.5 s before and 0.8 s after
the actual kick. The attacking team plays from left to right and is shown in red. The ball path is displayed in black.

3 https://www.sportec-solutions.de/, accessed 07/22/2021.

4 https://chyronhego.com/, accessed 07/04/2021.
4 G. Anzer et al., Individual role classification for players defending corners in football

1.2 Definitions for defender roles

Before professional football matches the coaching staff typically provides each player with a defensive
assignment covering almost all scenarios (substitutions, different scores, expulsions etc.). However, this
information is only available to the respective team’s coaching staff. Furthermore, players may adapt during
corner kicks, or just not follow the original instructions. Therefore our focus is to detect the strategy for all
defensive players during a corner kick, specifically until the ball passes the 16 m vertical boundary of the
box.
It is important to note that strategies can be (and often are) changed dynamically depending on
the progression of the ball, either from the original kick, or from interactions with players. Players also
start approximating where the ball will land and consequently change their strategy. Furthermore, not
all players position themselves for the assigned role at the same time. Some start marking their zone or
opponent immediately after the ball went out of bounds for the corner, other players can take a while to
find their position. However, everything begins with initial role assignment, which defines how the corner
kick progresses. The possible assignments for the defending team are defined as:
– Player-marking (PM): A player’s objective during the corner is to prevent a specific opposing player
from getting the ball. PM defenders position themselves close to their assigned attacker and follow
their movement into the target area. In some cases, the defender may initially position themselves so
that they intercept the intended run of the attacker. As the corner progresses, the perfect positioning is
usually as close as possible to the attacker, on the side of their own goal.
– Zonal-marking (ZM): A player’s objective is to defend (and usually clear) the ball from a specific
area. ZM players assume static positions close to the goal. They can attempt to intercept a ball by
running towards the crossed ball in the direction of play.
– Near-/Far-post (NP/FP): A player is post-marking if they are positioned directly next to the post
with the primary aim to block a shot resulting from a corner. A post-marking player can either defend
the NP or the FP, relative to which post is closer to the side of the corner kick. Player’s are only
considered NP/FP if they hold their position until the ball is cleared.
– Short-defender (SD): A player is classified as a SD if they are positioned in such a way as to either
prevent a short pass from being played, or to attack the respective players in case the corner is played
short. Players that start in a ZM position and step out towards the corner only if the ball is played
short are considered as ZM, since this is their primary intention.
– Back-space (BS): A player is positioned close to the horizontal 16 m line (i.e. outside the usual target
area of corners). It is their primary objective to get either the so called ”second ball” after an initial
header, or to defend longer distance shots from the back-space.
– Counterattack (CA): A player typically aims to be a pass option in the case of turnover-situations
and/or to receive long-distance clearances resulting from the corner. They are not defensive players in a
corner kick.

Not all roles are mandatory during corners, and any combination of the above labels can occur. To ensure
a consistent understanding (an “inter-labeller accordance”), we consolidated definitions for each strategy
among the labellers.

2 Methods
2.1 Rule-based baseline model and rule-supported labelling

Domain experts (professional football match analysts and coaches; see Acknowledgements) use the above
definitions to accurately label individual players’ roles by observing the video footage. labelling a single
 G. Anzer et al., Individual role classification for players defending corners in football 5

corner means assigning 10 roles5 for a time-window of 1.3 seconds, which can be technically demanding.
Thus, we make use of rule-based functions to create an initial dataset of ”weakly” labelled data. These rules
range from purely geometric “cuts” (i.e. if a player is standing more than away 25 m from the goal-line
they are a CA player), to geometric properties of the player in relation to the opposition and/or goal.
Table 1 summarises the domain rules. At times rules may conflict with each other (i.e., a player that is
“post-marking” according to rule 1 could also by labelled as “player-marking” qua rule 2). Thus, a player is
assigned a final label via a majority vote based on the different rule predictions. It is therefore beneficial to
have as many rules as possible for non-geometric classes. Geometric rules are prioritised. For example, if a
player is more than 25 m away from the goal, they are a CA player, regardless of other classifications. Both
majority vote and the prioritisation of geometric rules optimised the balanced accuracy on the hand-labelled
data.
This initial dataset serves two valuable purposes: First, we use the rule-based predictions as a baseline
model to evaluate the outcome of the CNN/LSTM approach. Second, it allows expert labellers to more
quickly label data and by simply correcting labels they deem incorrect. A typical setup comparing the 2-D
visualisation, video footage, and labels is shown in Figure 2. This setup drastically decreased labelling time,
compared to one without any precomputed recommendations.
In order to avoid biases during our custom setup labelling process, for the purpose of this paper, two
experts independently created the initial dataset. One expert labelled the data without the use of the
framework, thus not seeing (or being “manipulated”) by the rule-based labels, while the other one did.
The labellers agree on 94.9% of all data-points. Each labeller had the chance to annotate players of an
unclear assignment with “abstain”. Excluding all scenes which were marked as abstain by at least one of the
labellers yields an accordance of 98.1%. The labeller without rule-based support needed on average about
30% longer per corner.

Fig. 2: labelling configuration showing the radar visualisation, video footage (tactical camera), and output from the custom
framework.

The inter-labeller reliability study showed that the definitions are reliable and the distinct strategies can
be identified. Especially relevant for our use-cases is the detection of the player-marking strategy and the
respective assignment. Both can be accurately detected by domain experts (PM-role: 94.3%; PM-assignment:
87.1%6 ; see Section 2.5). The discussions showed that ZM is often mixed up with other static roles like BS
or short defender. In all considered cases, PM was only confounded with ZM. For some ”static” strategies

5 Note that experts were not interested in strategies conducted by goalkeepers for this study.
6 Calculated as the accuracy of obtaining PM (0.943) times the accuracy of obtaining the correct assignment (0.924).
6 G. Anzer et al., Individual role classification for players defending corners in football

Tab. 1: Domain rules applied to raw positional data.

Nr. Type Rule

Distance to post: A player is labelled as a NP or FP defender if they stay

1 Post-marking† within a radius of 3 m from a post for more than half the considered time
of the corner kick.

Distance to corner: A player positioned at the corner of the pitch (less

2 Short-defender†
then 20 m away from side line and less then 20 m away from goal line).

Distance to goal: A defensive player who is more than 25 m away from

3 Counterattack †
the goal in the x-coordinate direction.

Backspace corridor: A player who is within a corridor of 10 m above and

4 Back-space † below the centre line of the pitch (y-direction), as well as between 15 m
and 25 m from the goal line (x-direction).

5 Zonal-marking Average speed: The average speed of the defender is below 5 km/h.

Distance covered: The distance covered during the considered time is

6 Zonal-marking
less than 4 m.

Horizontal (Vertical) movement difference: The difference between the

7 Player-marking summed x/(y)-coordinate values for the defender and the closest attacking
player are  10 m.

Close to teammate: A defending player is within 3.5 m of their own

8 Player-marking
teammate. The closest attacking player is designated as the marked player.

Opposition tracking: The closest opposing player to a defender (within

9 Player-marking a 4 m radius) is the same at the beginning and end of the time window
considered.

Bezier curve comparison: The defender’s bezier curve shares a very small
10 Player-marking
difference (determined by inspection) with the attacking player’s curve.

Placement relative to goal: The defender has a closer mean distance to

11 Player-marking goal centre and a smaller mean angle to the line between attacker and
goal centre. A cutoff is determined via inspection.

Distance to opposition: If a defender’s mean distance to the closest

opposing player is below a threshold, assume player-marking, otherwise
assume zonal-marking. The threshold is determined via an exponential fit
between 8
two data points that satisfy the condition
12 Player/Zonal-marking
>
< 1.8 m, if x = goal line
y(x) = where the boundary conditions
>
: 7 m, at half-way line,
were determined by inspection.

Naive x-position: If the defender starts within 6 m from the goal line
14 Player/Zonal-marking assume zonal-marking. Otherwise, retrieve the closest opposition player
and assume player-marking.
† = Geometric rules
 G. Anzer et al., Individual role classification for players defending corners in football 7

the pure positioning of a player is enough to classify the role of the player (i.e. NP, FP, CA). However,
more complex roles need contextual information from the corner. In the static cases (NP, FP and CA) the
geometric rules already allow for such an accurate classification, that these are excluded in the following
steps. An example strategy where simple geometric rules are not appropriate for an accurate detection is
defending BS. Confusion often occurs between BS, SD, and ZM defenders and can not be inferred by solely
looking at the positioning on the pitch. However, the respective role of a player is clear to experts when
considering the context of each corner (e.g. trajectory of the ball, positioning of other players).

2.2 Data augmentation

We augment our data to increase the number of overall data points to be used in the training process, and
to expose the NN to numerous new and unique scenarios that may occur. Based on how far a player is
away from the goal line d, the x-coordinate of that player is perturbed such that x œ [x ≠ f (d), x + f (d)],
where f (d) = 0.1 ◊ 1.09d . Similarly, this is done for the y-coordinate. For example, a player who is covering
the NP will be perturbed very little. However, other players will be exponentially perturbed the further
away from the goal line they are, resulting in a different datapoint for the NN. Coordinates that fall outside
of the field are set to be the field limits in the respective direction. The chosen coefficients are based on
discussing their effects in different scenarios with experts.
Figure 3 shows an example of the original data compared to a perturbed dataset for the same corner.
Differences can particularly be seen between the cluster of players at the approximate coordinates (40, 5).

Fig. 3: Original corner (left figure) versus an example of the same augmented corner (right figure).

2.3 Model
A major extension in our approach compared to Power et al. [26] and Shaw et al. [29] is to classify the
assignments of individual players in the case of PM. To enable such classification, two models with two
separate inputs are trained simultaneously and joined in a final concatenation layer. The first NN, called
the role-classification network (RCN), consists of stacked convolutional layers. The second network, called
the player-assignment network (PAN), consists of stacked LSTM layers. A final dense layer combines the
outputs of both NNs to provide a prediction such that y œ {BS, SD, ZM, PM}, while the PAN determines
8 G. Anzer et al., Individual role classification for players defending corners in football

if the defender was assigned to a given attacker (y œ [0, 1]), if the RCN’s output is PM. Consequently, a
defender is assigned as a PM of an attacker if the RCN classifies them as PM. If that happens, the highest
probable attacker from the output of the PAN is assigned as the player the defender is marking.
The input to the RCN takes the form of RGB images where a single defender’s (x, y) coordinates are
plotted against the whole attacking team in a 130 ◊ 125 grid. The three color channels reflect each team
and the ball. The PAN takes sequential frames of the (x, y) coordinates and speed of the same defending
player and a single attacking player. Thus, we construct 10 ◊ 10 data-points out of a single corner (before
data augmentation is taken into consideration). Figure 4 shows the overall architecture, while Figure 5
shows examples of inputs to the RCN.

Fig. 4: Illustration of the overall model architecture. The PAN takes as input the features of a single attacking and a single
defending player. The RCN takes RGB images of size 130 ⇥ 125 consisting of the full attacking team, a single defender, and
the ball.

Fig. 5: RCN input examples for three different corner scenarios. Each team and a ball dominates the respective RGB chan-
nel, with light to dark trails indicating direction of travel, scaled by speed. The defender is shown in blue, the attacking
team in red and the ball is displayed through the green channel.

2.4 Training

The model was trained on a NVIDIA T4 GPU. A grid search for optimum parameters was performed.
Models were trained for 75 epochs, or the loss function showed negligible improvement over 15 epochs.
Table 2 shows the final hyperparameters selected. A series of jobs were run, where Bayesian optimization
was used to select successive regions, thus narrowing down the phase space. The final model was trained
 G. Anzer et al., Individual role classification for players defending corners in football 9

Tab. 2: Final configuration for RCN and PAN parameters. Stride and pooling sizes are square.

Hyperparameter Description Final Value

1 Learning rate Step size per iteration 9 ⇥ 10 4
2 Batch size Number of training examples per iteration 800
3 RCN conv 1 (size, stride) First convolutional layer (12, 5)
4 Max pooling 1 stride Downsample feature map using max values 2
5 RCN conv 2 (size, stride) Second convolutional layer (12, 3)
6 Max pooling 2 stride Downsample feature map using max values 2
7 RCN conv 3 (size, stride) Third convolutional layer (12, 3)
8 Average pooling stride Downsample feature map using average values 2
9 RCN dense Fully connected layer 12
10 RCN dropout Random fraction of nodes dropped in RCN 0.2
11 PAN dropout Random fraction of nodes dropped in PAN 0.4
12 PAN LSTM size Number of units in LSTM 12
13 RCN+PAN dense Fully connected layer combining RCN and PAN 12

using an Adam optimiser [36], with a learning rate of 9 ◊ 10≠4 and batch size of 800.7 An exponential
learning rate decay was used after five epochs showed no improvement in the loss function. The model
finished training when no improvement was seen in the loss function. The dataset was split into training
(50%), validation (25%), and test (25%) sets, where the ratio of each class mentioned in Section 1.2 is
maintained. Metrics obtained from the validation dataset are used to determine early stopping in the
training process, as well as selection of new parameter space in the hyperparameter optimisation. During
training, class weights for each of the defending categories are calculated on a batch-by-batch basis and
applied to the categorical cross entropy loss function. According the augmentation described in section
2.2 we create in total ten augmented data-files for each corner. The separation between training and test
datasets is maintained. To reduce the imbalance of the data for the PAN (a binary classifier for PM or
“not-PM”), of the ten extra augmented data files per match, only data points that are considered PM are
used for five of those augmentations.

2.5 Results

When evaluating the model, the order of how the NN makes predictions is important. First, the data point
is predicted to be in one of the 4 classes, {BS, SD, ZM, PM}.8 If the outcome is PM, only then is the
PAN consulted and its prediction obtained. Since a wrong player-assignment of the PAN is not taken into
consideration in the case of non-player-marking, an important metric for the PAN is in fact recall, while for
the RCN it is a weighted or “balanced” accuracy. Figure 6 shows the confusion matrix using the test dataset
for the RCN. A good classification can be seen among the different classes. Table 3 summarises these results
per class. The overall balanced accuracy is 89.3%. A player is correctly classified as PM 90.3% of the time,
additionally with the correct offensive player assignment occurring 80.8% of the time.9 The accuracy for
each class using the rule-based approach from Section 2.1 is included as a baseline model for comparison.

7 Due to how a single data point is constructed (Section 2.3) and fed into the NN, the total number of data points ends
up being quite large with similar or even identical attributes (in the case of the inputs to the RCN). For this reason,
large batch sizes and regularisation methods are essential.
8 For the purpose of this paper, the classes FP, NP and CA are removed since obtaining them geometrically is just as
accurate and significantly more stable given the limited number of data points in these categories.
9 Calculated as the probability that PM is correctly predicted (0.895), multiplied by the recall of the PAN (0.903).
10 Calculated as the accuracy of obtaining PM (0.943) times the accuracy of obtaining the correct assignment (0.924).
11 Calculated as the accuracy of obtaining PM (0.726) times the accuracy of obtaining the correct assignment (0.855).
10 G. Anzer et al., Individual role classification for players defending corners in football

Fig. 6: Confusion matrix for the RCN on unseen test data. Numbers indicate the players in each class.

Tab. 3: Summary of the different classes and the corresponding number of data for each used in the training phase, as
well as the overall classification achieved by our model on a test dataset. The inter labeller accordance is calculated on two
matches.

Action Data Aug-mented Inter-labeller Rule-based Test data

points data accordance accuracy accuracy
Player-marking Class 743 6,634 94.3 % 72.6% 90.3%
,! Correct Assignment 87.1%10 62.1%11 80.8%
Zonal-marking 793 4,843 89.6% 62.7% 86.9%
Near-post 15 105 93.3% 93.3% -
Far-post 14 134 100% 100% -
Counterattack 73 493 100% 100% -
Short Defender 184 1,234 95.2% 85.7% 93.7%
Back-space 274 2,084 94.7% 70.3% 92.5%

3 Practical Application
Our algorithm is able to accurately determine individual defender roles in detail. This allows for a wide
range of applications that can further be explored. We aim to outline a few of these use cases.

Use-Case 1: Automated match-report to monitor corner performance

Teams spend vast amounts of resources to prepare their strategy during corners for their upcoming matches.
However, when analysing the performance of a team post-match, an objective quality assurance is often
neglected or simply focuses on the few (possibly random) events, where goals were actually scored. Figure
7 shows an excerpt of a more granular match-report12 that is used by the German national teams in
post-match analysis. It provides an overview of the roles detected across all corners and can help to extract

12 The plot shows the post-match analysis of the U21 match Germany against Denmark in the round of the last 16 at
the U21 European Championship 2021.
 G. Anzer et al., Individual role classification for players defending corners in football 11

insights at a fraction of the time it would take match analysts. Player #10 (green) of the German team
for example, was marked by #20 (red) of the Danish team during two corners. Interestingly, he scored a
goal during the one corner where he was marked by #15 of the Danish team. From a defensive perspective,
an insightful performance metric for ZM is how often they were able to reach the ball first (“first touch”).
Figure 7 shows that number #13 and #15 (red), detected as zonal defenders (ZM), touched the ball first in
three of the corners. Another insight that can easily be retrieved from this figure is that #19, #8 and #7
on the German team were not player-marked by an opponent, with #19 getting a first touch, and all three
attempting a shot on goal following a corner kick.

Fig. 7: Excerpt from the match-report of a European Championship match of Germany U21 against Denmark. The plot
gives an overview of the player assignments of all corners where Denmark was defending, as well as their outcomes. For the
defending team (jersey numbers in red) the role is indicated by the lines to the blue role-boxes. In case of player-marking,
the player assignment is displayed by the connection to the green boxes (jersey numbers of the attacking team).
12 G. Anzer et al., Individual role classification for players defending corners in football

Use-Case 2: Analysis of individual players

Only about one in every 25 corners leads to a goal. It is not necessarily the best offensive players that take
shots but rather depends on a number of factors such as the trajectory of the ball, how well a player is
marked, and many other factors. Therefore, it is hard to objectively evaluate the performance of individual
players. By automatically analysing several seasons of a player, an indication of individual quality for aerial
duels can be given. By comparing how many shots and/or goals player-marked attackers create per defender,
we can compare attackers and defenders to each other. Our approach allows us to spot players with the
most first touches (per corner) when ZM, or the best PM-defenders preventing their assigned attackers to
reach the ball or even create a shot/goal. This is a major extension compared to existing literature.
Individual analysis can also be used for opponent scouting. Match analysis departments analyse the
upcoming opponent’s corner strategies by observing such situations on a weekly basis. Our approach can
support that process and gives more insights by looking at specific situations that occur sporadically over a
season, but cannot be covered in detail by manual annotation due to time constraints (e.g. how does a team
react after red-cards, by score, or in the last minutes of the game). Figure 8 shows an overview of individual
player strategies of the German U21 national team across their matches in the European championship
in 2021. This can be used efficiently for opponent analysis in order to get a first overview of the player’s
strategies.

Fig. 8: The figure shows the defensive corner roles of the German U21 National Team across all matches during the 2021
European Championship. The values in parentheses indicates the number of corners a player participated in.

Use-Case 3: Long-term analysis of strategies and their efficiency

Due to the effort of manual detection, no detailed statistical investigations about whether some general
strategies are more effective than others have been conducted so far. While Power et al. [26] presented a
 G. Anzer et al., Individual role classification for players defending corners in football 13

first indication about which strategy was applied more efficiently in the English Premier League season
2016/2017 on a team level, our approach enables us to understand the dominant category of hybrid marking
(80% of the cases) in greater detail. Using our automated detection, we can compare the efficiency of all
presented strategies for a team and/or on the individual player level on a long term basis. Shaw et al. [29]
already pointed out this use case by raising the question “Which attacking routines are most effective against
a certain defensive set-up?”, which could be answered using a sufficient sample size. Using our model, we
can go one step further and suggest strategies on a player-level (i.e. which defender should be assigned to
which attacker to minimize the goal scoring probability) instead of purely suggesting team strategies.

4 Discussion and Future Work

Currently, defensive strategies are chosen intuitively and heuristically in professional football. Coaches try
to make informed decisions on whether to use PM, ZM or a hybrid form, and which opponents should be
player-marked by whom. These decisions are typically prepared per match, where they try to optimise
specific role-assignments based on strength and weaknesses of their own players as well as the opponent.
Manually annotating corner strategies, or their results, has a long history in sport sciences [13–25].
These hand-created annotations are limited in size, and therefore do not allow for long-term analysis. A
Bundesliga team concedes on average only about 5.3 goals from corners per season (34 matches).13 Hence,
analysing only a single season or even only a single international tournament does not yield many significant
insights into the efficiency of different corner strategies. Power et al. [37] performed a first step towards
automating some manual annotations by detecting PM versus ZM on a team level. Hybrid marking is
predominant nowadays, but team strategies vary between two and eight zonal-markers within that hybrid
category. Power et al. [37] found that a hybrid strategy is used in 80.0% of the cases (English Premier
League, 2016/2017 season), while our dataset (from matches after 2019) consists purely of corners defended
in a hybrid formation. Shaw et al. [29] presented an approach detecting roles on a player level, however,
individual player-marking assignments were not detected. For ZM, they quote 81.0% for both precision and
recall on a training dataset. We improve upon in this score, achieving an accuracy of 86.9% on an unseen
test dataset. Our work extends previous studies and solves a practical and relevant problem with sufficient
accuracy. In the following, we lay out two major possibilities to further improve our current approach using
recent developments in machine learning.

Using weak supervision to reduce the labelling effort

The combination of positional and event data provide a detailed reproduction of professional football
matches, but most tactical strategies are too complex to detect them using a purely rule-based strategy.
Nevertheless, rules can serve as a solid starting point. Elaborate supervised machine learning models (trained
on human-made labels) are well established in football analytics literature—not only for corners [26, 29], but
also for open-play strategies like counterattacks [38, 39], counterpressing [40] or patterns like overlapping
runs [41]. Many rules aiming to detect tactical patterns require thresholds (see table 1), which can be
subjective among experts. In the special case of corners, those thresholds (e.g. the distance to the goal
until a player is considered a BS defender) cannot be set once for all corners but rather depend on several
factors (e.g. whether the corner is an in-swinger or an out-swinger), or on the philosophy and instruction
of the coach (i.e. some coaches want their BS defenders closer to the goal than others). Expert-labelling,
used for supervised learning to overcome this issue, is very time-consuming and not ideal for rapidly
evolving strategies. Individual strategies can even vary in nuances depending on different team-philosophies.

13 This average is calculated based on six seasons (2014/2015 until 2019/2020) of German Bundesliga and German 2.
Bundesliga. Goals are counted whenever they occur within 18.0 seconds after the corner was executed.
14 G. Anzer et al., Individual role classification for players defending corners in football

Accordingly, approaches using semi-supervised methodologies are well suited for football specific problems.
Ratner et al. [42] designed a weakly-supervised learning approach that significantly reduces the amount
of required labels with minimal trade-off in accuracy for different application areas (e.g. natural language
processing Hoffmann et al. [43], or saliency detection Zeng et al. [44]).
In our scenario, the rules formulated in section 2.1 to define our baseline model, could be used for
such a weak-supervision approach. We tested this methodology by training our RCN/PAN on a randomly
selected 75% of the hand-labelled data, with the remaining 25% of the labels coming from a majority vote
of the rules defined in Table 1. We achieved similar results (balanced accuracy RCN: 89.6 ± 1.6%; balanced
accuracy PAN: 85.1 ± 1.3%). This means that the labelling time can be further reduced without sacrificing
accuracy.

Using graph neural networks to reduce the computing complexity

Another recent methodology that can be used to improve our work in future investigations are graph neural
networks (GNN’s) carefully discussed by Battaglia et al. [45]. Graphs model football as a multi-agent set-up,
in which 22/23 agents (typically the ball is also modelled as an agent) interact with each other. A problem
when modelling invasion sports is that no trivial ordering is given for the players. This issue with multi-agent
classification problems on a team and a player level have been overcome in sports analytics literature using
CNN’s [46] and artificial orderings [47, 48]. An advantage of CNN approaches in football is that outliers in
player trajectories are efficiently eliminated by the CNN. This is especially helpful in set-piece scenarios
where many players interact with each other in a small space and tracking systems face many occlusions. A
GNN that models players/agents as nodes and their interactions as edges, can overcome this problem very
efficiently by creating permutation-invariant embeddings [49]. Combining a graph structure with recurrent
units allows to model temporal sequences efficiently as well [50, 51].
Yeh et al. [52], Sun et al. [53], Kipf et al. [54], and Games [55] used graphs to predict players’ trajectories
in invasion sports (i.e. basketball and football), taking all player interactions into consideration. On static
data, Stöckl et al. [56] used GNN’s to predict future events in football via node-predictions. Dick et al. [57]
were the first to present edge predictions on football data using a graph recurrent neural networks (GRNN)
in order to perform classification tasks on player interactions. Given the recent development in the theory
around graph neural networks, a combination of node-predictions (replacing the RCN) and edge-predictions
(solving the task of the PAN), could drastically reduce computation power when compared to our current
approach.

5 Conclusion
We detect defending corner strategies on a player level with a high accuracy using a combined NN approach.
Data augmentation helped us to achieve a high generalisation, despite a low sample-size. Simple geometric
rules reduced significant amounts of labelling time and can be used as labelling functions in a weak-
supervision scenario. This allows us to add new data to our approach and improve the accuracy further
with limited additional labelling effort. We improve on the accuracy compared to existing the literature, as
well as include a novel extension, which is of high value for practitioners in professional football.

Acknowledgment: This work would not have been possible without the perspective of professional match
analysts from world class teams who helped us to define relevant features and spend much time evaluating
(intermediate) results. We would cordially like to thank Dr. Stephan Nopp and Christofer Clemens (head
match analysts of the German men’s National team), Jannis Scheibe (head match analyst of the German
U21 men’s national team), Leonard Höhn (head match analyst for the German women’s national team) as
well as Sebastian Geißler (former match analyst of Borussia Mönchengladbach).
 REFERENCES 15

Funding: Please insert information concerning research grant support here (institution and grant number).
Please provide for each funder the funder’s DOI according to https://doi.crossref.org/funderNames?mode=list.

Ethics and Reproducibility By informing all participating players, all tracking is compliant to the general
data protection regulation (GDPR)14 . An ethics approval for wider research program using the respective data is
authorized by the ethics committee of the Faculty of Economics and Social Sciences at the University of Tübingen. The
data are property of the DFL e.V. / DFB e.V. and cannot be shared public. However, interested researchers can request
samples of data under non-disclosure agreement constraints at the respective institutions. With the description of the
respective tracking vendors and systems, peers working in the football industry can reproduce the results by using any
kind of professional football data.

References
[1] Ramy Elitzur. “Data analytics effects in major league baseball.” In: Omega (United Kingdom) 90 (Jan. 2020),
p. 102001. issn: 03050483. doi: 10.1016/j.omega.2018.11.010 (cit. on p. 1).
[2] Tom MacLennan. “Moneyball: The Art of Winning an Unfair Game.” In: The Journal of Popular Culture 38.4 (May
2005), pp. 780–781. issn: 0022-3840. doi: 10.1111/j.0022- 3840.2005.140- 11. url: https://onlinelibrary.wiley.
com/doi/full/10.1111/j.0022- 3840.2005.140_11.xhttps://onlinelibrary.wiley.com/doi/abs/10.1111/j.0022-
3840.2005.140_11.xhttps://onlinelibrary.wiley.com/doi/10.1111/j.0022-3840.2005.140_11.x (cit. on p. 1).
[3] Jeremy Hochstedler and Paul T Gagnon. “American Football Route Identification Using Supervised Machine Learn-
ing.” In: MIT Sloan Sports Analytics Conference, Boston (USA) (2017), pp. 1–11 (cit. on p. 1).
[4] Changjia Tian et al. “Use of machine learning to automate the identification of basketball strategies using whole
team player tracking data.” In: Applied Sciences (Switzerland) 10.1 (2020), pp. 1–16. issn: 20763417. doi: 10.3390/
app10010024 (cit. on p. 1).
[5] Matthew van Bommel and Luke Bornn. “Adjusting for scorekeeper bias in NBA box scores.” In: Data Mining and
Knowledge Discovery 31.6 (2017), pp. 1622–1642. issn: 1573756X. doi: 10.1007/s10618-017-0497-y (cit. on p. 1).
[6] Daniel Cervone et al. A Multiresolution Stochastic Process Model for Predicting Basketball Possession Outcomes.
Vol. 111. 514. 2016, pp. 585–599. isbn: 8750142011. doi: 10.1080/01621459.2016.1141685 (cit. on p. 1).
[7] F. R. Goes et al. “Unlocking the potential of big data to support tactical performance analysis in professional soccer:
A systematic review.” In: European Journal of Sport Science 0.0 (2020), pp. 1–16. issn: 15367290. doi: 10.1080/
17461391.2020.1747552. url: https://doi.org/10.1080/17461391.2020.1747552 (cit. on p. 1).
[8] Gennady Andrienko et al. “Constructing Spaces and Times for Tactical Analysis in Football.” In: IEEE Transactions
on Visualization and Computer Graphics 27.4 (2019), pp. 2280–2297. doi: 10.1109/TVCG.2019.2952129. url:
https://ieeexplore.ieee.org/document/8894420 (cit. on p. 1).
[9] M. Herold et al. “Machine learning in men’s professional football: Current applications and future directions for
improving attacking play.” In: International Journal of Sports Science and Coaching 14.6 (2019). issn: 2048397X.
doi: 10.1177/1747954119879350 (cit. on p. 1).
[10] Robert Rein and Daniel Memmert. “Big data and tactical analysis in elite soccer: future challenges and opportuni-
ties for sports science.” In: SpringerPlus 5.1 (2016). issn: 21931801. doi: 10.1186/s40064- 016- 3108- 2 (cit. on
p. 1).
[11] Diego Brito Souza et al. “A new paradigm to understand success in professional football: analysis of match statistics
in LaLiga for 8 complete seasons.” In: International Journal of Performance Analysis in Sport 19.4 (2019), pp. 543–
555. issn: 14748185. doi: 10.1080/24748668.2019.1632580. url: https://doi.org/10.1080/24748668.2019.1632580
(cit. on p. 1).
[12] Gabriel Anzer, Pascal Bauer, and Ulf Brefeld. “The origins of goals in the German Bundesliga.” In: Journal of Sport
Science (2021). doi: 10.1080/02640414.2021.1943981. url: https://www.tandfonline.com/doi/full/10.1080/
02640414.2021.1943981 (cit. on p. 1).
[13] Vasilis Armatas et al. “Analysis of the set-plays in the 18th World Cup in Germany.” In: Physical training October.1
(2007) (cit. on pp. 1, 13).
[14] Robert H. Schmicker. “An application of satscan to evaluate the spatial distribution of corner kick goals in major
league soccer.” In: International Journal of Computer Science in Sport 12.2 (2013), pp. 70–79. issn: 16844769
(cit. on pp. 1, 13).

14 https://gdpr-info.eu/, accessed 07/20/21

16 REFERENCES

[15] Craig Pulling, Matthew Robins, and Thomas Rixon. “Defending corner kicks: Analysis from the English premier
league.” In: International Journal of Performance Analysis in Sport 13.1 (2013), pp. 135–148. issn: 14748185. doi:
10.1080/24748668.2013.11868637 (cit. on pp. 1, 13).
[16] Pilar Sainz de Baranda and David Lopez-Riquelme. “Analysis of corner kicks in relation to match status in the
2006 World Cup.” In: European Journal of Sport Science 12.2 (2012), pp. 121–129. issn: 17461391. doi: 10.1080/
17461391.2010.551418 (cit. on pp. 1, 13).
[17] Toni Ardá Suárez et al. “Análisis de la eficacia de los saques de esquina en la copa del mundo de fútbol 2010. Un
intento de identificación de variables explicativas.” In: Revista de Psicologia del Deporte 23.1 (2014), pp. 165–172.
issn: 1132239X (cit. on pp. 1, 13).
[18] Craig Pulling. “Long Corner Kicks in the English Premier League.” In: Kinesiology 47.2 (2015), pp. 193–201. issn:
13311441 (cit. on pp. 1, 13).
[19] Claudio A. Casal et al. “Analysis of corner kick success in elite football.” In: International Journal of Performance
Analysis in Sport 15.2 (2015), pp. 430–451. issn: 14748185. doi: 10.1080/24748668.2015.11868805 (cit. on pp. 1,
13).
[20] Ali Onur Cerrah, Barı Özer, and Ismail Bayram. “Quantitative Analysis of Goals Scored from Set Pieces: Turkey
Super League Application.” In: Turkiye Klinikleri Journal of Sports Sciences 8.2 (2016), pp. 37–45. issn: 1308-0938.
doi: 10.5336/sportsci.2016-50745 (cit. on pp. 1, 13).
[21] Craig Pulling and Jay Newton. “Defending corner kicks in the English Premier League: Near-post guard systems.”
In: International Journal of Performance Analysis in Sport 17.3 (2017), pp. 283–292. issn: 14748185. doi: 10.1080/
24748668.2017.1331577. url: http://doi.org/10.1080/24748668.2017.1331577 (cit. on pp. 1, 2, 13).
[22] Daniel Fernández-Hermógenes, Oleguer Camerino, and Antonio García De Alcaraz. “Acciones ofensivas a balón
parado en el fútbol.” In: Apunts. Educacion Fisica y Deportes 129 (2017), pp. 78–94. issn: 20140983. doi: 10.5672/
apunts.2014-0983.es.(2017/3).129.06 (cit. on pp. 1, 13).
[23] C. A. Casal et al. “Influence of match status on corner kicks tactics in elite soccer.” In: Revista Internacional de
Medicina y Ciencias de la Actividad Fisica y del Deporte 17.68 (2017), pp. 715–728. issn: 1577-0354. doi: 10 .
15366/rimcafd2017.68.009 (cit. on pp. 1, 13).
[24] Ben William Strafford et al. “Comparative analysis of the top six and bottom six teams’ corner kick strategies in
the 2015/2016 English Premier League.” In: International Journal of Performance Analysis in Sport 19.6 (2019),
pp. 904–918. issn: 14748185. doi: 10.1080/24748668.2019.1677379. url: https://doi.org/10.1080/24748668.2019.
1677379 (cit. on pp. 1, 13).
[25] Raif Zileli and Mehmet Söyler. “Analysis of corner kicks in FIFA 2018 World Cup.” In: Journal of Human Sport and
Exercise 17.1 (2020). issn: 1988-5202. doi: 10.14198/jhse.2022.171.15 (cit. on pp. 1, 13).
[26] Paul Power et al. “Mythbusting Set-Pieces in Soccer.” In: MIT Sloan Sports Analytics Conference, Boston 102.2
(2018), pp. 1–12. issn: 15730565 (cit. on pp. 2, 7, 12, 13).
[27] Daniel Linke, Daniel Link, and Martin Lames. “Validation of electronic performance and tracking systems EPTS
under field conditions.” In: PLoS ONE 13.7 (2018), pp. 1–20. issn: 19326203. doi: 10.1371/journal.pone.0199519
(cit. on pp. 2, 3).
[28] Daniel Linke, Daniel Link, and Martin Lames. “Football-specific validity of TRACAB’s optical video tracking sys-
tems.” In: PLoS ONE 15.3 (2020), pp. 1–17. issn: 19326203. doi: 10.1371/journal.pone.0230179 (cit. on pp. 2,
3).
[29] Laurie Shaw and Sudarshan Gopaladesikan. “Routine inspection: A playbook for corner kicks.” In: MIT Sloan Sports
Analytics Conference, Boston (USA) (2021) (cit. on pp. 2, 7, 13).
[30] Jan Van Haaren et al. Machine Learning and Data Mining for Sports Analytics. September. 2013, p. 2013. isbn:
9783030649111. doi: 10.1007/978-3-030-64912-8 (cit. on p. 2).
[31] Alexander Franks et al. “Characterizing the spatial structure of defensive skill in professional basketball.” In: Annals
of Applied Statistics 9.1 (2015), pp. 94–121. issn: 19417330. doi: 10.1214/14-AOAS799 (cit. on p. 2).
[32] Matt Taberner et al. “Interchangeability of position tracking technologies; can we merge the data?” In: Science
and Medicine in Football 4.1 (2020), pp. 76–81. issn: 24734446. doi: 10 . 1080 / 24733938 . 2019 . 1634279. url:
https://doi.org/10.1080/24733938.2019.1634279 (cit. on p. 3).
[33] A. Redwood-Brown, W. Cranton, and C. Sunderland. “Validation of a real-time video analysis system for soccer.” In:
International Journal of Sports Medicine 33.8 (2012), pp. 635–640. issn: 01724622. doi: 10.1055/s-0032-1306326
(cit. on p. 3).
[34] Gabriel Anzer and Pascal Bauer. “Expected Passes—Determining the Difficulty of a Pass in Football (Soccer) Using
Spatio-Temporal Data.” In: Data Mining and Knowledge Discovery, Springer US (2022). issn: 1573-756X. doi:
10.1007/s10618-021-00810-3 (cit. on p. 3).
[35] Gabriel Anzer and Pascal Bauer. “A Goal Scoring Probability Model based on Synchronized Positional and Event
Data.” In: Frontiers in Sports and Active Learning (Special Issue: Using Artificial Intelligence to Enhance Sport
Performance) 3.0 (2021), pp. 1–18. doi: 10.3389/fspor.2021.624475. url: https://www.frontiersin.org/articles/10.
3389/fspor.2021.624475/full (cit. on p. 3).
 REFERENCES 17

[36] Diederik Kingma and Jimmy Ba. “Adam: A Method for Stochastic Optimization.” In: International Conference on
Learning Representations (Dec. 2014) (cit. on p. 9).
[37] Paul Power et al. “"Not all passes are created equal:" Objectively measuring the risk and reward of passes in soccer
from tracking data.” In: Proceedings of the ACM SIGKDD International Conference on Knowledge Discovery and
Data Mining Part F1296 (2017), pp. 1605–1613. doi: 10.1145/3097983.3098051 (cit. on p. 13).
[38] Dennis Fassmeyer et al. “Toward Automatically Labeling Situations in Soccer.” In: Frontiers in Psychology (Special
Research Topic on Collective Behaviour in Team Sports; accepted for publication) (2021) (cit. on p. 13).
[39] Jennifer Hobbs et al. “Quantifying the Value of Transitions in Soccer via Spatiotemporal Trajectory Clustering.” In:
MIT Sloan Sports Analytics Conference, Boston (USA) (2018), pp. 1–11 (cit. on p. 13).
[40] Pascal Bauer and Gabriel Anzer. “Data-driven detection of counterpressing in professional football—A supervised
machine learning task based on synchronized positional and event data with expert-based feature extraction.” In:
Data Mining and Knowledge Discovery 35.5 (2021), pp. 2009–2049. issn: 1573-756X. doi: 10.1007/s10618- 021-
00763-7. url: https://link.springer.com/article/10.1007/s10618-021-00763-7 (cit. on p. 13).
[41] Gabriel Anzer et al. “Detection of tactical patterns using semi-supervised graph neural networks.” In: MIT Sloan
Sports Analytics Conference, Boston, USA (Accepted for Research paper track 2022) 16 (2022), pp. 1–3 (cit. on
p. 13).
[42] Alexander Ratner et al. “Snorkel: Rapid training data creation with weak supervision.” In: Proceedings of the VLDB
Endowment 11.3 (2017), pp. 269–282. issn: 21508097. doi: 10.14778/3157794.3157797 (cit. on p. 14).
[43] Raphael Hoffmann et al. “Knowledge-based weak supervision for information extraction of overlapping relations.” In:
ACL-HLT 2011 - Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human
Language Technologies 1.May 2014 (2011), pp. 541–550 (cit. on p. 14).
[44] Yu Zeng et al. “Multi-source weak supervision for saliency detection.” In: Proceedings of the IEEE/CVF Conference
on Computer Vision and Pattern Recognition. (2019), pp. 6074–6083. issn: 23318422 (cit. on p. 14).
[45] Peter W. Battaglia et al. Relational inductive biases, deep learning, and graph networks. 2018 (cit. on p. 14).
[46] Uwe Dick and Ulf Brefeld. “Learning to Rate Player Positioning in Soccer.” In: Big Data 7.1 (2019), pp. 71–82.
issn: 2167647X. doi: 10.1089/big.2018.0054 (cit. on p. 14).
[47] Hoang M. Le et al. “Coordinated multi-agent imitation learning.” In: 34th International Conference on Machine
Learning, ICML 2017. Vol. 4. 2017, pp. 3140–3152. isbn: 9781510855144 (cit. on p. 14).
[48] Keisuke Fujii. “Data-Driven Analysis for Understanding Team Sports Behaviors.” In: Journal of Robotics and Mecha-
tronics 33.3 (2021), pp. 505–514. issn: 0915-3942. doi: 10.20965/jrm.2021.p0505 (cit. on p. 14).
[49] Luana Ruiz, Fernando Gama, and Alejandro Ribeiro. “Gated Graph Recurrent Neural Networks.” In: IEEE Transac-
tions on Signal Processing 68 (2020), pp. 6303–6318. issn: 19410476. doi: 10.1109/TSP.2020.3033962 (cit. on
p. 14).
[50] Ehsan Hajiramezanali et al. “Variational Graph Recurrent Neural Networks.” In: 33rd Conference on Neural Informa-
tion Processing Systems (NeurIPS 2019), Vancouver, Canada. NeurIPS (2019) (cit. on p. 14).
[51] Yujia Li et al. “Gated graph sequence neural networks.” In: 4th International Conference on Learning Representa-
tions, ICLR 2016 - Conference Track Proceedings 1 (2016), pp. 1–20 (cit. on p. 14).
[52] Raymond A. Yeh et al. “Diverse generation for multi-agent sports games.” In: Proceedings of the IEEE Computer
Society Conference on Computer Vision and Pattern Recognition 2019-June (2019), pp. 4605–4614. issn: 10636919.
doi: 10.1109/CVPR.2019.00474 (cit. on p. 14).
[53] Chen Sun et al. “Stochastic prediction of multi-agent interactions from partial observations.” In: Seventh Inter-
national Conference on Learning Representations (ICLR), New Orleans (USA) (2019), pp. 1–15. issn: 23318422
(cit. on p. 14).
[54] Thomas Kipf et al. “Neural relational inference for Interacting systems.” In: 35th International Conference on Ma-
chine Learning, ICML 2018 6 (2018), pp. 4209–4225. issn: 1938-7228 (cit. on p. 14).
[55] Multi-agent Sports Games. “A Graph Attention Based Approach for Trajectory Prediction in Multi-agent Sports
Games.” In: Preprint (arXiv) (2019) (cit. on p. 14).
[56] Michael Stöckl et al. “Making Offensive Play Predictable - Using a Graph Convolutional Network to Understand
Defensive Performance in Soccer.” In: MIT Sloan Sports Analytics Conference, Boston (USA) (2021) (cit. on p. 14).
[57] Uwe Dick, Maryam Tavakol, and Ulf Brefeld. “Rating Player Actions in Soccer.” In: Frontiers in Sports and Active
Learning (Special Issue: Using Artificial Intelligence to Enhance Sport Performance) 3 (2021), p. 174. doi: 10.3389/
fspor.2021.682986. url: https://www.frontiersin.org/article/10.3389/fspor.2021.682986 (cit. on p. 14).
F Appendix—Study VI: Toward Automatically La-
beling Situations in Soccer

200
 ORIGINAL RESEARCH
published: 03 November 2021
doi: 10.3389/fspor.2021.725431

Toward Automatically Labeling

Situations in Soccer
Dennis Fassmeyer 1 , Gabriel Anzer 2,3† , Pascal Bauer 2,4† and Ulf Brefeld 1*
1
Machine Learning Group, Leuphana University of Lüneburg, Lüneburg, Germany, 2 Department of Sport Psychology and
Research Methods, Institute of Sports Science, University of Tübingen, Tübingen, Germany, 3 Sportec Solutions AG,
Subsidiary of the Deutsche Fußball Liga (DFL), Munich, Germany, 4 DFB-Akademie, Deutscher Fußball-Bund e.V. (DFB),
Frankfurt, Germany

We study the automatic annotation of situations in soccer games. At first sight, this
translates nicely into a standard supervised learning problem. However, in a fully
supervised setting, predictive accuracies are supposed to correlate positively with
the amount of labeled situations: more labeled training data simply promise better
performance. Unfortunately, non-trivially annotated situations in soccer games are
scarce, expensive and almost always require human experts; a fully supervised approach
Edited by: appears infeasible. Hence, we split the problem into two parts and learn (i) a meaningful
Rui Marcelino, feature representation using variational autoencoders on unlabeled data at large scales
University Institute of Maia, Portugal
and (ii) a large-margin classifier acting in this feature space but utilize only a few (manually)
Reviewed by:
Paizis Christos, annotated examples of the situation of interest. We propose four different architectures
Université de Bourgogne, France of the variational autoencoder and empirically study the detection of corner kicks,
Hendrik Meth,
crosses and counterattacks. We observe high predictive accuracies above 90% AUC
Stuttgart Media University, Germany
irrespectively of the task.
*Correspondence:
Ulf Brefeld Keywords: sports analytics, soccer, tracking data, variational autoencoders, labeling situations
brefeld@leuphana.de

† ORCID:

Gabriel Anzer INTRODUCTION

orcid.org/0000-0003-3129-8359
Pascal Bauer The acquisition of tracking/positional and event data has become ubiquitous in professional
orcid.org/0000-0001-8613-6635 football. The benefits of the resulting digital reproduction of a match, widely available in
professional leagues, are twofold: Firstly, coaches, analysts and other decision makers in clubs may
Specialty section: use data as an objective and quantitative alternative to traditional analyzes of performance, and,
This article was submitted to secondly, the collected data enables media to tell automated stories, to provide data-driven insights
Elite Sports and Performance in what is happening on the pitch.
Enhancement,
For example, match-analysis departments have historically spend vast amounts of time
a section of the journal
Frontiers in Sports and Active Living
analyzing their upcoming opponent before each match by manually evaluating video footage. This
work intensive approach is nowadays being supported or even partially replaced by automatic
Received: 15 June 2021
insight generation based on available data. While some information is easily accessible from the
Accepted: 06 October 2021
Published: 03 November 2021
collected data, e.g., extracting the preferred formation of a team (Shaw and Glickman, 2019),
other (rather tactical) pieces of information cannot be automatically computed yet, either because
Citation:
Fassmeyer D, Anzer G, Bauer P and
they are too complex (e.g., how teams behave during counterattacks), depend on the actual
Brefeld U (2021) Toward Automatically game philosophy of a team, require large amounts of tactical knowledge, or are considered a
Labeling Situations in Soccer. niche with only few interested followers. Detecting such events and patterns automatically offers
Front. Sports Act. Living 3:725431. a huge potential for performance analysis and may revolutionize current pre- and post-match
doi: 10.3389/fspor.2021.725431 performance analyses in professional football.

Frontiers in Sports and Active Living | www.frontiersin.org 1 November 2021 | Volume 3 | Article 725431
Fassmeyer et al. Toward Automatically Labeling Situations in Soccer

When speaking about data in soccer, we differentiate between and spatial dependencies of positional data. Existing body of
positional/tracking and event data. Positional data, describing research on extending VAEs to sequential data mainly focuses on
player and ball positions at any point in time of a match, the generative aspects of the models rather than on their potential
are collected automatically via computer vision algorithms and benefits in the context of semi-supervised learning (Chung et al.,
dedicated tracking cameras. Event data, on the other hand, 2015; Goyal et al., 2017).
provides basic annotations of game events (mainly on ball actions In this paper, we propose novel VAE-based feature extraction
like passes, shots, tackles, etc.) and is still acquired manually methods. Starting from the vanilla VAE, we begin with proposing
by human operators. The manual collection of such events is a rather straight forward generalization that can be applied
unsurprisingly labor and cost intensive and involves up to five to positional data. A second contribution incorporates existing
operators per game. The goal of this article is to bridge the auxiliary labels in the training process. The idea of the auxiliary
gap from the status quo toward fully-automatic annotations of labels is to foster discriminative causes of variation in the inferred
soccer games. latent feature representation. The main contribution however
There are several recent studies aiming to detect basic events is the development of sequential counterparts of the two VAEs
directly out of video footage (Ekin et al., 2003; Wickramaratna to match the spatiotemporal problem domain. After one of the
et al., 2005; Kolekar and Palaniappan, 2009) or positional data VAEs has been trained using unlabeled or auxiliary labeled data,
(Zheng and Kudenko, 2010; Motoi et al., 2012; Richly et al., only a few of the feature representations, for which labels of
2016; Stein et al., 2019) and others focus on the identification interest exist, are fed into a support vector machine to train
of sophisticated tactical patterns (Hobbs et al., 2018; Andrienko the final classifier. We empirically evaluate the effectiveness of
et al., 2019; Shaw and Sudarshan, 2020; Anzer et al., 2021; our approach on three different detection tasks, involving the
Bauer and Anzer, 2021). The proposed approaches provide detection of cornerkicks, crosses (labels obtained from event
useful solutions for their respective tasks. However, they are data), and counterattacks (labels manually annotated by experts).
also restricted to either a particular data source or type of We observe detection rates above 90% AUC for all tasks and
events or pattern that is to be detected; none of the above discuss several findings on methodological issues derived from
approaches offer an all-encompassing framework to deal with further experimentation.
general detection problems. The remainder is structured as follows. Section Problem
A challenge for designing a general detector of game situations Setting introduces the formal problem setting. The static and
is the available data structure. While vast amounts of positional sequential models are presented in sections Static Models,
data of players and ball exist, collecting the associated labels of Sequential Models, respectively. We report on our empirical
interest is an expensive endeavor and requires manual annotation findings in section Empirical Evaluation and provide a discussion
by human experts. For example, counterattack detection first in section Discussion. Section Related Work reviews related work
involves defining strict criteria and definitions of counterattacks and section Conclusion concludes.
before engaging in extensive search processes to annotate the
matching game snippets. Consequently, it is vital to reliably
PROBLEM SETTING
extract the game situations with little external supervision. In
that sense, classical supervised learning methods fail to be a Positional data from professional soccer is introduced as follows.
viable candidate since the algorithms typically require large Let A be the set of agents (i.e., players and ball) and T be
amounts of annotated data to achieve a good generalization the set of timesteps. For each element of the cartesian product
error (Erhan et al., 2010). However, a strategy to mitigate the A × T , whereabouts of all agents on the pitch in form of two-
necessity of a large number of labels is to incorporate abundantly dimensional coordinates (g, h) ∈ R2 are observed. It will be
available unlabeled data into the training process. While there are convenient to further divide the set of agents into three disjoint
many conceivable ways to operate within such a semi-supervised subsets, A1 , A2 , and A3 , corresponding to the players on teams
framework, we focus particularly on the variational autoencoder 1, team 2, and the ball1 , respectively.
(VAE) (Kingma and Welling, 2013; Rezende et al., 2014) family Individual spatiotemporal movements of the agents allow to
of methods. augment the positional data with additional pieces of information
Variational autoencoders learn implicit low-dimensional dg
such as the (approximated) velocity of players ( dt , dhdt ). More
feature representations for input data by jointly training a
precisely, linearized motion for agent a ∈ A is computed via
probabilistic encoder and decoder network. The idea is that
the original observations can be reconstructed (approximately)
from this lower-dimensional feature space. In fact, our semi- (!gt(a) , !h(a) (a) (a) (a) (a)
t ) = (gt # − gt , ht # − ht )
supervised strategy relies on inferring these semantically salient
representations for annotated situations, hence reducing the with t # > t and (!gt(a) , !h(a) # / T,
t ) = (0, 0) for the case of t ∈
need to solve a large supervised learning problem in feature i.e., using a small time window between two consecutive frames.
space. Our instance of semi-supervised learning achieves a Further defining Y as an auxiliary label space that consists of
substantial increase in generalization ability in cases where only inexpensive labels (e.g., provided by event data), we are given a
a few observed labels are available (Kingma et al., 2014). An subset of event annotations TY ⊂ T s.t. |TY | & |T |, referred
essential contribution of this paper is to lift the underlying
principles to spatiotemporal structures to capture the temporal 1 We have A
i ⊂ A s.t. A1 ∪ A2 ∪ A3 = A and A1 ∩ A2 ∩ A3 = ∅.

Frontiers in Sports and Active Living | www.frontiersin.org 2 November 2021 | Volume 3 | Article 725431
Fassmeyer et al. Toward Automatically Labeling Situations in Soccer

to as yS := {yt : t ∈ TY }. We further denote Yb as the (binary) 1. The training of a VAE-based feature extraction module to
target space described by an action value of interest and a “no transform the high-dimensional tensor data xt into a low-
action” value with TYb ⊂ T (|TYb | & |T |) defining the set dimensional embedding space.
yB : = {yt : t ∈ TYb }2 . We denote the composite of all pixel 2. The training of a classifier using the derived embeddings and
coordinates and velocity values of agents a at a certain timestep as the available label information.
xt := {(gt(a) , h(a) (a) (a)
t , !gt , !ht )}a∈A and formulate our objective as Irrespective of the first step’s choice, we use a support-vector
quantifying the probability over Yb given the state representation machine (SVM) (Cortes and Vapnik, 1995) for the second step.
xt for all t ∈ T . The technical contributions of this paper address the first stage
An emerging issue is to find a pertinent representation and introduce novel feature extraction methods in sections Static
of the described data for model training. A plain random Models and Sequential Models. See Figure 1 for an illustration of
concatenation of the agents’ coordinates and velocities at time t the information flow.
is clearly inappropriate in the sense that divergent instantiations
of agent orderings also translate into divergent representations
for the exact same state. Accordingly, the function that
STATIC MODELS
tranforms instances of {(gt(a) , h(a) (a) (a)
t , !gt , !ht )}a∈A into an In this section, we present static models that operate only on a
input representation of a neural network needs to be invariant single timestamp to predict a labeling of the encoded situation.
under permutation of the agents. Since the locations of the The term static stems from an equivalence class of model
agents are given as pixel coordinates, we choose to convert architectures whose resulting optimization targets are derived
these coordinates into an image-based representation, resulting based on the assumption that each tensor frame is iid., i.e., the
in a consistent representational structure across different computation factors across the individual timesteps of a game.
game settings. Note, however, that the data points themselves contain sequential
The mechanism for capturing position and motion information due to the inclusion of motion vectors for each agent.
information in a 3-dimensional image representation xt is We discard the time subscripts for the tensor representations x
based on the approach presented in Dick and Brefeld (2019). since we operate within a static domain.
Here, the pitch size (105 × 68) defines the axes in the horizontal
and vertical directions, with each channel of the tensor encoding Preliminaries
a different subset of the available information. The first 3 The idea of a variational autoencoder (VAE) (Kingma and
channels capture positional information of A1 , A2 and A3 (in Welling, 2013; Rezende et al., 2014) is to learn a deep generative
that very order) by assigning constant 1 s to the coordinates model pθ (x, z) = p(z)pθ (x|z) by maximizing the marginal
defined by (gt(a) , h(a) t ) ∀a ∈ A and the corresponding channel.
log-likelihood of the training data D. Due to intractabilities
Since agent positions live in real-world coordinates, a transfer that arise from the integration over the latent variables z, the
into image pixels requires a translation (gt(a) , h(a) marginal likelihood is substituted by some variational lower
t ) + t with
t = ( 105 , 68
), effectively shifting the origin from the center bound to infer the model parameters. This requires introducing a
2 2
of the image to the top left corner. The remaining channels variational approximation qφ (z|x), which is used to approximate
track motion information, with velocity values acting as the intractable true posterior. The resulting (negative) evidence
value assignments for the indices instead of constant 1 s. lower bound (ELBO) denotes the VAE training criterion and
The speed values in g direction (!gt ) is covered for A1 , enables concurrent optimization of θ and φ,
A2 , and A3 in channels 4, 6 and 8; the information in h
log pθ (x) ≥ Eqφ (z|x) [log pθ (x|z)] − KL[qφ (z|x) , p(z)]
direction (!ht ) is handled by channels 5, 7 and 9. All other
values in the resulting input representation xt ∈ R105×68×9 ≡ −LVAE (θ , φ; x). (1)
are 0.
The first term of (1) quantifies the reconstruction error and
In summary, the final dataset representing a soccer game is
the second term measures the distance between variational
a collection of tensor representations for each timestep D =
approximation and the pre-defined prior in terms of the
{x1 , .., x|T | } with additional label sets yS (auxiliary labels) and
KL divergence. The learned variational distributions qφ (z|x)
yB (target labels). The goal is to use the available evidence and
capture semantically meaningful low-dimensional feature
auxiliary labels to construct detectors that work effectively to
representations of the higher-dimensional observations x.
identify situations of interest defined in Yb . To this end, we
This encoded information facilitates finding a generalizable
adopt a two-stage optimization procedure, which relies on the
discriminator, especially when labels are scarce. The merits
derivation of semantically meaningful feature representations.
of such a semi-supervised instance are e.g., explored in the
This instance of semi-supervised learning is advantageous in
M1 model in Kingma et al. (2014), where samples from the
the present context because a large part of the model training
approximate posterior distribution over the latent variables
is already accomplished independently of the specific game
qφ (z|x) are used as input data for a downstream classifier (e.g.,
situation of interest. Consequently, the general detection design
an SVM) to learn a decision boundary in latent space.
can be described based on the following stages:
SoccerVAE
2 A description of the exact form and type of the label information used in this work We begin with a rather straight forward application of VAEs to
is given in section Experimental Setup. the problem at-hand. The SoccerVAE uses the same optimization

Frontiers in Sports and Active Living | www.frontiersin.org 3 November 2021 | Volume 3 | Article 725431
Fassmeyer et al. Toward Automatically Labeling Situations in Soccer

FIGURE 1 | Detection worflow. The building blocks outlined in orange are presented in sections Static Models, Sequential Models and denote the technical
contributions of this paper. The proposed LabelVAE methods for feature extraction operate on Y, while the SoccerVAE methods are trained fully unsupervised.

target as the vanilla VAE (cf. Equation 1) so that only the input The role of the encoder is to transform a static game
and resulting choices on distribution type and architecture design situations into fixed-size vector representations. We use strided
need to be considered3 . Regarding the former, the generating convolutions with the leaky rectified activation (Maas et al., 2013;
distribution of the generative model pθ (x, z) is modeled as a Xu et al., 2015) and batch normalization to process the input
multivariate distribution of independent Bernoulli parametrized tensors. A fully-connected layer is dedicated to mapping the
by a decoder neural net with parameters θ : final representation onto the parameter space of qφ (z|x), i.e., to
the mean and standard deviation vector of a diagonal Gaussian,
D
! which are used in conjunction with N (!|0, I) to generate the
pθ (x|z) = Bernoulli(x|µ(z; θ )) = Bernoulli(xj |µj (z; θ )), latent vector z.
j=1
LabelVAE
where D is the dimensionality of x and µ(µ1 , . . . , µD ). The goal is to infer continuous latent embeddings that capture
aggregates the individual µj ∈ [0, 1] parameters for each pixel. beneficial properties to detect a predefined (generally speaking:
This consitutes a reasonable design choice as we constrain the rarely occurring) game situation of interest. Hence, the quality
observed values to lie in the interval [0, 1]. of our approach is not primarily measured by reconstruction
Our generative and inference network definitions can be seen errors but in terms of the ability to discriminate between different
as instantiations of the class of CNN proposed by Radford et al. types of situations in the subsequent supervised learning task.
(2015). Specifically, the network µ(z; θ ), which incrementally The second static model thus aims at directly optimizing a
converts a sampled vector z to the observation space x ∈ classification network. The model uses a VAE over the input
R105×68×9 , is implemented using fractional-strided convolutions variables that serves an effective regularizer. However, our
with ReLU activations (Nair and Hinton, 2010) and a sigmoid envisaged optimization strategy is based on the extraction of
activation for the output layer, as well as batch normalization general feature representations via pre-trained parameters to
layers to reparametrize the intermediate layer activations (Ioffe enable flexible adaption to the task at-hand.
and Szegedy, 2015; Bjorck et al., 2018). Each of the convolutional The generative model reflects that causal factors of the
layers has kernels of the same size, with the number of kernels observed x can be broadly categorized into label-specific and
per layer decreasing proportionally to the depth of the network. label-unspecific factors,
All four proposed models deal with continuous priors given in pθ (x, a, z) = pθ (x|a, z)p(z)p(a), (2)
form of standard multivariate Gaussians. The inference model
qφ (z|x) is a diagonal Gaussian parametrized by an encoder neural where we assume that a encapsulates all relevant label-specific
net with parameters φ, information and z the remaining label-unspecific characteristics.
The dependency structure of the inference model embodies the
qφ (z|x) = N (z|µ(x; φ), diag(σ 2 (x; φ))). consideration that the data-specific latent information z may vary
with respect to the class-specific information of a, that is,
3 Unlessexplicitly stated, these choices are reused in the derivation of the
other models. qφ (a, z|x) = qφ (a|x)qφ (z|a, x). (3)

Frontiers in Sports and Active Living | www.frontiersin.org 4 November 2021 | Volume 3 | Article 725431
Fassmeyer et al. Toward Automatically Labeling Situations in Soccer

The above approximate posterior is amenable to approximating Although direction of movement and velocities may add context
the true posterior over the latent variables to provide a tractable to the otherwise isolated situation, the idea of processing
lower bound on the log-likelihood log pθ (x). The resulting short sequences around these situations may add important
(negative) ELBO is the optimization target of an unsupervised information. Hence, in this section, we present sequential
data point variants of the previously introduced models.
" " We denote a slice of consecutive frames from the game
log pθ (x) = log pθ (x, a, z)dzda (4) D as x≤T , where T denotes the length of the game segment.
Importantly, this implies that the time specifications of the
frames xt refer more narrowly to the timestep in a segment within
#
= Eqφ (a|x) Eqφ (z|x,a) [− log pθ (x|z, a)] the soccer game x≤T = x1 , ..., xT and no longer to the timestep in
$ the overall game (as we describe it in section Problem Setting).
(5)
− KL[qφ (z|x, a) , p(z)] − KL[qφ (a|x) , p(a)] SeqSoccerVAE
A viable avenue for inferring sequence-level features is to
≡ −Lu (θ , φ; x) reconstruct the input sequence using a single global latent
variable z. While most approaches from the literature have
To encourage the model to capture the most relevant variational been developed for modeling data distributions, we revisit this
factors in the representations obtained via inference, we embed approach primarily to aggregate game sequences/multi-agent
the available supervised learning signals concurrently with the trajectories into informative vectors. Here we simply adapt the
unsupervised learning signals by means of an auxiliary classifier. static VAE objective (1) to a sequential definition by assigning a
Thus, the learning process is given by jointly maximizing the temporal dimension to the data points:
probability of each frame log pθ (x) and minimizing the auxiliary
loss given the latent space realizations a, LSeqSoccerVAE (θ , φ; x≤T ) = Eqφ (z|x≤T ) [log pθ (x≤T |z)
−KL[qφ (z|x≤T ) , p(z)]. (8)
Ls (θ , φ, ξ ; x, y) = Lu (θ , φ; x) − αEqφ (a|x) [log qξ (y|a)], (6)
To model the components constituting Equation (8), we
where ξ are the parameters of the classifier, α is a hyperparameter generalize the parameter functions for a given point to
encoding the trade-off between generative and discrimative architectures suitable for sequential data. Accordingly, the
learning and qξ (y|a) = Cat(y|π(a; ξ )). Equation (6) is essentially parameters of the approximate posterior qφ (z|x≤T ) are
a regularized classification objective. More precisely, the second obtained from the last hidden state of an encoder RNN
term quantifies the performance of a deep classification network (parameterized by φ) working on the input sequence, and the
with injected noise from the sampling operation a ∼ qφ (a|x) generating distribution pθ (x≤T |z) is modeled by a decoder
and the variational loss Lu can be viewed as a form of RNN (parameterized by θ ) conditioned on the sampled hidden
regularization imposed on the learned representations of the code alongside the previous data point, yielding the generating
supervised prediction model. distribution pθ (x≤T |z) = Tt=1 pθ (xt |z, x<t ). Thus, we force the
&
The full training criterion is then given by collecting Ls model to encode all information about the data into the latent
and Lu for the supervised and unsupervised data points of the variable since it is the only source of information available for
evidence D: data reconstruction. The overall workflow of the SeqSoccerVAE
% is illustrated in Figure 2.
LLabelVAE (θ , φ, ξ ; Du , Ds ) = Ls (θ , φ, ξ ; x, y)
(x,y)∼Ds SeqLabelVAE
% The static LabelVAE in section LabelVAE seeks to leverage
+γ Lu (θ , φ; x), (7)
discriminative information already existing in the data by
x∼Du
injecting them into the latent space via a classification network
to facilitate the detection of game situations. In this section,
where Ds := {(xt , yt ), ∀t ∈ TY } and Du := D \ Ds , and trade-
we propose a sequential generalization of the LabelVAE that
off γ balances the contribution of the unsupervised term to the
builds upon the dependencies in inference and generative parts
overall objective. This can be advantageous in situations where
of its peer. Accordingly, the SeqLabelVAE utilizes a label-specific
the labeled data is very sparse (Nl & Nu ) and therefore aim
partition of the latent space into at and z t , describing two
to externally impinge on the relative weight that is otherwise
distinct pieces of information about the data. We address the
implicitly given by the data set (Siddharth et al., 2017). We define
temporal dependency for successive observations by generating
the feature vector for SVM training by concatenating the derived
conditional independence for the random variables (the data
variables a and z into a single vector: [a, z].
and the latent variables) given the hidden states of two separate
RNN networks,
SEQUENTIAL MODELS
henc enc
t = fφ (xt , ht−1 )
A clear limitation of the static models of the previous section
is that their input is solely a single snapshot of the game. hdec dec
t = gθ (at , z t , ht−1 ),

Frontiers in Sports and Active Living | www.frontiersin.org 5 November 2021 | Volume 3 | Article 725431
Fassmeyer et al. Toward Automatically Labeling Situations in Soccer

FIGURE 2 | The SeqSoccerVAE. The model takes a sequence of frames as input and then extracts corresponding feature representations per timestep, which are
passed to an LSTM network that outputs a global summary of the sequence. The sampled latent representation is used to reconstruct the sequence.

where henct denotes the recurrent state for the inference model the ELBO
dec
and ht denotes the recurrent state for the generative model.
T
#
The latent variables of the generative model at time t %
encode the observation xt indirectly via the state representation Ju (θ , φ; x≤T ) = Eqφ (z ≤T ,a≤T |x≤T ) − log pθ (xt |z ≤t , a≤t )
hdec
t , yielding the conditional distribution pθ (xt |z ≤t , a≤t ). As
t=1
$
in the previous models, we restrict ourselves to standard
multivariate Gaussian priors for both latent variables per + KL[qφ (z t |x≤t , at ) , p(z t )] + KL[qφ (at |x≤t ) , p(at )] .
timestep. Using unconditional prior distributions may reduce
the approximability of observation sequences, but our focus is Also, we enforce the latent variables to encode discriminative
on obtaining informative feature representations rather than on information by introducing an auxiliary classifier for the
generating sequences. For the inference model, we condition the supervised training loss
LabelVAE dependency structure of the posterior approximation
on the RNN state henc
t , resulting in the factorization Js (θ , φ; x≤T , y) = Ju (θ , φ; x≤T )
T
'% (
T −αEqφ (a≤T |x≤T ) log qξ (yt |a≤t ) ,
!
qφ (z ≤T , a≤T |x≤T ) = qφ (z t |at , x≤t )qφ (at |x≤t ). t=1
t=1
where log qξ (yt |a≤t ) is the per timestep classification loss and
α is the hyperparameter that controls the trade-off between
The derivations in the remainder of this section is analogous classification and generation. Note that the label y ∈ Y denotes
to the derivation of the static LabelVAE objective. Specifically, the event annotation for the game situation x≤T , such that each
we optimize an unsupervised training instance by maximizing frame is assigned an identical label: y1 = ... = yT = y.

Frontiers in Sports and Active Living | www.frontiersin.org 6 November 2021 | Volume 3 | Article 725431
Fassmeyer et al. Toward Automatically Labeling Situations in Soccer

FIGURE 3 | The SeqLabelVAE. Dashed lines indicate modules performing inference, solid lines denote components performing generation, and dotted lines point out
the auxiliary classifier.

We define the feature vector for classifier training on Yb by the tensor representations of the games are computed as follows.
concatenating the derived variables a≤T and z ≤T into a single Firstly, the origin centered representation of the player position
vector: [aT1 , ..., aTT , z T1 , ..., z TT ]. The SeqLabelVAE architecture is is transformed into pixel values of the tensor representation.
sketched in Figure 3. This is done by adding half of the size of the pitch along the
horizontal and vertical direction to the position of the agents.
To approximate the velocities of the players and the ball at each
EMPIRICAL EVALUATION
timestep, we compute differences in positions over the last five
Data frames (corresponding to a time lag of 0.2 s), yielding movement
We operate on two matches of the German national team. The vectors of the form (!gt , !ht ) = (gt+5 − gt , ht+5 − ht ). Since
tracking data consist of (g, h) positions of all players and ball, we assume the outputs to be Bernoulli distributed, we map the
sampled at 25 frames per second. Following Dick et al. (2018), resulting speed values onto the range [0, 1]. To obtain the final

Frontiers in Sports and Active Living | www.frontiersin.org 7 November 2021 | Volume 3 | Article 725431
Fassmeyer et al. Toward Automatically Labeling Situations in Soccer

input representation described in section Problem Setting, we seconds in most cases. We obtain TP values (FP values) if
incorporate the coordinates and velocity values into a 0-tensor of any (no) element within the extracted sequences is assigned
the size of the target shape (105, 68, 9). The updated tensor forms the label of interest. Further, we define FN values as true
the input for a single timestamp. Every game consists of about action frames that remain undetected, i.e., do not occur within
140, 000 such frame representations. the positively predicted regions. We compute F1-scores for 50
distinct threshold values in the range between 0.6 and 0.98 and
Experimental Setup only report the maximum F1-score in the subsequent section5 .
As described in section Problem Setting, we define our setup We compare our approaches to a fully supervised deep
with two different label spaces: the auxiliary label space Y convolutional network that directly processes the tensor frames.
that includes all available (inexpensive) labels and the binary The architecture of the baseline is identical to the feature
(expensive) label space Yb that indicates occurrences of the extraction modules of our inference models, i.e., it consists of
game situation of interest. The auxiliary label space Y defines convolutional and batch normalization layers with LeakyReLu
the label information yS and originates in our study from the activation functions. The output dimensionality equals 1, and
event data of the respective games (roughly 4000 observations we use the standard binary cross-entropy loss for training.
per game). Note that only LabelVAE and SeqLabelVAE make That is, the baseline directly computes the prediction of the
use of these inexpensive labels in the training process to desired label without a need for an additional SVM but lacks
capture discriminative variations in the respective feature the reconstruction part of the proposed networks. We train the
spaces. For simplicity, we focus on 5 auxiliary lables, Y = model with RMSprop (Tieleman and Hinton, 2012) and a batch
{shot, cross4 , ground, pass, other}. If more than one auxiliary label size of 4. All methods are implemented with Tensoflow 2.0 (Abadi
is active in a snapshot, we select the minority label for the et al., 2016)6 .
observations in question. To ensure clarity regarding the used baseline architecture, we
By contrast, the label space Yb defines the label information replaced “feature extraction modules of our inference models”
yB used for SVM training and depends on the task at-hand. Our with “encoder network of the SoccerVAE.” We report the
exemplary use cases target game actions of increasing difficulties comparison with this supervised baseline in Table 1.
by predicting variables encoded already in the available auxiliary
labels or annotated manually by human experts. Accordingly, Predictive Accuracies
when employing fully unsupervised feature extraction methods We showcase the expressivity of our approaches on three tasks
(i.e., SoccerVAE and SeqSoccerVAE), targets yB are the only label with gradually increasing difficulty, the first one being the
information required. We elaborate on the exact construction automatic detection of cornerkicks. The task should be the
of the set yB when discussing the predictive results in the easiest one as the spatial distribution of agents is very indicative
following section. and event data provides ground-truth labels. The second task
We use one game for training and model selection and is the detection of crosses. Again, ground-truth is provided by
the other game for testing. In the training process, parameters event data, however, the spatial distribution of the agents is not
of the static and sequential VAEs are optimized as well as as obvious as for cornerkicks. For both tasks, we train the models
parameters of the support vector machine which serves as the on one game and use another one for testing and evaluation.
final classifier. After training, the best parameters are fixed and The third task is the detection of counterattacks and clearly
used for processing the test game. For every frame in the test more involving than the former two. The task is more difficult
game, probabilities of the quantities of interest are computed as than the previous two as many different temporal aspects need
follows: A static (section Static Models) or sequential (section to be learned by the model, including gaining and maintaining
Sequential Models) approach computes the embedding of the ball possession, etc. Labels for this task are provided by human
situation which is then used as input to the support vector experts. Since the effort of labling is tedious, we train the models
machine which computes the prediction of interest and a softmax only on the first half of a game and evaluate on the second.
turns this prediction into a probability. We begin with the detection of cornerkicks. For this straight
To assess the detection performance, we mainly use two forward task, the variational autoencoders are trained on a single
different performance metrics: the area under the ROC curve game. The subsequent SVM is trained on 16 labeled examples
(AUC) and the F1 score. We calculate the relevant components per class (cornerkick vs. no cornerkick), where the negative
that constitute the F1 score (true positives (TPs), false positives examples are randomly drawn from the training game. The test
(FPs), and false negatives (FNs)) as follows. To identify an action, game contains 26 cornerkick situations. The baseline uses the
we apply a threshold to the derived probability estimates for same training and testing set as the downstream SVM. Table 1
each frame of the test game. The independently detected frames (top rows) summarizes the results for the different metrics
are then converted into coherent game situations (or positive on the test/validation game. All semi-supervised approaches
prediction instances), defined as a set of detected consecutive outperform the fully-supervised baseline with SeqLabelVAE
frames where the time gap between 2 successive frames is less being the best predictor in this task. Comparing the static
than 10 s. The average length of the detected sequences depends
on the concrete application, but it is in the range of a few 5 Therefore, unlike the reported AUC values, the F1 scores are validation values as

we engage in threshold optimization.

4 The auxiliary label cross also includes corners and freekicks. 6 The source code is available at https://github.com/fassmeyer/labeling-situations.

Frontiers in Sports and Active Living | www.frontiersin.org 8 November 2021 | Volume 3 | Article 725431
Fassmeyer et al. Toward Automatically Labeling Situations in Soccer

TABLE 1 | Results for the detection of cornerkicks, crosses and counterattacks.

Task Model AUC TP-Rate Precision F1 Length

Baseline 0.909 0.904 0.478 0.620 13.624

SoccerVAE 0.944 0.940 0.578 0.716 14.445
Cornerkick LabelVAE 0.967 0.877 0.670 0.760 8.451
SeqSoccerVAE 0.975 0.886 0.792 0.824 11.054
SeqLabelVAE 0.986 0.920 0.785 0.850 14.560
Baseline 0.827 0.765 0.507 0.606 20.070
SoccerVAE 0.920 0.933 0.575 0.707 24.229
Cross LabelVAE 0.924 0.927 0.577 0.711 24.812
SeqSoccerVAE 0.931 0.983 0.578 0.728 19.138
SeqLabelVAE 0.940 0.812 0.683 0.739 16.750
SeqSoccerVAE 0.835 0.855 0.533 0.651 7.586
Counterattack
SeqLabelVAE 0.912 0.745 0.726 0.730 3.712

The highest values are indicated in bold face. The average length of the detected segments is given in seconds. All numbers are averages on the test game.

models shows decent improvements of the LabelVAE over the Analyzing LabelVAE
SoccerVAE. Furthermore, the average length of the detected To shed light on the effect of the auxiliary labels used in LabelVAE
sequences is significantly lower for the LabelVAE. Since the and SeqLabelVAE, we visualize the latent space of the former
average length is a good indicator concerning the width of the using t-SNE (Van der Maaten and Hinton, 2008) in Figure 4.
predicted amplitudes, the value can be interpreted as a confidence Recall that the generative model of LabelVAE makes use of
measure of the predictions. Though LabelVAE performs worse two latent variables a and z. The former encodes label-specific
than the sequential models, the static models provide solid information while the latter captures all label-unspecific traits.
results in this task, presumably because the agents’ distribution Thus, both latent variables are supposed to capture different
on the playing field is easily distinguishable from other game properties which actually holds true for the trained models as
situations. When comparing the sequential models, we find can be seen in the figure. Every point in the figure corresponds
that the SeqLabelVAE performs better than SeqSoccerVAE. This to a game situation and its color indicates the attached auxiliary
improvement however comes at the cost of detection lengths. label. The difference of the two latent variables is clearly visible
Next, we study the detection of crosses using the same and accentuated by a clear separation into action clusters (right
extracted features as for cornerkick detection. The classifier part of figure) for a and the absence of any class structure
is trained on 33 examples per class (cross vs. no cross), and (left part) for z. Since both variables are used to reconstruct
the test game consists of 38 cross situations. Table 1 (center the tensor frames, but merely variable a concurrently needs to
rows) summarizes the results for the different metrics for the accurately discriminate between the different actions, it stands
test/validation game. The trends are largely consistent with those to reason that z captures position-specific information useful for
of the corner detection task but at a lower overall level. The drop frame reconstruction.
in performance stems from the variance in spatial distributions of Recall, that the empirical results for the LabelVAE in Table 1
agents that render the detection of crosses naturally more difficult are based on concatenating the two latent variables a and z into a
than cornerkicks. single feature vector yielding an AUC of 96.7% for cornerkicks.
For the detection of counterattacks, static methods cannot Passing on only a single variable to the SVM decreases the
sensibly be applied as the sequential nature and complexity performance to 94.0% for z and 90.1% for a, respectively. Hence,
of the situation (change of ball possession, maintaining ball the two variables complement one another and focus on different
possession thereafter, etc.) cannot be captured by focusing on aspects of the problem.
only a single point in time. Consequentially, we only evaluate
the sequential models using the first half of a manually annotated Qualitative Assessment
game with 27 counterattack situations for training and use the To shed light on the nature of the proposed methodology, we
second half containing 33 situations for testing the classifier. compare the structure of correctly and incorrectly predicted
The inherent complexity of counterattacks render the task much examples for the detection of counterattacks on the example
more challenging compared to the detection of cornerkicks or of SeqSoccerVAE. We begin with a correctly identified counter
crosses. Table 1 (bottom rows) shows the results. As in the attack in Figure 5. The upper part of the figure shows the
previous cases, the SeqLabelVAE emerges as the model of choice. detection probabilities computed from the output of the SVM.
Albeit detection performances are below previous ones, the The black indicator on top of the figure at timestamp 68.330
findings show the potential of the models in challenging domains indicates the true label by the experts. The SeqSoccerVAE
with manual labels. The detection rate of counterattacks is still classifies the indicated segment above the threshold (dashed
above 91% AUC. line) between timestamps 68.265 and 68.355 as a successful

Frontiers in Sports and Active Living | www.frontiersin.org 9 November 2021 | Volume 3 | Article 725431
Fassmeyer et al. Toward Automatically Labeling Situations in Soccer

FIGURE 4 | t-SNE visualizations of variable z (left) and variable a (right) of the LabelVAE. Each point in the plot is a tensor frame and the color represents a particular
game action.

FIGURE 5 | TP example. (A) Predicted probability values for an approximately 70 s excerpt from the validation game. The horizontal line indicates the threshold
applied on the probabilities to recognize a counterattack. The vertical lines mark the beginning and the end of the extracted sequence. The mark at the probability
value 1 indicates the counterattack annotation. (B) The corresponding snapshot visualizations for the beginning (left) and the end (right) of the detected scene.

counterattack. The two figures below display the snapshots at constantly below the threshold and consequentially, the turnover
the beginning and end of the detected scene and clearly show is missed by the classifier. Interestingly, the expert annotation
the successful counterattack that over both halves of halfes of is at a position, where the probability for a counterattack
the pitch. has decreased entirely and stays around zero. We credit
By contrast, Figure 6 shows a false negative. The detection this poor performance to the rather crowded origin of
probabilities shown in the upper part of the figure stay the situation and the many defending players behind the

Frontiers in Sports and Active Living | www.frontiersin.org 10 November 2021 | Volume 3 | Article 725431
Fassmeyer et al. Toward Automatically Labeling Situations in Soccer

FIGURE 6 | FN example. (A) Predicted probability values for an approximately 80 s excerpt from the validation game. The horizontal line indicates the threshold
applied on the probabilities to recognize a counterattack. The mark at the probability value 1 indicates the corner annotation. (B) Snapshot visualizations of a potential
start and end point around the counterattack annotation.

ball. The situation is clearly different from the one shown a much better structured feature space that allows the SVM to
in Figure 5. learn an accurate hyperplane with only a few labeled instances.
Last but not least, Figure 7 shows a false positive. As can This, in turn, renders the approach useful for practitioners
be seen, the situation resembles the one agent in Figure 5 but as they only need to provide manual labels for a handful of
here, the turnover fails and correspondingly, there is no expert situations.
annotation. This result expresses both, the strength and the To investigate the models’ applicability in a practical context,
limitation of the SeqSoccerVAE, and possibly the use of VAEs we quantify the (human) labeling effort to achieve accurate
in general for such tasks. By using an autoencoder, we implicitly performance for the detection of cornerkicks and counterattacks,
assume that similar situations in feature space will have a similar respectively. Figure 8 shows the results. The y-axis shows AUCs
outcome in the real world. On one hand, this assumption and the x-axis depicts the number of positive training examples
allows to use many unlabeled situations to extract meaningful which are (manually) labeled. In addition, the same amount of
features and render the entire classification approach with only negative examples are introduced, however, these are randomly
a handful of (expert) labels feasible. On the other hand, once the drawn from the training games and do not need manual
feature representation is fixed, the subsequent SVM is unlikely to attention. To reduce the effect of the randomness in the training
differentiate neighboring situations although their labels suggest sets, we report on averages over five runs; error bars indicate
separation. However, the overall performance impressively shows standard error. The left part of the figure shows the results for the
that the latter case does not occur very often, resulting in an SoccerVAE and the detection of cornerkicks. A training set with
excellent total detection rate. only six instances, three (manually) labeled positive and three
randomly drawn negative ones, is sufficient to obtain optimal
Importance of Labeled Data performance. Adding more instances to the training set does not
The idea of the paper grounds on splitting the original problem of lead to further improvements.
labeling situations in soccer into two: an unlabeled7 grouping of For the detection of counterattacks with the SeqLabelVAE
similar situation by a variational autencoder (VAE) and feeding (right part of figure), the performance stabilizes for about
the learned feature representation into a support vector machine seven manually labeled data points. Increasing the size of the
(SVM) to compute the final prediction. This approach promises training set further reduces the variance that is introduced
by selecting only a few positive and negative examples and
7 Recall that we use auxiliary labels in (Seq)LabelVAE to enforce sensible groupings. renders the classifier more robust. However, the key message is

Frontiers in Sports and Active Living | www.frontiersin.org 11 November 2021 | Volume 3 | Article 725431
Fassmeyer et al. Toward Automatically Labeling Situations in Soccer

FIGURE 7 | FP example. (A) Predicted probability values for an approximately 100 s excerpt from the validation game. The horizontal line indicates the threshold
applied on the probabilities to recognize a counterattack. The vertical lines mark the beginning and the end of the extracted sequence. (B) The corresponding
snapshot visualizations for the beginning (left) and the end (right) of the detected scene.

FIGURE 8 | Resulting AUC scores for different training set sizes of the SVM. Left: SoccerVAE for cornerkicks. Right: SeqLabelVAE for counterattacks.

that only seven manual annotations suffice to accurately detect more complicated pattern, namely counterattracks, is addressed
counterattacks with a detection rate (AUC) of over 90%. in Hobbs et al. (2018) using an unsupervised clustering. By
making use of a few expert-labels, we combine a data-driven
DISCUSSION approach with expert guidance. The autoencoder-based approach
introduced by of Karun Singh8 is improved in two ways: First,
Our approach allows us to detect basic events (cornerkicks and we use a variational autoencoder and second, we extend the
passes) as well as more complicated patterns (i.e., counterattacks)
without requiring massive sets of annotated data and without 8 Opta Analytics Pro Forum, 2019 London https://www.youtube.com/watch?v=

falling back to rule-based approaches. The detection of a H1iho17lnoI.

Frontiers in Sports and Active Living | www.frontiersin.org 12 November 2021 | Volume 3 | Article 725431
Fassmeyer et al. Toward Automatically Labeling Situations in Soccer

approach to use time series instead of static snippets of positional additional input at each timestep. Chung et al. (2015) apply a
data. Bauer and Anzer (2021) compare a rule based model, to similar model termed VRNN to speech data, sharing parameters
a machine learning based one to identify the tactical pattern between the RNNs for the generative model and the inference
of counterpressing automatically across 20, 000 labels from 97 network. In Goyal et al. (2017), the latent variable participates
matches. For their trained model they extract 137 hand-crafted to the prediction of the next timestep, and the variational
features. The advantage of our approach is that it not only posterior is informed about the whole future in the sequence
requires far fewer labeled observations, but also works with very modeled by an RNN processing the sequence backwards. While
simple basic features. It is easily reproducible for any other the previously mentioned methods sample a separate latent
pattern, and, can be adjusted quickly even if definitions of variable at each timestep, Bowman et al. (2015) propose an RNN-
patterns slightly change when the game-philosophy shifts (e.g., based VAE to derive global latent representations for sentences.
because of a coaching change). The approach to modeling human-drawn images discussed
Besides the potential to reduce the costs of manual event in Ha and Eck (2017) shares many architectural similarities
data collection, our approach enables several team performance to Bowman et al. (2015), but uses an additional backward
affecting applications: The automatic detection of relevant RNN encoder. Teng et al. (2020) introduce a semi-supervised
patterns saves coaches and match-analysis departments not only training objective for modeling sequential data where the model
time, but furthermore increases consistency and offers scalability. specification draws inspiration from Kingma et al. (2014)
This can consequently be used to perform long-term analysis and Chung et al. (2015).
across multiple seasons or even leagues. Furthermore, besides
match-analysis this methodology could also be integrated in CONCLUSIONS
the player scouting process, by identifying certain beneficial
individual action patterns and finding players that exhibit We studied automatic annotation of non-trivial situations
these frequently. in soccer. We proposed to separate the problem into an
While our work describes the technical framework to achieve unsupervised autoencoder to learn a meaningful feature
these results, for it to be usable in a club environment, one would representation and a supervised large-margin classification.
need to integrate it in an application that fits in seamlessly into The advantage of this separation lied in the use of
daily routines of match-analysis or scouting departments. abundant unlabeled data that allowed for learning a nicely
structured feature space so that only a few labeled examples
RELATED WORK were needed in the classifier to learn the target concept
of interest.
This paper explores issues related to VAE-based semi-supervised We proposed two variants of autoencoders, a straight
learning, with the main contribution in this field introduced forward application of existing results (SoccerVAE) and a more
by Kingma et al. (2014). Our SoccerVAE and LabelVAE are sophisticated variant that used auxiliary labels and allowed
clearly inspired by their proposals M1 and M2. Specifically, for even more discriminative feature spaces (LabelVAE). In
the authors integrate label information into the assumption of addition to these two static variants, we devised their sequential
the data generation process, thereby obviating the necessity for peers to account for the spatiotemporal nature of soccer.
the otherwise required supervised learning task on extracted Empirically, we studied the performance of the four approaches
label-feature pairs. Recent work by Joy et al. (2020) argues on three different detection tasks, involving cornerkicks, crosses,
that explicitly modeling the connection between labels and and counterattacks. The SeqLabelVAE turned out the best
their corresponding latent variables improves the classification competitor and outperformed all others with detection rates
accuracy compared to the M2 approach and allows to of 91% AUC or higher in all problems for only a few
learn meaningful representations of data effectively. Maaløe labeled examples.
et al. (2016) also improve M2 classification performance While our methods emerged as valuable tools for detection
by introducing an auxiliary variable that leaves the original tasks in soccer, there are some shortcomings that could be
model unchanged but increases the flexibility of the variational addressed in future work. A possible starting point is to
posterior. This can result in convergence to a parameter compare the implicit regularization of our semi-supervised
configuration that is closer to a local optimum of the actual data approach against supervised sequential models with alternate
likelihood (due to potentially better fits to the complex posterior) regularization methods (Semeniuta et al., 2016). From the
while maintaining the computational efficiency of fully factorized perspective of achieving the lowest possible generalization
models. Siddharth et al. (2017) choose a more generalized error, there are several avenues for potential variations.
formulation of semi-supervised learning with VAE compared to Future work might include alternate probabilistic assumptions
the models in the work by Kingma et al. (2014). Their framework (Goyal et al., 2017; Joy et al., 2020) such as conditioning
allows choosing complex models, such as when a random variable the variational distribution on the full input sequence
determines the number of latent variables itself. (Goyal et al., 2017), novel regularization techniques for
In addition to static semi-supervised tasks, this work VAE (Tolstikhin et al., 2017; Ma et al., 2019; Deasy et al.,
methodologically touches a branch of research that describes 2020), other approaches to semi-supervised learning (Kingma
methods involving autoencoders to model sequential data. Bayer et al., 2014; Dai and Le, 2015) such as transfer learning
and Osendorfer (2014) incorporate stochasticity into vanilla (Fabius and Van Amersfoort, 2014; Srivastava et al., 2015),
RNNs by making the independently sampled latent variables an or to achieving consistent agent representations such as

Frontiers in Sports and Active Living | www.frontiersin.org 13 November 2021 | Volume 3 | Article 725431
Fassmeyer et al. Toward Automatically Labeling Situations in Soccer

graph-networks (Sun et al., 2019; Yeh et al., 2019) and tree-based AUTHOR CONTRIBUTIONS
role alignments (Lucey et al., 2013; Sha et al., 2017; Felsen et al.,
2018). All authors listed have made a substantial, direct and intellectual
contribution to the work, and approved it for publication.
DATA AVAILABILITY STATEMENT
ACKNOWLEDGMENTS
The original contributions presented in the study are included
in the article/supplementary materials, further inquiries can be We would like to thank the German Football Association (DFB)
directed to the corresponding author/s. for providing the data for this study.

REFERENCES Ioffe, S., and Szegedy, C. (2015). “Batch normalization: Accelerating deep network
training by reducing internal covariate shift,” in International Conference on
Abadi, M., Agarwal, A., Barham, P., Brevdo, E., Chen, Z., Citro, C., et al. Machine Learning (Lille: PMLR), 448–456.
(2016). Tensorflow: Large-scale machine learning on heterogeneous distributed Joy, T., Schmon, S. M., Torr, P. H., Siddharth, N., and Rainforth, T.
systems. arXiv [Preprint]. arXiv:1603.04467. (2020). Rethinking semi-supervised learning in vaes. arXiv [Preprint].
Andrienko, G., Andrienko, N., Anzer, G., Bauer, P., Budziak, G., Fuchs, G., et al. arXiv:2006.10102.
(2019). Constructing spaces and times for tactical analysis in football. IEEE Kingma, D. P., Rezende, D. J., Mohamed, S., and Welling, M. (2014).
Trans. Vis. Comput. Graph. 27, 2280–2297. doi: 10.1109/TVCG.2019.2952129 Semi-supervised learning with deep generative models. arXiv [Preprint].
Anzer, G., Bauer, P., and Brefeld, U. (2021). The origins of goals in the German arXiv:1406.5298.
Bundesliga. J. Sports Sci. doi: 10.1080/02640414.2021.1943981 Kingma, D. P., and Welling, M. (2013). Auto-encoding variational bayes. arXiv
Bauer, P., and Anzer, G. (2021). Data-driven detection of counterpressing [Preprint]. arXiv:1312.6114
in professional football—A supervised machine learning task based on Kolekar, M. H., and Palaniappan, K. (2009). Semantic concept mining based
synchronized positional and event data with expert-based feature extraction. on hierarchical event detection for soccer video indexing. J. Multimedia 4,
Data Min. Knowl. Disc. 35, 2009–2049. doi: 10.1007/s10618-021-00763-7 298–312. doi: 10.4304/jmm.4.5.298-312
Bayer, J., and Osendorfer, C. (2014). Learning stochastic recurrent networks. arXiv Lucey, P., Bialkowski, A., Carr, P., Morgan, S., Matthews, I., and
[Preprint]. arXiv:1411.7610. Sheikh, Y. (2013). “Representing and discovering adversarial team
Bjorck, J., Gomes, C., Selman, B., and Weinberger, K. Q. (2018). Understanding behaviors using player roles,” in Proceedings of the IEEE Conference
batch normalization. arXiv [Preprint]. arXiv:1806.02375. on Computer Vision and Pattern Recognition, (Portland, OR: IEEE),
Bowman, S. R., Vilnis, L., Vinyals, O., Dai, A. M., Jozefowicz, R., and Bengio, 2706–2713.
S. (2015). Generating sentences from a continuous space. arXiv [Preprint]. Ma, X., Zhou, C., and Hovy, E. (2019). Mae: Mutual posterior-divergence
arXiv:1511.06349. doi: 10.18653/v1/K16-1002 regularization for variational autoencoders. arXiv [Preprint]. arXiv:1901.01498.
Chung, J., Kastner, K., Dinh, L., Goel, K., Courville, A., and Bengio, Y. (2015). Maaløe, L., Sønderby, C. K., Sønderby, S. K., and Winther, O. (2016). “Auxiliary
A recurrent latent variable model for sequential data. arXiv [Preprint]. deep generative models,” in International Conference on Machine Learning
arXiv:1506.02216. (New York, NY: PMLR),1445–1453.
Cortes, C., and Vapnik, V. (1995). Support-vector networks. Mach. Learn. 20, Maas, A. L., Hannun, A. Y., and Ng, A. Y. (2013). “Rectifier nonlinearities improve
273–297. doi: 10.1007/BF00994018 neural network acoustic models,” in International Conference on Machine
Dai, A. M., and Le, Q. V. (2015). Semi-supervised sequence learning. arXiv Learning (ICML). 30.
[Preprint]. arXiv:1511.01432. Motoi, S., Misu, T., Nakada, Y., Yazaki, T., Kobayashi, G., Matsumoto, T., et al.
Deasy, J., Simidjievski, N., and Liò, P. (2020). Constraining variational inference (2012). Bayesian event detection for sport games with hidden Markov model.
with geometric jensen-shannon divergence. arXiv [Preprint]. arXiv:2006.10599. Pattern Anal. Appl. 15, 59–72. doi: 10.1007/s10044-011-
Dick, U., and Brefeld, U. (2019). Learning to rate player positioning in soccer. Big Nair, V., and Hinton, G. E. (2010). Rectified linear units improve restricted
Data 7, 71–82. doi: 10.1089/big.2018.0054 boltzmann machines. In Icml.
Dick, U., Tavakol, M., and Brefeld, U. (2018). Rating continuous actions in spatial Radford, A., Metz, L., and Chintala, S. (2015). Unsupervised representation
multi-agent problems. learning with deep convolutional generative adversarial networks. arXiv
Ekin, A. A., Tekalp, M., and Mehrotra, R. (2003). “Automatic soccer video analysis [Preprint]. arXiv:1511.06434.
and summarization,” in IEEE Transactions on Image Processing 12.7, 796–807. Rezende, D. J., Mohamed, S., and Wierstra, D. (2014). “Stochastic backpropagation
Erhan, D., Courville, A., Bengio, Y., and Vincent, P. (2010). “Why does and approximate inference in deep generative models,” in International
unsupervised pre-training help deep learning?” in Proceedings of the Thirteenth Conference on Machine Learning (Bejing: PMLR), 1278–1286.
International Conference on Artificial Intelligence and Statistics, JMLR Richly, K., Bothe, M., Rohloff, T., and Schwarz, C. (2016).
Workshop and Conference Proceedings (Sardinia), 201–208. “Recognizing compound events in spatio-temporal football
Fabius, O., and Van Amersfoort, J. R. (2014). Variational recurrent auto-encoders. data,” in IoTBD 2016-Proceedings of the International Conference
arXiv[Preprint]. arXiv:1412.6581. on Internet of Things and Big Data March 2018 (Funchal),
Felsen, P., Lucey, P., and Ganguly, S. (2018). “Where will they go? predicting 27–35.
fine-grained adversarial multi-agent motion using conditional variational Semeniuta, S., Severyn, A., and Barth, E. (2016). Recurrent dropout without
autoencoders,” in Proceedings of the European Conference on Computer Vision memory loss. arXiv [Preprint]. arXiv:1603.05118.
(ECCV) (Munich), 732–747. Sha, L., Lucey, P., Zheng, S., Kim, T., Yue, Y., and Sridharan, S. (2017). Fine-
Goyal, A., Sordoni, A., Côté, M.-A., Ke, N. R., and Bengio, Y. (2017). Z-forcing: grained retrieval of sports plays using tree-based alignment of trajectories. arXiv
Training stochastic recurrent networks. arXiv [Preprint]. arXiv:1711.05411. [Preprint].arXiv:1710.02255.
Ha, D., and Eck, D. (2017). A neural representation of sketch drawings. arXiv Shaw, L., and Glickman, M. (2019). “Dynamic analysis of team strategy in
[Preprint]. arXiv:1704.03477. professional football,” in Barca Sports Analytics Summit, 1–13.
Hobbs, J., Power, P., Sha, L., Ruiz, H., and Lucey, P. (2018). “Quantifying the value Shaw, L., and Sudarshan, G. (2020). “Routine inspection: A playbook for corner
of transitions in soccer via spatiotemporal trajectory clustering,” in MIT Sloan kicks,” in International Workshop on Machine Learning and Data Mining for
Sports Analytics Conference, 1–11. Sports Analytics (Cham: Springer).

Frontiers in Sports and Active Living | www.frontiersin.org 14 November 2021 | Volume 3 | Article 725431
Fassmeyer et al. Toward Automatically Labeling Situations in Soccer

Siddharth, N., Paige, B., Van de Meent, J.-W., Desmaison, A., Goodman, Xu, B., Wang, N., Chen, T., and Li, M. (2015). Empirical evaluation of rectified
N. D., Kohli, P., et al. (2017). Learning disentangled representations activations in convolutional network. arXiv [Preprint]. arXiv:1505.00853.
with semi-supervised deep generative models. arXiv [Preprint]. arXiv:170 Yeh, R. A., Schwing, A. G., Huang, J., and Murphy, K. (2019). “Diverse generation
6.00400. for multi-agent sports games,” in Proceedings of the IEEE/CVF Conference on
Srivastava, N., Mansimov, E., and Salakhudinov, R. (2015). “Unsupervised learning Computer Vision and Pattern Recognition (Long Beach, CA), 4610–4619.
of video representations using lstms,” in International Conference on Machine Zheng, M., and Kudenko, D. (2010). Automated event
Learning (Lille: PMLR), 843–852. recognition for football commentary generation. Int. J. Gaming
Stein, M., Seebacher, D., Karge, T., Polk, T., Grossniklaus, M., and Keimet, D. Comput. Mediated Simulat. 2, 67–84. doi: 10.4018/jgcms.2010
A. (2019). From movement to events: improving soccer match annotations. 100105
Lecture Notes Comput. Sci. 11295, 130–142. doi: 10.1007/978-3-030-05710-
5247_11 Conflict of Interest: The authors declare that the research was conducted in the
Sun, C., Karlsson, P., Wu, J., Tenenbaum, J. B., and Murphy, K. (2019). absence of any commercial or financial relationships that could be construed as a
Stochastic prediction of multi-agent interactions from partial observations. potential conflict of interest.
arXiv [Preprint]. arXiv:1902.09641.
Teng, M., Le, T. A., Scibior, A., and Wood, F. (2020). Semi-supervised sequential Publisher’s Note: All claims expressed in this article are solely those of the authors
generative models. arXiv [Preprint]. arXiv:2007.00155. and do not necessarily represent those of their affiliated organizations, or those of
Tieleman, T., and Hinton, G. (2012). “Lecture 6.5-rmsprop: Divide the gradient by
the publisher, the editors and the reviewers. Any product that may be evaluated in
a running average of its recent magnitude,” in COURSERA: Neural Networks for
this article, or claim that may be made by its manufacturer, is not guaranteed or
Machine Learning, Vol. 4, 26–31.
Tolstikhin, I., Bousquet, O., Gelly, S., and Schoelkopf, B. (2017). Wasserstein endorsed by the publisher.
auto-encoders. arXiv [Preprint]. arXiv:1711.01558. Copyright © 2021 Fassmeyer, Anzer, Bauer and Brefeld. This is an open-access article
Van der Maaten, L., and Hinton, G. (2008). Visualizing data using t-sne. J. Mach. distributed under the terms of the Creative Commons Attribution License (CC BY).
Learn. Res. 9, 2579–2605. The use, distribution or reproduction in other forums is permitted, provided the
Wickramaratna, K., Chen, M., Chen, S.-C., and Shyu, M.-L. (2005). “Neural original author(s) and the copyright owner(s) are credited and that the original
network based framework for goal event detection in soccer videos,” in publication in this journal is cited, in accordance with accepted academic practice.
Proceedings-Seventh IEEE International Symposium on Multimedia (Irvine, CA: No use, distribution or reproduction is permitted which does not comply with these
IEEE), 21–28. terms.

Frontiers in Sports and Active Living | www.frontiersin.org 15 November 2021 | Volume 3 | Article 725431