Department of Economics and Finance

Chair of Empirical Finance

ORDERFLOW IMBALANCE AND

HIGH FREQUENCY TRADING

CANDIDATE SUPERVISOR
Matteo Pecchiari Prof. Paolo Santucci
De Magistris

CO-SUPERVISOR
Prof. Stefano Grassi

Academic Year 2019/2020

1
 Acknowledgements

I would like to thank my Supervisor Paolo Santucci De Magistris for the

countinuous support of my study and research.
I would like to express my sincere gratitude to my mentors, Salvatore Scarano
and Fabio Michettoni, for all they have taught me about the industry and about
trading.
Finally, I would like to thank my family for their unending support throughout
this journey.

2
 Orderflow Imbalance and High Frequency Trading

Matteo Pecchiari

Abstract

This thesis aims to provide a complete description of the relationships

between orderflow, trading volume and market returns. Firstly, it was
described the trading process and it was done a review of the market
microstructure literature in order to explain the differents approaches
to the empirical investigations. Thereafter, there are illustrated the
properties of the variables and the statistical methods such quantile
regression and multiple regression that enable us to highlight many
aspects of interconnectedness among these variables. From the
outcomes of these empirical studies was possible to build a trading
strategy and consequently introduce some improvements.

3
Contents

Introduction 1

1 Framework of Financial Markets

1.1 Framework of Financial Markets 3
1.2 History and Evolution 6
1.3 Activity of Market Partecipants 9

2 Microstructure and Orderflow

2.1 Introduction to market microstructure 14
2.2 Tauchen and Pitts Model 17
2.3 Limit Order Book 20
2.4 Orderflow Imbalance 25
2.5 High Frequency Trading strategies 29
2.6 Market inefficiencies 32

3 Statistical studies and Trading Strategy

3.1 Data and Variables 36
3.2 Theory and Statistical Methods 38
3.3 Empirical Study 43
3.4 Trading Strategy 57

4 Conclusion 60

Bibliography

4
Introduction

The technological develpment and the innovations in the financial

markets made possible to collect and study the information descending
from the market activity in the microstructure.
The information obtained can be extremely useful in order to explain
the price movements in the short-term.
The first empirical studies highlight the relevance of trading volume and
successively the attention shifted to the orderflow. The orderflow is a
more advanced concept respect to the trading volume because it
attributes an additional feature to each trade executed making possible
to identify where a trade is buyer-initiated or seller-initiated.
Consequently, an investor able to get and exploit this information may
have an advantage respect the other partecipants who do not, at least in
the short-term. The partecipants who, for definition, exploit and trade
relying on these signals are the High Frequency Trader (called ‘HFT’).
They trade at extreme speed against the other investors and aim to gain
small profits countless times.
In order to support the thesis that the market activity is a driver of
market returns, we investigated with statistical tools the relationships
between returns, orderflow and trading volume for four time intervals
(30 seconds, 1 minute, 2 minutes, 5 minutes) and we were able to
develop a trading strategy based on signals deriving from these
variables.
In the first chapter it is illustrated the market framework, the evolution
during this last decades and the way how investors trade.
The second chapter describes with the market activity from a more
technical view. There are described the methods use to assign the sign
to every trade, the first model that riassumes the relations between

5
trading volume, returns with information flow that comes to the
market. In this part some studies has been mentioned in order to
understand how previous authors deal with these concepts and what
they found. Moreover, we describes the most common high frequency
trading strategies and analyzed some market inefficiencies.
The third chapter is the main part of this thesis. We studied the
distributions of the three variables taken into account, the
autocorrelation function of the orderflow and the correlation between
returns and orderflow. The first tool used to get a description of these
relationships is a multiple linear regression that shows that the effects
of the independent variables (orderflow and trading volume) are
significant. To examine more deeply the relation between returns and
orderflow we availed of quantile regression and found that positive
returns are more influenced by orderflow than negative returns.
In the light of these findings a profitable trading strategy has been
developed exploiting the orderflow and trading volume statistical
features and the relationship with returns. This strategy considers these
variable and takes in account the returns of one period ahead.
After having shown that choosing specific thresholds of orderflow and
volume can be profitable in the sample considered, another rule is
introduced in the strategy deriving from the relation between orderflow
and returns. Until the orderflow variable remains positive the trade will
not be closed in order to exploiting the buying momentum until it ends.
This improves the results of the strategy and shows how the
information deriving from orderflow and trading volume can be used
in a proper way.

6
CHAPTER 1

Framework of Financial Markets

1.1 Framework of Financial Markets

Exchanges are the centralized marketplaces for transacting and clearing

securities orders. That place may be a physical trading floor, or it may
be an electronic system in which traders can easily communicate with
each other. The trading rules and the trading system used define the
structure.
A regulated market is a system where proposal of buy or sell are put
from intermediaries, on behalf of their customers or on their own.
They are executed one against the other one match them with opposite
sign. This kind of market are managed by a market management
company, authorised from the financial autority.
An alternative to the regulated market are the multilateral trading
system (MTF), even these follow the rules imposed by the financial
autority but they can be managed by different entities from the market
management company autorized to provide investment services. There
are less information related to the securities listed.
Systematic Internaliser are intermediary enabled to investment services
on their own that in a organized, frequent and systematic way trade
financial instruments executing clients orders. It is a bilateral trading
system because there is only on intermediary to be the counteparty of
every trade. There are no specific rules about information of who issue
the securities.

7
Financial markets can organize the trading sessions in two ways:
continuous market sessions and call market sessions.
Continuous market sessions, the most common form, entail in
acontinuous trading activity for a lot of hours between traders and
permit the price discovery during the day. This allows to determine a
significant real price for a given security and consequently to increase
the efficiency of the market.
The technology covers a fundamental perspective in the organization
of the market with continuous trading session for the purpose of permit
the immediate transmission to the market of price movements. It also
covers a not less important aspect of collection and trasmission of the
trades.
In the continuous market session, a problem that oftenarises is the
excess of offer or demand for a security that can produce high volatility
in the quotation. The presence of a specialist, or liquidity provider, can
resolve thisproblem.
In call market sessions, an auctioneer, in a specified time, invites all
traders to trade a security. After all the proposals of buy and sell are
made, a price is fixed at a level to satisfyall the offers and demands.
This price is considered teorically the equilibrium price for all operators.
The biggest defect of this kind of trading session is that is notable to
satisfy the continuous flow from the market and entails the the creation
of parallel markets before and after that session. There is no interaction
between the market and the traders so the price is notable to react
promptly to the arrival of new information making preventing to the
market to be efficient and consequently less attractive. After the
introduction of automatic quotation this kind of session was
abandoned.

8
It is essential to specify also the way which the market allows to
partecipants to interactive among them and how their orders are
processed.
The first kind of markets is defined as order driven: in the order driven
market can exist limit orders and market orders. The difference between
these two orders is that the first can be executed only at the indicated
price by the trader, or at a better price (smaller for a buy order, higher
for a sell order) and they remain on the order book till the end of the
trading session. Contranstly, market orders are executed immediately at
the best current price in the other side of the market (best ask for a
market buyorder, best bid for a market sell order).
For the same price level, the limit order are sorted with a cronological
way, from the oldest to the most recent.
The other kind of market iscalled ‘quote driven’ is led from the
quotation offered by market makers. The market makers exposes
continuously two prices: a bid price at which the specialist is willing to
buy and an ask price at which it is willing to sell. The difference between
these two prices is called bid-ask spread. The market maker forms a
fluidifying element for the market because it assumes own positions
against other investors. Its profit, deriving from the continuous
quotation activity, is the bid-ask spread. The activity represents also an
obligation to the other market partecipants in order to permit the trade
at the best possible conditions. The order driven is charaterized to be
more transparent in terms of liquidity displayed. Contrastly, in the quote
driven market ispossible to see only the bid and ask of the market
maker.
In this market, the principal advantage is the immediacy in finding a
counterpart.

9
Nasdaq and London stock exchange are examples of quote driven
markets.
Recentdevelpments in the financial ecosystem led some markets to
assume a hybrid form, combining some aspects of quote driven to other
of order driven market.
For example, New York Stock Exchange adopted this form,
fundamentally it is an orderdriven market but it requires the presence
of a specialist as liquidity provider.

1.2 History and Evolution

In the beginning, the market was based on the Open Outcry auction, a
method for communication among traders, and that happened on the
pit, or floor, of the market. The orders for buying or selling securities
were written down on paper tickets, which needed to be processed by
the exchange and each counterparty’s clearing firm.
The professional partecipants of the floor trading are the broker-
dealers, figures engaged in trading securities for its own account or on
behalf of its customers. When executing trade orders on behalf of a
customer, the institution is said to be acting as a broker. Brokers help
their clients find traders who are willing to trade with them and their
profit is the commission charged on the trade.
When executing trades for its own account, the institution is said to be
acting as a dealer.
Dealers trade with their clients when their clients want trade. They will
buy and sell at bid and ask prices.
Who collect and disseminate trade ticket information from trader to
trader, or broker, are the “Runners”. The runner will usually be

10
responsible for delivering the trade order to the company’s broker
located in the exchange’s trading pit.
Brokers can take advantage of runners to execute trade orders or
personally taking, writing and executing them in the exchange’spit, for
a compensation represented by the spread charged on the trade.
Broker-dealers provide trading service to different kind of clients:
Institutional clients, large corporations, commercial clients and high-
net-worth individuals. For this reason, they played a central role in the
financial system.
Arrival of internet and the technological progress transformed the
financial markets decentralizing the multilayer framework dealers-
clients.
The competition between exchanges increased the trading liquidity and
consequently decreased the spread between the same security listed in
different exchanges.
The first innovation for the exchanges was the introduction of the
electronic trading; for the first time the matching engine of orders was
managed by a computer network. This happened in 1971, when the
National Association of Securities Dealers introduced an automated
quotation system that gives the name to the first electronic stock
market: NASDAQ.
After this event, in the 1992 the Chicago Mercantile Exchange (CME)
launched its first electronic platform, called ‘Globex’, where was
possibile trade with an electronic system for order quotation on the
futures listed on CME.
The fragmentation of electronic trading platform caused a strong
development in the Electronic Communication Network (ECN),
Alternative Trading System (ATS) and Dark Pools and allowed access

11
to trading securities outside the traditional stock exchanges (NYSE or
NASDAQ).
The ECN is an electronic system that widely disseminates orders
entered by market makers to third parties and permits the order to be
executed against in whole or in part. Trading in the ECN eliminates the
need for an intermediary to access to the market.
The alternative trading system are automated trading systems alternative
to the official exchanges used mainly from the institutional investors
and not available for the retail traders. The main feature consists in the
opportunity to access to trading activity without the presence of a
specialized.
intermediary. The advantages of these trading venues are the speed of
execution, reduction of transaction costs, the anonymity of the active
players.
Dark Pool are private forum for trading securities in a confidential way.
This makes these alternative markets very attractive for institutional
investors for buying and selling large block of securities while remaining
anonymous. Also the prices and the number of securities traded remain
hidden. This aspect makes also appealing the use of this routes because
big investors that have the need of buy or sell large blocks of securities
are sure that they will not impact the market price. The disadvantage of
using these alternative trading venues is that Some studies of FINRA,
in the United States there are 43 Dark Pools owned and managed by
investment banks and brokers such Credit Suisse, Morgan Stanley,
Goldman Sachs.
In the 2012 about 32% of the trades were done in these Alternative
Trading System.

12
1.3 Activity of Market Partecipants

In the financial markets there are different types of operators that trade
for different purposes. In this chapter it is highlighted the features and
the differences bewteen these and why they trade in that way.
Brokers are agents who arrange trades for their clients charging a
commission on every order. They also act as financial advisers for their
clients about investments ideas or organize financial plans.
In order driven and quote driven markets, brokers get orders from their
customers and match them with orders and quotes displayed by the
other partecipants. This activity is related to trades of small or medium
size.
In brokered markets, their role is to looking for traders that are willing
to buy or sell the securities; in this case the main difference is that they
don’t act as liquidity provider but they search liquidity.
In the primary market, on the occasion of issue of new securities, they
look for traders that want to buy on their behalf.
Lastly, in mergers and acquisition deals, their role is to find a
counterparty to conclude the deal.
The dealers are the other big players in the trading industry. Their
purpose is to make money trading with their own account, and not for
customers differentiating themself from brokers.
Many of them are professional traders who work on the floor or in
trading firms and the exchanges often register them
Their activity is to buy and sell to their clients supplying liquidity in the
market. This behaviour improves the efficiency and the liquidity
presence in the market making easier for the traders that want
immediacy in finding a counterparty and executing their trades.

13
They are considered as passive traders because they trade when other
want to trade.
In addition to offering liquidity as brokers do, they speculate on the
price changes and consequently they trade as position traders, buying
low and selling high.
Dealers quote bid and ask prices, two side market and that usually
happens on equity markets. On other hand they can quote only one
side, usually in the bond market. Theyaim to capture the bid-ask spread
and their profit is called ‘realized spread’, calculated after that all their
positions on the order book are filled with traders’ orders. The quotes
displayed by dealers have a time expiration, after that they are cancelled
to put other orders at better condition.
The best situation for a dealer is when the orderflow on the market is
strong so they continuously get their orders filled and update newone.
Consequently, they construct positions on their portfolios, called
‘inventories’, and they try to keep them in balance to mitigate the market
risk.
One of the biggest risk of this partecipants is the adverse selection risk.
They may trade against informed traders causing an inventory
imbalance after the future price movements.
The way they reduce risk on their portfolios is a continuous hedging on
correlated instrument or rising the ask side or decreasing the bid size.
Sometimes, some investors, called block initiators, need to trade block
trades and the sizes are too large to be filled using the liquidity present
on the market. The demand for such liquidity is found thanks to
presence of block dealers, block brokers or large buy-side traders,
alsocalled ‘block liquidity suppliers’. The consequence of these trades
are significant for the volumes and the prices of the security required.

14
The New York Stock Exchange defines a block trade as 10.000 shares
or more and most of them usually exceeded a quarter of a day’s average
trading volume.
There are some problems related to block trades: latent demand
problem that makes it hard to find necessary liquidity, the order
exposure problem that makes reluctant the initiators of the trade to
show the size for fear to scar the market for a better informed trader
and the price discrimination problem that makes the liquidity suppliers
reluctant to trade because they that more size will follow after that.
Block Dealers, or Block Positioner, or Block Facilitators take charge
these trades from their clients and break them into smaller parts and
distribute in the market or they look for other traders interested to do
the counterparty of them. They make transaction cost analyses to
ensure they will get a profit from these trades.
Block Brokers don’t take charge these big positions and consequently
they don’t assume any risk for the operation but they look for traders
that will fill these orders. Usually it becomes easier when the block
broker assemble many traders to fill that and that gives the name to
them as Block assemblers. Their profit is given by the commission
charged for this service.
The biggest typology of investors are the value traders. They buy or sell
instruments that they believe are undervalued or overvalued using all
the available information on the security that they want to trade. They
also act as liquidity providers even if they don’t considered themselves
in this way. They get into a trade if the price of a security differs from
its fundamental value and this can happen in response of the arrival of
new information or the movement caused by uninformed traders. The
situation when the price moves from its fair value is given by the fact
that the dealers do not recognize that the orderflow on the market

15
thatpushes the security price is given by uninformed traders and
consequently the adjust their inventory in order to converge to their
target levels. In a situation when a large number of uninformed traders
demand liquidity selling stocks the the dealers will offer them a bid to
satisfy that demand. This will cause a decrease of the bid prices and so
the carrying value of the dealers will be lower than the price before the
massive sell. Therefore they accumulate long position in their
inventories and they will adjust the prices accordingly. At that point the
value traders will take action buying the stocks to the dealers that are
ready to liquidate their position and get some profits. The price of the
security will return to the fundamental value.
The main problem of a value trader is the adverse selection: infact a
value trade can be a counterparty of a better-informed traders that
possess information not yet disclosed to the market.
A particular category of investors there are the arbitrageurs, speculators
that trade on information about relative values. They follow strategies
in order to exploit differences in correlated instruments. The correlation
is given by the fact that their values depend on common factors and so
they tend to behave in the same way. The divergence from this relation
gives birth to arbitrage opportunity and it is defined arbitrage spread.
In a pratical view, an arbitrageur buy the instruments which is relatively
cheaper and sell the instrument which is relatively expensive. The
arbitrageurs construct so called ‘hedge portfolios’, constitute by
different legs, one or more long and one or more short in order to
reduce the total risk of portfolio. The size of one leg is identify as the
arbitrage numerator and the relation between one leg and the other
opposite is called ‘hedge ratios’ of portfolio. There are different
strategies around this concept: pure arbitrages for instruments that their
value of hedge portfolio is mean reverting. Risk arbitrages or speculative

16
arbitrages are strategies involving a non-stationary component of hedge
portfolio. A famous and very used arbitrage strategy is the statistical
arbitrage: if there are two securities with a non-stationary feature and
their difference is stationary (in the econometric terminology it’s called
‘cointegrated time-series’) for the variance, it means that the spread
between the two securities is bounded. The arbitrageur will buy and sell
the two instruments when the difference arrives at maximum level
observed exploiting the mean reverting feature of the spread.

17
 CHAPTER 2

Market Microstructure, Orderflow and High

Frequency Trading

2.1 Introduction to Market Microstructure

In the financial markets there are a lot of different securities quoted,

but all of themhave a common feature: the way howthey are traded. The
area of financial economics that focuses on the trading processis the
market microstructure. This branch studies different types of
aspectssuchas market structure and design, price formation and price
discoveryprocesses, transaction and timing costs, information and
disclosure and market maker and investor’s behavior.
The market structure and design focuses on the relationship between
the process of the prices determination and the trading rules and trading
protocols adopted by the exchange. The research addresses to study
how the market structure affects the trading costs, the efficiency of the
market and the disclosure of the information.
The price formation and price discovery processes focuses on how the
fair price is determined in the market studying the behavior of market
partecipants. Nowadays, a very important and actual issue aboutthis
concept is how the high frequency traders affect the price discovery
process because they are able to collect and exploit the arrival of new
information faster than everyone else in the market universe.
The transaction and timing costs focuses on how the execution
methods impact the returns of an investment. It might be possible to

18
classify the processing costs, adverse selection costs, and inventory
holding costs under the heading of transaction costs.
The information and disclosure factor focuses on the availability of
market information, the transparency and the impact of the information
on the behavior of the market partecipants. The market information
widespread are price, breadth, bid-ask spread, reference data, trading
volumes, liquidity and securities information.
A fundamental aspect of the market that implies the efficiency is the
liquidity. It is considered the immediacy to convert an asset into cash
or viceversa and reflects the theability to trade immediately by
executingat the best available price.
It has three main features: depth, tightness and resiliency. The depthis
the quantity of buy and sell orders of one asset around the current price.
A security with a deep orderbook enable to trade large orders without
causing large price movements.
Tightness refers to the bid-ask spread and it can be considered the cost
of trading. Tight spread means efficiency for the market and represent
a small cost to get it or out of a position.
Resiliencyis the ability of the market to recover from an event. If a
market is considerable resilient the price discrepancies are less frequent
and it can be considered efficient for the price discovery process. The
role of liquidity is fundamental to attract investors and to raise the
reputation of the market because it leads to have an orderbook full of
proposals, it tightens the spread and help to bring back the market price
to a fair price after a shock.
The trading ennvironment is defined with rules and protocols to place
all the partecipants in the same fair position. These rules cover different
aspect of the process such order precedence, requirements for trade

19
sizes, pricing increments, opening/closing procedures and the reactions
to shock events.
The order precedence specify how the market orders execute with the
limit orders on orderbook. Most markets give the priority to orders with
the best price (higher for buy limit orders, lower for sell limit orders)
and for the orders put on the same price level, the precedence is given
by the time of entry.
The minimum price increments, called ‘tick sizes’, is the minimum
difference between two prices. Bigger the tick size, more profitable is
to provide liquidity, or trading with limit orders. In the US equity
markets, the stocks have 1 cent of tick (for some penny stocks the SEC
imposed the ‘5 tickspilotprogram’, that increases the tick size to 5 cents
makes more expensive get in and out of a trade).

20
2.2 Tauchen and Pitts model

The trading activity raised an interest in the financial research in order

to analyze the relation between the trading volume and the returns.
Tauchen and Pitts (1983) presented a structural model that concerns on
the relationship between the variability of daily price change and trading
volume.
The market is composed by J active traders that can take long or short
position. There are I trades. That number of traders is assumed to be
fixed during the day The market moves from one equilibrium price to
another one reacting to the arrival of new information, called
‘Information Flow’.
The desired position 𝑄𝑖,𝑗 of the 𝐽𝑡ℎ trader is given by the relation

∗
𝑄𝑖𝑗 = λ (𝑝𝑖,𝑗 - 𝑝𝑖 )

with λ > 0 and 𝑝𝑖 is the transaction price.

∗
𝑝𝑖,𝑗 is the 𝐽𝑡ℎ trader’s reservation price.

A positive value of 𝑄𝑖𝑗 represent the willing of the 𝐽𝑡ℎ trader to take a
long position, a negative value means a desired short position. The J
traders on the market have different reservation prices.
The market clearing condition requires that ∑𝐽𝑗=1 𝑄𝑖𝑗 =0 that implies
that the

1
𝑝𝑖 = ∑𝐽𝑗=1 𝑝𝑖,𝑗
∗
𝐽

21
and defining

∗ ∗ ∗
∆𝑝𝑖,𝑗 = 𝑝𝑖,𝑗 - 𝑝𝑖−1,𝑗
∆𝑝𝑖 = 𝑝𝑖 - 𝑝𝑖−1

It is possible to derive the trading volume

λ ∗
𝑣𝑖 = 2 | ∆𝑝𝑖,𝑗 - ∆𝑝𝑖 |

The assumptions support that the changes in the reservation values are
due to the changes in global information and trader-specific
information:

∗
∆𝑝𝑖,𝑗 = ϕ𝑖 + 𝜓𝑖,𝑗

where ϕ𝑖 is the common component and 𝜓𝑖,𝑗 is the individual

component.
The moments are:

E[𝜙𝑖 ] = E[𝜓𝑖 ] = 0 and Var[𝜙𝑖 ] = 𝜎𝜙2 and Var[𝜓𝑖 ] = 𝜎𝜓2

It is possible define the return as

1
𝑟𝑖 = ∆𝑝𝑖 = ϕ𝑖 + 𝐽 ∑𝐽𝑗=1 𝜓𝑖,𝑗

If a good news hit arrive to the market, the common component

ϕ𝑖 increases causing an increase in 𝑟𝑖 .
If a trader has a good private news, the return increases but it is
1
averaged by other traders (from the component 𝐽 ).

22
 The trading volume becomes

λ
𝑣𝑖 = 2 ∑𝐽𝑗=1 | 𝜓𝑖,𝑗− ̅̅̅̅
𝜓𝑖 |

The unconditional moments of 𝑟 and 𝑣 implied by the Tauchen and

Pitts model are:

𝜇𝑟 = E [𝑟𝑖 ]
𝜎𝑟2 = 𝜎𝜙2 + 𝜎𝜓2 / 𝐽
1
λ 𝐽 2 𝐽−1
𝜇𝑣 = 2
√ 𝜎𝜙 (
𝜋 𝐽
)2

𝜆
𝜎𝑣2 = ( 2 )2 Var[| 𝜓𝑖,𝑗 − ̅̅̅̅
𝜓𝑖 |]

These results lead to the conclusion that as the number of J traders

increases, the expected volume increases and the volatility of returns
decreases.
The model supports the idea that the information flow that comes into
the market in form of private or common information is reflected by the
activity of the traders and influences the moments of the returns and the
trading volume.
This is one of the first models that give importance to the trading volume
and treats it as a dependent variable. Furthermore, the trading volume
will be studied as explanatory variable of the returns or the magnitude of
the returns.

23
2.3 Limit Order Book

The different instructions to buy or sell are represented by the different

types of orders sent to the market. It is important to highlights these
two types because they influence the trading process in different ways:
market orders and limit orders.
A market order is an instruction to trade immediately at the best price
available. The trader that chooses to buy or sell a security with this order
is considered aggressive and he wants immediacy to enter in the market
wihout having certainty about the execution price. It is called ‘price
taker’ and from a liquidity perspective, it removes liquidity.
Instead, a limit order is an instruction to buy at a maximum price or to
sell at a minimum price. The trader that sends a limit order is considered
passive and he waits for the execution have certainty about the price he
will get. It is called ‘price maker’ and he adds liquidity to the market
In the financial ecosystem, the exchanges allow to the listed securities
to be traded with the engine that matches buyers to sellers: the Limit
Order Book.
It is a centralized database of pending limit orders for a specified price
or better and with a given size. It is a transparent system for the market
partecipants and the matching algorithm for limit orders with market
orders works with a price time priority algorithm: it works with
matching priority for price, time and lastly visibility.
A buy limit order with the highest price represents that buyer is willing
to pay more than the other traders and he gets the priority of execution
when a sell market order will arrive. Consequently, a market order is
executed on the best proposal, best bid or best ask (in the US market
the first levels are guaranteed by NBBO, National Best Bid Offer).
After that, it follows the FIFO principle: First in First out.

24
The limit order that is submitted earlier will be executed before the later
arrival.
Therefore the queue of limit orders follows this chronological rule.
In most markets, there is the possibility to put hidden orders, that
stands on the book without being displayed. They are the last limit
orders to be executed because the precedence is given to the visible
orders.
The ordebook events can be mainly identified in three types:
submission of limit orders, cancellation of limit order, execution of
market orders.
The empirical studies in the financial literature focus their attention on
studying the price impact of the orderbook events. The outstanding
limit orders, called Market Depth, concept related to the market
liquidity, is associated with the magnitude of price changes. In a thin
orderbook, where the market depth is low given by the small size and
the low number of limit orders, it is easier to cause a bigger price change
given to the fact that there is less need of market orders, respect a thick
orderbook, to buy or sell all the first level proposals and than to move
the price.
The price dynamics are also conditioned by the arrival, the size and the
direction of the market orders.
There are other kinds of orders in order to address to the different
needs of investors. For example, conditional orders allow to traders to
put one or more conditions before the order can be submitted to the
market. These conditions may activate an order if the price is above or
below one price level, or before or after a specified time. A stop order,
usually used by traders to close a position at target price or at maximum
loss, is a conditional order that activated a market order when the
market price reaches a point.

25
Limit if-touched order is a standing limit order but it remains hidden
from the order book until the the price reaches a specified price level.
Investors that need to submit large volume orders may wish to hide the
full size of their orders to avoid adverse behaviours from other market
partecipants given by the big quantity displayed. To overcome this
problem, they can use iceberg orders, that submit a specified portion of
the total size. After this part is executed, another part of the total order
is submitted until all the quantity is executed.
To get a more precise empirical investigation of the relationship
between the market response and the microstructure activity, is not only
important to take into account the trading volume, but also to make a
buy-sell classification of the ordeflow.
There are different trade classification algorithms: Tick Rule, Lee-Ready
Algorithm and Bulk Volume Classification.
The Tick Rule is a level-1 algorithm that classifies a trade as buy if the
trade price is higher than the price of the trade before, called ‘uptick’.
Conversely, if the trade price is lower than the price of the trade before
it is considered a seller-initiated trade, called ‘downtick’. In the event
where the price is equal, it is called ‘zero-tick trade’ and it needs to look
the closest prior trade price to attribute a zero-uptick trade or zero-
down trade.
This kind of classification requires only trade data without the need to
have the current bid and ask quote.
The Lee and Ready’s algorithm [8] is a level-2 algorithm and it is a more
complex trade classification. It needs the TAQ Data (Trades and
Quotes, trade price, best bid and best ask price) to assign a feature to
the trade.
If a trade occurs in the mid-point between bid and ask, the classification
works using the tick rule. Instead, it uses a quote rule that assigns the

26
sign to the trade if it has an execution price above or below the
midpoint, classifying buyer-initiated or seller-initiated feature.
The bulk volume classification is a probabilistic method that assigns the
sign of trading volume aggregating the trading intervals into bars
composed of equal size of volume. For each bar, a fraction of volume
is determined to be buyer-initiated, while the seller-initiated amount is
given by the difference between the total volume and the buyer-initiated
volume. The formula of the BVC algorithm takes in account the
standardized price change for the reason that if the last trade price of a
volume bar increases respect to the last trade price of prior volume bar
the buyer-initiated volume increases.
The feature of limit orderbook that is the presence of bid and ask price
raises an important issue for the empirical market microstructure
research and for who is involved in working with high frequency data:
the bid-ask bounce.
The first problem is to determine the true price of the security and the
presence of two prices at the same moment makes it tricky.
In some researches or in the phase of planning of a trading strategy the
price used is the mid-quote between ask and bid.
The second problem, illustrated by Roll (1984), is the impact of bid-ask
bounce on the time series of returns. He proved that this feature
induces negative autocorrelation in observed returns.

27
 Figure 2.1: Limit Order Book

Figure 2.2: Arrival of Limit Buy Order on bid side

Figure 2.3: Arrival of Market Sell Orders executed on the best bid

28
2.4 Orderflow Imbalance

The limit order book contains all the information available to investigate
how the events of submission or cancellation of limit orders and arrival
of market orders influences the securities returns.
In the market microstructure literature some authors focus their
empirical studies modeling the orderbook events through a variable
called ‘Order flow Imbalance’. The orderflow imbalance variable
captures the events on the orderbook and it is shown that has an
explanatory power of the traders’ intentions.
Chordia and Subrahmanyam [4] in their study developed a trading
strategy based on the order imbalance. They focus on the daily relation
between the order imbalance and stock returns. They compute the order
imbalance variable signing every trade with the Lee and Ready algorithm.
Bringing together all the trades, they obtained the daily order imbalance
for each stock of the sample. Moreover, they defined two variables as
order imbalance: OIBNUM, as number of buyer-initiated trades minus
the seller-initiated trades scaled by the total number of trades, and
OIBVOL, as the buyer-initiated dollar volume less the sell-initiated dollar
volume scaled by the total dollar volume. The order imbalance exhibit a
positive autocorrelation of the time-series and the authors explained this
results supporting the idea that traders sent their orders splitting them
over time to minimize price impact. This also entails a positive
intertemporal correlation between the order imbalance and price
changes. The imbalance in terms of number of transactions exhibited a
stronger autocorrelation respect the dollar imbalance. The time-series
return regression used market adjusted open-to-close returns in order to
reduce cross-correlation in error term across stocks. The

29
contemporaneous imbalance and four lags of order imbalance were used
as explanatory variable.
The results showed that about 75% of the coefficients of the first lag
were positive and statistically significant, instead the latter coefficients
were smaller or negative. Given these results, they developed a trading
strategy that buys a stock at ask price at open and close the position
selling at bid price at close if the previous day’s order imbalance was
positive. The results present a statistically significant daily average return
of 0.09% for the entire sample of stocks.
Another important study that deserves to be mentioned is made by Rama
Cont [5], ‘Order book price events’, where a more complex and complete
form of order imbalance was used to explain the price action, called
‘Order flow Imbalance’. It represents the difference between the order
flow at the best bid and the order flow at the best ask taking in account
the changes in size of these first two levels. In this way, the market orders
are also taken in account because a market sell order decreases the bid
size and conversely a market buy order decrease the ask size. The
computing of order flow imbalance variable is made by two part: one
represents the events related to a buy pressure and the other one is
related to a sell pressure. The buy pressure is made by the cancelation of
a limit order on the ask, a submission of a limit order on the bid and the
arrival of a market buy order.
On the other hand, the sell pressure is made by the cancelation of limit
order on the bid, a submission of a limit order on the ask and the arrival
of a market sell meritaorder.
The limit orders events can generate three situations and it is important
to identifying them in order to highlight the different behaviour of the
orderflow function.

30
If the bid price is lower than the previous bid price, this means that a
market sell order arrived and filled all the previous buy limit orders or
the previous limit order were cancelled. If the bid price is equal to the
previous bid price, the function takes in account the difference between
the current quantity on the bid and the previous quantity on the same
level. If the bid price is greater than the previous bid price, this means
that there was a submission of buy limit orders at a higher level. They
divided the data in interval of 30 seconds and run the regression for
clusters of 30 minutes.
They found, with an OLS regression, that the coefficients of the Order
Flow Imbalance, considered as price impact coefficients, explaining the
Price change be positive and the results were statistically significant in
98% of sample. Moreover, it was regressed the price change with Trade
Imbalance, computed as the difference between the buyer-initiated
trades and seller-initiated trades, but the results were much weaker than
the regressions with Order Flow Imbalance as explanatory variable.
From this study conducted by Rama Cont that highlights a positive
relation between
Another prominent study was done by Huang [7] that studied the
relations between the order imbalance and the return and order
imbalance and price volatility for a sample of 150 stocks in the tender
offer announcement day.
The order imbalance is computed for three different timeframes, 5-10-
15 minutes and the trade assignment is computed using Lee and Ready
algorithm.
First, they employ a multi-period regression model to find the
coefficients of contemporaneus and lagged order imbalances. These
coefficients found in this studiy are negative for lagged order imbalance
on current returns. They interpretate the results attribuiting the reason to

31
the the market makers’ behaviour to accomodate big price impacts
caused by traders that react to the news. Consequently, the inventory
levels of market makers are a fundamental aspect to consider to attribuite
an interpretation to returns in these days. Instead, they found the
coefficients of the contemporaneus order imbalance be positive and
significant for the 59.30% of the results. As result of that, they developed
a trading strategy based on this findings. The strategy buys a share at
trading price when the order imbalance is positive and sell it at the trading
price when it turns negative. The daily strategy returns outperform the
daily returns of the underlying.
Another interesting aspect studied in this paper was the order imbalance-
volatility relation for the three intervals. They employ a Garch (1,1)
model to estimate it and they found that the order imbalance has negative
impact on price volatility. As the time interval getting longer the order
imbalance effect on volatility became weaker. The explanation in this
case is given by the fact that the market makers have an obligation to
reduce price volatilities and they reduce the effect of the news induced
by the arrival of orders by traders. They face the variability of the stock
after the release of the news increasing the bid ask spread in order to
make more profit.
From the empirical studies emerges a strong interest from the academic
community to study the impact on the prices of the orderbook events.
The focus is on the orderflow imbalance considered one of the main
driver for the price change. It captures the traders’ intent to get in to a
long or short position and consequently has an explanatory power to
anticipate short-term movement.

32
2.5 High Frequency Trading strategies

The technological process enable some professional market

partecipants to deploy the information flow in the microstructure and
build algorithmic trading strategies characterised by high speed of
execution and rapid-decay alpha, defined ‘High Frequency Trading
strategies’. The introduction of these sophisticated and low latency
machines raised a strong interest in the academic researches and lead
the financial regulator to face this new challenge.
The first definition was given in June 2012 by the U.S. Commodity
Futures Trading Commission (CFTC) as follow:

“High-frequency algorithmic trading is a trading technique

characterised by describing these strategies as algorithms for decision
making, order initiation, generation, routing, or execution for each
individual transaction without human direction. The high frequency
trading firms benefit from low-latency technology that is designed to
minimize response times, including proximity and co-location services,
high speed connections to market for order entry and high message rate
(orders, quotes or cancellations).”

The co-location services allow the high frequency trading firms to gain
a speed advantage on the other partecipants thanks to the trading
servers situated in the same facility of exchange’s matching servers.
Given this advantages of milliseconds, the information contained in this
small time interval can be extremely remunerative if exploited with the
right strategies and the the right models
The HFT become therefore some of the most important market
partecipants, especially in the microstructure environment.

33
 Their activity can be divided in two main ways: aggressive trading and
passive trading.

Latency Arbitrage

Latency Arbitrage is the first strategy, born thanks to the use of the
fastest technologies that allow these big partecipants to receive the
signals from different markets before everyone else. Given the Law of
One Price, when the same security is listed in more than one exchange,
the difference between prices among different exchange produce an
arbitrage opportunity for the fastest trader that realizes that. The fiber
that the big players of Wall Street adopted allow them to have speed
advantage and exploiting that. For example, if in one market the security
has its ask price quoted at 74 usd and in another one the same security
has its bid price quoted at 74.10 usd, the opportunity is exploited by an
algorithm is to buy at ask price at 74 usd and sell it at the bid price of
74.10 usd.

Spread Capturing

This is a passive strategy and the high frequency trader acts as a market
maker. It provides liquidity in the order book when is more convenient
and it generates profits when the incoming market orders hit the pending
limit orders. The profit is given by the spread between the best ask and
the best bid. It is particularly profitable in high liquid markets when the
probability of being executed is higher
than less liquid markets. Obviuosly, this strategy is subject to inventory
and adverse selection risks linked to the market making activity risks.

34
Quote Matching

The arrival of limit orders can cause short term price movements and the
High Frequency Trader is able to detect the market impact, takes
advantage of this information and takes profit after few ticks of
movements. The success of this strategy relies on HFT’s ability to
understand which limit orders will generate positive or negative market
impact.

Spoofing

This strategy is considered market manipulation because it modify

artificially the other trader’s view in order to to obtain a better execution
price. The high frequency trader submits a big limit order in the
orderbook in order to induce other traders to get a distorted information
from the orderbook. For example, the HFT intents to buy will send a sell
limite order, with the intent of not being executed, and a hidden buy limit
order. The other traders, seeing a big sell order will sell consequently
causing a downturn in price. In this way, HFT will be able to get a better
purchase price. After the execution, the pending sell order will canceled.
This strategy is illegal in the United States.

Momentum Ignition

In this strategy, the High Frequency Trader sends market orders in order
to get a large position. This behaviour induce other traders that see the
incoming flow to take a position on the same side emphasizing the price
movements. The HFT that have bought or sold at the beginning of the
move, is able to liquidate the position and get some profits. This strategy

35
is most profitable if the price gets to levels where a lot of stop orders are
positioned. The activation of those orders accelerates the movement and
increase the potential profit of the strategy.

Pinging

The way to benefit from other liquidity provision is the pinging strategy
or Liquidity Detection. The HFT aims to discover hidden orders or price
levels that trigger orders submitted by algorithmic strategies and to
exploit the price movements consenquently to these incoming flow.
Once in the price level, for example a bid price, is detected the presence
of a hidden order (called also ‘Iceberg’) the HFT will buy just above in
order to have a risk-reward trade favourable.

2.6 Inefficiencies of market microstructure

The market microstructure presents some cases of inefficiencies that

cause large price movement and increments the short term volatility or
deviations from the fair price. These efficiencies can be the reasons of
losses or can be the alpha of trading strategies, especially for High
Frequency Trading strategies.
The most famous kind of inefficiency is the so called ‘Flash Crash’, an
event in which there is a deep withdrawal of limit orders on the bid size
followed by a big flow of sell market orders, executed at prices much
lower than expected. This causes a strong lowering of prices and a
negative return for the security. The most famous case of flash crash
happened the 6th May of 2010 on the Dow Jones Futures.

36
Some authors, Christensen and Kolokolov [2], analyzed the
microstructure features such as trading volume, market depth, bid ask
spread during the flash crashes on a sample of French Stock. They found
an expected behaviour of these variables. While the price was dropping
faster the trading volume increased, due to the fact that the market
partecipants reacted immediately to the ‘Panic Selling’. The market depth
and bid ask spread increased during this event, due to the market makers’
behaviour to protect their inventory levels from the high volatility.
The widening of the spread causes some ‘jumps’ on the orderbook, the
price goes from on level to the next without passing for every tick of the
movement.

Figure 4: Jump of the order book on the Future Dax Dec 19 at 11.932,5 – Volcharts
Trading Platform

The figure x.x represents 1 minute chart of Future Dax. The grey square
indicated by the arrow is an inefficiency of the market. Given the auction
process that characterize the way which the trading activity is carried out,

37
all the price levels have to be traded before passing across. If the market
let an empty space, it means that an auction has not been done and the
price moves jumping some price levels. There is an increase in the spread
between bid and ask and some sell market orders hit the bid. They
execute all the buy limit orders and the best ask moved down.
There were not buy market orders that hit the ask and consequently the
price moved down. Letting undefeated the first ask the first levels of the
book, the price moved down letting the auction at the precedent price
level incompleted.

Figure 5: Jumps of the order book on the Future Dax Dec 19 at 11.819,5 11.816
11.814 – Volcharts Trading Platform

The figure x.2 shows the same kind of inefficiences. The price, before
going up, did not pass all the price levels, it let some price levels not
traded causing an inefficiency in the market.

38
 Figure 6: Flash Crash on Future Nasdaq Dec 19 – Volcharts Trading Platform

The last image, figure x.3, shows the most famous form of inefficiency,
the “flash crash”.
The chart represents the Nasdaq futures on 1 minute chart. The limit
orders standing on the bid size were fastly removed leaving some price
levels on the order book empty. All the sell market orders hit limit orders
on the bid side much lower than expected (the investor who send a limit
order is a price maker, who send a market order is a price taker). This
fast removal caused a strong price downturn followed by a sharp
recovery in the quotations.

39
CHAPTER 3

Empirical studies and trading strategy

3.1 Data and variables

In this work, the data used for the empirical analysis are relative to the
Dax Futures (FDAX, with expiration on September 2019) quoted on
EUREX Exchange. The Future Dax contract has a value of 25 EUR
for each point and the tick value is 0.5 (12,5 EUR for one tick).
These data are downloaded by IQFeed Data Feed in comma separated
values (CSV) format and they have tick-by-tick timestamp so it was
recorded every trade with a microsecond precision. The information
provided are:
- Trade identification number
- Trade time
- Trade volume
- Trade price
- Current Bid price
- Current Ask price
It has been taken into account 62 days of trading from the front future
contract, being the most liquid and traded contract and so the order
flow effects can be clearer and better than future contracts with longer
expirations.
The trading hours of the contract is from 2.00 A.M. to 10.00 P.M. but
the index is moved from the underlying stocks from 9.00 A.M. to 5.30
P.M. In the statistical study, it was considered only the data between the

40
9.00 A.M. and 5.00 P.M. in order to study only the part of the day where
the liquidity is more present and the trading activity is intense.
Thanks to the granularity of the TAQ data (Trades and Quotes) it was
possibile to compute the mid-quote between bid and ask price, variable
used in computing the order flow variable with the Lee and Ready’s
algorithm.
It was assigned a positive sign to the trade volume if the trade occured
above the mid-quote and negative sign if the trade occured below the
mid-quote.
I separated all the data in days and I divided that in 30 seconds, 1
minute, 2 minutes and 5 minutes intervals in order to study how the
order flow, used as explicative variable, has effect on the return of the
instrument at high frequency level.
The percentage returns were computed as 𝑅𝑡 = (𝑃𝑡 − 𝑃𝑡−1 )/𝑃𝑡−1
considering the close price of the intervals. Consequently, I deleted the
first element of the order flow vector in order to work with two vectors
with the same length. To obtain an order flow variable for the intervals,
all the trades occured in the same time interval are summed. If the result
is positive, it means that there was more buying pressure on the price
(more and bigger buy market orders than sell market orders). The data
used for the empirical study are the returns and the orderflow. In the
trading strategy, there is an additional variable used as trigger for the
trade: the trading volume. The trading volume gives the information
about the intensity of trading activity and associated with orderflow can
be useful to exploit the short-term momentum.

41
3.2 Theory and statistical methods

Quantile Regression

To investigate the features of a random variable we need to know the

moments of its distribution. These measures give each a singular
information on the shape of the probability function and they can be
divided for the charateristic that explain. The location measures are the
mean and the median and they characterize the average and the center of
the distribution. The dispersion measures as the variance indicate how
far the observations are spread out from their average value. The third
moments as skewness and kurtosis show the asymmetry of the
distribution about its mean and the thickness of the tails of the
distribution.
The conventional methods that describe the behaviour of y conditional
on x, are the Least Square, that minimizes
̂𝒕
∑𝑇𝑡=1(𝑦𝑡 − 𝑥𝑡′ β) 2 to obtain 𝜷
and the Least Absolute Deviation, that minimizes
̂𝒕
∑𝑇𝑡=1 |𝑦𝑡 − 𝑥𝑡′ β| to obtain 𝜷
These methods approximates respectively the conditional mean and the
conditional median of y given x.
They assume that the distribution of the dependent variable is not
affected by the values of the covariates but only the central tendency of
the dependent variable. The estimate provides a partial description of the
relationship bewteen variables and it may be more useful to know the
relationship at different level of the conditional distribution of the
dependent variable. To investigate about the potential effects of the
independent variable on the shape of the probability distribution we need
a regression method robust to changes in the shape of distribution. The

42
idea to model the quantiles of a conditional distribution 𝑓 (𝑦|𝑥 ) of a
variable y, given the covariates x, was developed and introduced by
Koenker and Basset in 1978. Quantile regression provides a more
complete description of the distribution of the conditionated quantiles
of y, given the independent variable x.
This statistical tool is especially useful when the relationship between
the variables is asymmetric, for example at the tails of the distribution,
and it cannot be captured properly by the OLS method.
Before moving forward in the model description, it is essential to define
a statistical value, called Quantile, that divides the distribution in equal
parts.
Quantiles are a position indexes distribution, the τ quantile is the value
such that

𝑃(𝑌 ≤ 𝑦)= τ for any 0 < τ < 1.

Starting from the cumulative distribution function (CDF)

𝐹𝑌 (𝑦) = 𝐹 (𝑦) = 𝑃(𝑌 ≤ 𝑦)

we define the quantile function of Y for the τth quantile as the inverse
of the CDF
𝑄𝑌 (τ) = 𝑄 (τ) = 𝐹𝑦−1 (τ) = inf {𝑦 𝜖 ℝ ∶ 𝐹𝑦 (𝑦) ≥ τ}.

The quantile that divides the distribution in two simmetrical parts is the
median.
Hao and Naiman (2007) define the quantiles as a center of the
distribution c, obtained minimizing a weighted sum of absolute
deviations. The quantile q is:

43
𝑞τ = argmin 𝐸 [𝜌 τ (𝑌 − 𝑐)]
𝑐

Where 𝜌 τ is a loss function defined as:

𝜌 τ (𝑦) = [τ – I(y < 0)]y = [(1 − τ) I (y ≤ 0) + τI (y > 0)]|y|

This is an asymmetric loss function; it is a weighted sum of absolut
deviation where a weight (1- τ) is given to negative deviations and a
weight τ is used for positive deviations.
The optimization problem, in a discrete case, becomes:

𝑞τ = argmin { (1 − τ) ∑𝑦≤𝑐|𝑦 − c|𝑓(𝑦) + τ ∑𝑦>𝑐|𝑦 − c|𝑓(𝑦) }

𝑐

In the continuous case, it becomes:

c ∞
𝑞τ = argmin { (1 − τ) ∫−∞|𝑦 − c|𝑓(𝑦)𝑑 (𝑦) + τ ∫𝑐 | 𝑦 −
𝑐

𝑐 |𝑓(𝑦)𝑑 (𝑦) }

This formulation is useful to go straight in the application of quantile

regressive models.
In reference to the i-th sample unit, it is possible to write a linear model:

𝑦𝑖 = 𝑥𝑖𝑇 𝛽𝑞 + 𝑒𝑖 with i=1,…,n

Where n is the sample numerousness, 𝑦𝑖 is the response variable in the

i-th unit, 𝑥𝑖 is the vector of p explicative variables with first component
equal to 1 to ensure the presence of the intercept, β is the vector of
coefficients of the model and 𝑒𝑖 is the error term.

44
The assumption 𝑄𝑞 (𝑒𝑖 |𝑦𝑖 , 𝑥𝑖 ) = 0 is a necessary condition to get the
errors distribution centered on the q-th quantile (0 ≤ q ≤ 1).
The linear model for the q-th quantile of response variable 𝑦𝑖
conditioned to the explicative variables 𝑥𝑖 can be written as:

𝑄𝑞 (𝑦𝑖 | 𝑥𝑖 ) = 𝑥𝑖𝑇 𝛽𝑞

̂ is estimated minimizing the following object

The parameter vector 𝜷𝒒
function

̂ = 𝑎𝑟𝑔𝑚𝑖𝑛𝜷 ∑𝑛𝑖=1 𝜌𝑞 (𝑦𝑖 - 𝑥𝑖𝑇 𝛽𝑞 )

𝜷𝒒

where 𝜌𝑞 (u) is the loss function

𝑞𝑢 𝑢 < 𝑐
𝜌𝑞 (𝑢) = {
(1 − 𝑞) 𝑢 > 𝑐

̂ = argmin ∑ 𝑇
𝜷𝒒 𝑖:𝑦𝑖 ≥𝑥 𝑇 𝑞 | 𝑦𝑖 − 𝑥𝑖 𝛽𝑞 | + ∑𝑖:𝑦𝑖 ≤𝑥 𝑇 (1 − 𝑞) | 𝑦𝑖 −
𝛽𝑖 𝑖𝛽𝑞 𝑖𝛽𝑞

𝑥𝑖𝑇 𝛽𝑞 |

Computing the parameters β for every quantile q can be expressed as a

problem of linear programming. They can be computed with the
simplex algorithm or the interior points method when the sample size
is big. Quantile regression presents some advantages respect to the
conventional methods. The Ordinary Square Estimator can be
inefficient if the errors are not normal and the presence of outliers can
significantly affects the results of the linear regression method. The

45
quantile regression estimator is more robust because the model avoids
some assumptions on the errors distribution.
An important aspect of this method is the equivariance. If there is the
need to reparametrize the data to study the effect from a different
perspective, the estimates change in the same in way of the
reparametrization leaving the results invariant.
The estimator has four equivariance property:

- Scale Equivariance:
for any α > 0 and τ ∈ [0,1]
̂ (τ; αY, X) = α𝜷
𝜷 ̂ (τ; Y, X)
̂ (τ; −αY, X) = -α𝜷
𝜷 ̂ (1 − τ; Y, X)

- Shift Equivariance
for any γ ∈ 𝑅 𝑘 and τ ∈ [0,1]
̂ (τ; Y ˔ Xγ, X) = 𝜷
𝜷 ̂ (τ; Y, X) ˔ γ

- Equivariance to reparametrization of design

Let A be any p x p nonsingular matrix and τ ∈ [0,1]
̂ (τ; Y, XA) = 𝐴−1 𝜷
𝜷 ̂ (τ; Y, X)

- Invariance to monotone transformations

If h is a non decreasing function on R
h (𝑄𝑌|𝑋 (τ)) = 𝑄ℎ(𝑌)|𝑋 (τ)

46
 3.3 Empirical studies

The aim of this work is to detect the effect of the orderflow at different
level of the distribution of returns. Firstly, we used the percentage
returns of the underlying over 30 seconds, 1 minute, 2 minutes and 5
minutes period, denoted as r, as dependent variable. The variable
OrderFlow, denoted as OF, is used as independent variable.
The tables describes summary statistics of the returns, the orderflow
and the trading volume for the four time intervals.

30 seconds time interval

Returns Orderflow Volume
Mean 0.000 0.02 75
Median 0.000 0.00 55
Standard deviation 0.0002 20.60 73.48
Skewness 0.21 0.09 3.85
Kurtosis 23.52 14.65 32.7
1𝑠𝑡 Percentile -0.000675 -58 7
5𝑡ℎ Percentile -0.000365 -31 14
25𝑡ℎ Percentile -0.000121 -9 32
75𝑡ℎ Percentile 0.000121 9 94
95𝑡ℎ Percentile 0.000363 31 204
99𝑡ℎ Percentile 0.000659 59 367

Table 3.1: Summary statistics of return, orderflow and trading volume in the 30
seconds time interval

47
 1 minute time interval
Returns Orderflow Volume

Mean 0.000 0.06 98

Median 0.000 0.00 115
Standard deviation 0.0004 29.22 111.92
Skewness 0.7 0.05 3.22
Kurtosis 24.87 10.42 22.5
1𝑠𝑡 Percentile -0.000954 -81 21
5𝑡ℎ Percentile -0.000517 -45 35
25𝑡ℎ Percentile -0.000165 -14 70
75𝑡ℎ Percentile 0.000164 16 188
95𝑡ℎ Percentile 0.000511 45 390
99𝑡ℎ Percentile 0.000943 83 649

Table 3.2: Summary statistics of return, orderflow and trading volume in the 1
minute time interval

2 minutes time interval

Returns Orderflow Volume

Mean 0.000 -0.11 302

Median 0.000 0.00 239
Standard deviation 0.0005 41.65 234
Skewness 0.8 0.11 2.79
Kurtosis 21.52 7.4 18.2
1𝑠𝑡 Percentile -0.00133 -112 54
5𝑡ℎ Percentile -0.00072 -65 82
25𝑡ℎ Percentile -0.00024 -21 152
75𝑡ℎ Percentile 0.00024 21 377
95𝑡ℎ Percentile 0.00072 65 741
99𝑡ℎ Percentile 0.00130 116 1193

Table 3.3: Summary statistics of return, orderflow and trading volume in the 2
minutes time interval

48
 5 minutes time interval
Returns Orderflow Volume

Mean 0.000 -0.31 759

Median 0.000 0.00 624
Standard deviation 0.0008 66.36 511
Skewness 0.48 0.02 2.29
Kurtosis 14.90 5.8 13.3
1𝑠𝑡 Percentile -0.00217 -181 166
5𝑡ℎ Percentile -0.00115 -105 233
25𝑡ℎ Percentile -0.00036 -36 416
75𝑡ℎ Percentile 0.00010 35 957
95𝑡ℎ Percentile 0.00034 105 1712
99𝑡ℎ Percentile 0.00213 187 2664

Table 3.4: Summary statistics of return, orderflow and trading volume in the 5
minutes time interval

The returns over short time intervals are close to zero for the sample
considered. This result means that the intraday returns does not change
substantially over time. The orderflow presents a similar feature; the
mean and the median of the buying and selling pressure are close to
zero, meaning that the market, after strong movements, comes back to
its balance. This result is supported by the activity of the HFT that often
trade in extreme situations. Moreover, the returns, orderflow and
volume have high kurtosis, meaning that there is a good number of
extreme values over the sample and consequently their distribution has
bigger tails than a normal distribution. In particular, the orderflow and
the trading volume have higher kurtosis for the smaller time intervals.
This finding highlights that the market activity has less severe
observation respect its distribution in higher time frames. Contrastly, at
lower frequencies there are more extreme values.
The Figure 3.1 represents the autocorrelation function of the orderflow
variable for the four time intervals considered in the study.

49
These time-series exhibit a particular feature: the lags-1 are positive and
significant. This result indicates that positive orderflow (buying
pressure) is followed by positive orderflow and viceversa. This
characteristic is more accentuated for the smaller time intervals. The 30
seconds, 1 minute and 2 minutes time-series have also the second and
the third lags positive and significative.
The 5 minutes time-series instead has only the first lag positive and
significative. These results explain that the orderflow has a short term
persistence, more evident as we move through smaller time intervals.

Figure 3.1: Autocorrelation function of the orderflow for the 30 seconds time intevals

50
Figure 3.2: Autocorrelation function of the orderflow for the 1 minute time intevals

Figure 3.3: Autocorrelation function of the orderflow for the 2 minutes time intevals

51
 Figure 3.4: Autocorrelation function of the orderflow for the 5 minutes time intevals

Moreover, we computed the first difference of the orderflow variable

and I found that it has a significant lag-1 negative autocorrelation for
every time intervals considered. These results are consistent with the
results by Chordia [3].

Figure 3.5 Autocorrelation function of the first-difference of

the orderflow time-series for the 30 seconds time interval Figure 3.6 Autocorrelation function of the first-difference of
the orderflow time-series for the 1 minute time interval

52
Figure 3.7 Autocorrelation function of the first-difference of Figure 3.8 Autocorrelation function of the first-difference of
the orderflow time-series for the 2 minutes time interval the orderflow time-series for the 5 minutes time interval

Furthermore, I found a positive correlation between contemporaneous

return and orderflow variable. These interesting findings indicate that
more we move through smaller time intervals, stronger the correlation
between these two variables.
The correlation for the 30 seconds, 1 minute, 2 minutes and 5 minutes
time intervals is illustrated in the next table.

30 seconds 1 minute 2 minutes 5 minutes

Correlation 0.5739 0.5513 0.5320 0.4781

Table 3.5: Correlation between orderflow and market returns

Before I move in the quantile regression, I fit a linear regression on the

returns and the orderflow

𝑟𝑡 = α + 𝛽𝑡 𝑂𝐹𝑡 + 𝑒𝑡

53
to highlight the difference between the results of the two statistical
models and the benefits given by studying the relation for different
levels of the distribution. This linear model, given the positive values of
β describes the positive relation between orderflow and returns. In
addition, also this model gives a 𝑅 2 higher as we move through smaller
time intervals.

Linear Regression
30 seconds 1 minute 2 minutes 5 minutes
Intercept -0.000985 -0.00025 -0.00047 -0.00233
Beta 0.006912 0.00665 0.00637 0.00553
𝑅2 0.329 0.304 0.283 0.229

Table 3.6: Results of linear regression

Figure 3.9: Linear regression plot Figure 3.10: Linear regression plot
between orderflow and market between orderflow and market
returns for the 30 seconds time returns for the 1 minutes time
intervals intervals

54
 Figure 3.11: Linear regression Figure 3.12: Linear regression plot
plot between orderflow and between orderflow and market returns
market returns for the 2 minute for the 5 minute time intervals
time intervals

The order flow variable, denoted as OF, is used as independent variable.

Having obtained a first image of the relation between the orderflow and
the market returns, the effect was studied introducing an other
regressor, the trading volume 𝑉𝑂𝐿 , in order to have a more complete
picture among these variables. A multiple linear regression has been
used to discover the potential explanatory power. It can be defined as:

𝑟𝑡 = α + 𝛽1 𝑂𝐹𝑡 +𝛽2 𝑉𝑂𝐿𝑡 + 𝑒𝑡

This statistical method was applied to all the four time intervals. the
coefficients found of trading volume are positive and significative

55
different from 0 in all samples. The 𝑅 2 are slightly higher than the
previous regression, as expected.

Multiple Linear Regression

30 seconds 1 minute 2 minutes 5 minutes
Intercept -0.00016 -0.00080 -0.00170 -0.00430
𝐵1 0.00069 0.00066 0.00042 0.00023
𝐵2 0.00038 0.00011 0.00007 0.00004
𝑅2 0.367 0.349 0.316 0.248

Table 3.7: Results of Multiple Linear Regression

Market returns have a clear feature that characterize them: they exhibit
heteroscedasticity, or autocorrelation in the squared residuals, meaning
that the variance is a dynamic process. To identify such charateristic, an
Engle’s Arch (Autoregressive Conditional Heteroscedasticity) test has
been used on the returns time-series of all time intervals of the sample.
Starting from the basic time-series theory, the ARCH is illustrated as
follows:

𝑦𝑡 = 𝜇𝑡 + 𝜖𝑡

where 𝜇𝑡 is the conditional mean, 𝜖𝑡 is the innovation with mean 0.

𝜖𝑡 =𝜎𝑡 𝑧𝑡

where 𝑧𝑡 is an i.i.d. process with mean 0 and variance 1 and the

innovations are uncorrelated across time.
Now define the residual series

56
𝜖𝑡 =𝑦𝑡 - 𝜇̂𝑡

To test the possibility of a serial correlation between residuals we emply

an Engle’s Arch test

𝐻𝑎 : 𝜖 2𝑡 = 𝑎0 + 𝑎1 𝜖 2𝑡−1+…+𝑎𝑚 𝜖 2𝑡−𝑚 + 𝜇𝑡

Where 𝜇𝑡 is a white noise error process. The null hypothesis is

𝐻0 : 𝑎0 = 𝑎1 = 𝑎𝑚 = 0

The null hyphotesis of the test is that there is no autocorrelation in the

residuals time-series.

Moreover, I analyzed the market returns time-series in order to detect

some arch-effect. The Table 3.8 represents, as expected, the results. For
all the time intervals the market returns time-series exhibit
heteroscedasticity and the volatility clustering is well shown in these
outcomes.

Engle’s Arch Test on Market Returns

30 seconds 1 minute 2 minutes 5 minutes
Test results Reject 𝐻0 Reject 𝐻0 Reject 𝐻0 Reject 𝐻0
P-value < .001 < .001 < .001 < .001

Table 3.8: Results of Engle’s Arch Test

57
The final empirical study of this work aims to investigate the effect of
the orderflow to different levels of market returns.
In order to do that, I conduct a set of quantile regressions to examine
the relationship between contemporaneous orderflow and returns. The
model can be expressed as:

𝑟𝑞 = 𝛽𝑞 𝑂𝐹𝑞 + 𝑒𝑞

30 seconds time intervals

Quantiles Intercept Orderflow
0.01 (0.001) -0.055 (0.000) 0.000624
0.05 (0.000) -0.025 (0.000) 0.000620
0.10 (0.000) -0.016 (0.000) 0.000672
0.25 (0.000) -0.07 (0.000) 0.000678
0.50 (0.000) 0.000 (0.000) 0.000676
0.75 (0.000) 0.007 (0.000) 0.000676
0.90 (0.000) 0.016 (0.000) 0.000667
0.95 (0.000) 0.025 (0.000) 0.000659
0.99 (0.001) 0.55 (0.000) 0.000648
Table 3.9: Results of quantile regression on 30 seconds
time interval

1 minute time intervals

Quantiles Intercept Orderflow
0.01 (0.002) -0.082 (0.000) 0.000556
0.05 (0.001) -0.040 (0.000) 0.000657
0.10 (0.000) -0.027 (0.000) 0.000659
0.25 (0.000) -0.012 (0.000) 0.000664
0.50 (0.000) 0.000 (0.000) 0.000669
0.75 (0.000) 0.012 (0.000) 0.000669
0.90 (0.000) 0.027 (0.000) 0.000684
0.95 (0.000) 0.040 (0.000) 0.000679
0.99 (0.002) 0.08 (0.000) 0.000696

Table 3.10: Results of quantile regression on 1 minute

time interval

58
 2 minutes time intervals
Quantiles Intercept Orderflow
0.01 (0.004) -0.113 (0.000) 0.000556
0.05 (0.001) -0.058 (0.000) 0.000657
0.10 (0.001) -0.039 (0.000) 0.000659
0.25 (0.000) -0.017 (0.000) 0.000664
0.50 (0.000) 0.000 (0.000) 0.000669
0.75 (0.000) 0.017 (0.000) 0.000669
0.90 (0.001) 0.038 (0.000) 0.000684
0.95 (0.001) 0.056 (0.000) 0.000679
0.99 (0.003) 0.114 (0.000) 0.000696

Table 3.11: Results of quantile regression on 2 minutes

time interval

5 minutes time intervals

Quantiles Intercept Orderflow
0.01 (0.012) -0.194 (0.000) 0.000439
0.05 (0.003) -0.092 (0.000) 0.000580
0.10 (0.001) -0.063 (0.000) 0.000548
0.25 (0.001) -0.027 (0.000) 0.000553
0.50 (0.001) 0.001 (0.000) 0.000564
0.75 (0.001) 0.028 (0.000) 0.000552
0.90 (0.001) 0.061 (0.000) 0.000560
0.95 (0.003) 0.090 (0.000) 0.000532
0.99 (0.009) 0.188 (0.000) 0.000501

Table 3.12: Results of quantile regression on 5 minutes

time interval

The Tables reports the regression results. The coefficients of the

orderflow are positive and consistent in the quantile around the mean
for all the time intervals.
In the 30 seconds and 5 minutes charts, it is possible to notice a reverse-
U values for the orderflow coefficients. This results lead us to say that

59
the orderflow has less explanatory power for the extreme values of the
return distribution.
In the 1 minute and 2 minutes time intervals there is an interesting
feature. The orderflow coefficients are higher for the return quantiles
above the mean, with maximum values in the quantiles that reflect the
most positive values of the entire distribution of returns. These finding
leads us to say that the orderflow has stronger effect in the positive
returns. These results will be exploit to build the trading strategy. In the
most negative return quantiles the orderflow effect seems disappear at
all the frequencies considered. Not only the coefficients have valuable
information. The intercepts of the model are negative in the return
quantiles below the mean and positive in the return quantiles above the
mean.
These results are present in all the time intervals of the study.
Consequently, considering only above the mean part of the return
distribution, it can be possible to support the idea that the orderflow
variable has strong effect.
We attribute this effect to the fact that if the investors are willing to buy
the Futures contract, they want to increase their long exposure in the
market causing an upturn in the prices. On the contrary, the sell of a
Futures contract can be done to hedge a portfolio with long exposure
the underlying. The effect in this case is much weaker on the price, as
shown in the results of the quantile regressions.

60
 Figure 3.13: Intercepts for the 30 Figure 3.13: Beta for the 30 seconds
seconds time interval time interval

Figure 3.14: Intercepts for the 1 minute Figure 3.15: Beta for the 1 minute time
time interval interval

61
Figure 3.16: Intercepts for the 2 Figure 3.17: Beta for the 2 minutes time
minutes time interval interval

Figure 3.18: Intercepts for the 5 Figure 3.19: Beta for the 5 minutes time
minutes time interval interval

62
3.4 Trading Strategy

The goal of the previous statistical studies was to highlight the features of both
the orderflow and the trading volume, together with their short-term
explanatory power for the analyzed returns.
This trading strategy is based on two particular aspects emerged in the
precedent paragraph: the orderflow has a positive and significant first lag in its
autocorrelation function. The quantile regression on orderflow and returns
illustrates that the effect of the orderflow is stronger for the positive quantile
of returns’ distribtuion. Moreover, with the use of the multiple regression it is
shown that trading volume has a positive relation with returns.
In the light of these outcomes, the strategy will only open long positions due to
the fact that the coefficients of orderflow for positive returns are higher,
meaning that the effects are stronger. The trading volume is included as variable
of the strategy because we choose to identify the situations when the market
activity is intense and with a positive buying pressure. The strategy will perform
only on 1 time interval because the statistical results show clearer effects than
results based on the other time intervals.
It is necessary specify the assumptions underlying:

- When the strategy opens a trade, it sends a buy market order and
there is always a counterparty on the ask side.
- There is no latency between the arrival of market data and the
execution time.
- The strategy trade only 1 contract of Future Dax (12,5 euro/tick)
- There is not spread between bid and ask
- There are no commission or other transaction costs.

63
The trigger to open a trade has been choosen considering the orderflow and
trading volume distributions in order to identify situations with strong buying
pressure and intense trading activity.
A market buy order will be sent only if the orderflow and trading volume in the
1 minute time interval are bigger than specific thresholds. The position will be
opened at the beginning of the next time interval and closed at the end.
It was conducted an analysis of 17 percentiles of both distributions, moving by
5% from 10% to 90%.
To choose the most performing parameters it was conducted an investigation
for all the combinations of percentiles.

Figure 3.20: Trading results with all the combinations of parameters

The most performing parameters are the 85𝑡ℎ percentile of the orderflow
distribution and the 85𝑡ℎ percentile of the trading volume distribution, equal
respectively to 24 and 242.

64
As expected, when the orderflow is relatively high and positive and trading
volume assumes a big value respect to its distribution, the 1 minute ahead return
is often positive.
This situation meaning that the majority of investors are willing to open long
positions driving the prices up.
This strategy yields a positive returns of 4.1% in the 62 days of sample, it has
an average win ratio of 69% and an average risk-reward ratio of 1.18.
Considering the results deriving from the quantile regression on
contemporaneuos orderflow and returns, the trading strategy is improved
introducing the trailing stop. The same parameters are used as threshold for the
trades and the open positions will not be closed at the end of the time interval,
but it will be held until the orderflow variable is positive in order to exploit the
positive momentum.
The strategy opens less trades than the strategy without trailing stop but the
win-ratio is approximately the same.
The total return increase by a 1.1% for the sample considered and the driver is
the improvement of the risk-reward ratio. Now, the strategy obtains bigger
returns than the previous backtest and the size of stop losses are the same.
The results are illustrated in the table as follow:

Trading Strategy Results

Return Win-Ratio Risk:Reward Nr. Trades
4.12% 68.5% 1.18 496
5.27% 69.7% 1.37 462

Table 3.13: Results of the trading strategy and the trading strategy improved with the trailing
stop

65
4 Conclusion

In this paper was introduced the area of empirical market microstructure.

Starting from the description of the functioning of the limit orderbook, it was
done a literature review, were described some high frequency trading strategy
and showed some inefficiencies. The empirical study aims to show that using
the orderflow and the trading volume, a profitable trading strategy can be
developed. The multiple regression highlights the significativity of these
variables in the explanation of market returns. The market returns, as expected,
present some volatility clustering. This information can be derived with a
statistical method that tests the presence of arch effects in the residuals time-
series. The quantile regression shows a particular feature. The orderflow has a
stronger effect on the positive quantile of the returns’ distribution. Starting
from these results, a trading strategy that is able to produce consistent returns
was developed exploiting the outcomes of the previous statistical relations.
We can conclude that it is possible to developed an alpha generating strategy
exploiting the order flow that comes to the market.

66
Bibliography

[1] Irene Aldridge. High-Frequency Trading: A Practical Guide to Algorithmic

Strategies and Trading Systems. John Wiley & Sons, Inc, 2010.

[2] Mario Bellia, Kim Christensen, Aleksey Kolokolov, Loriana Pelizzon, and
Roberto Renò. High-frequency trading during flash crashes: Walk of fame or
hall of shame? Working Paper, 2018.

[3] Tarun Chordia, Richard Roll, and Avanidhar Subrahmanyam. Order

Imbalance, liquidity, and market returns. Journal of Financial Economics, 65:111-
130, 2002.

[4] Tarun Chordia and Avanidhar Subrahmanyam. Order imbalance and

individual stock returns: Theory and evidence. Journal of Financial Economics,
72:485-518, 2004.

[5] Rama Cont, Arseniy Kukanov, and Sasha Stoikov. The price impact of
order book events. Journal of Financial Econometrics, 12(1):47-88, 2014.

[6] Joel Hasbrouck and Gideon Saar. Low-Latency Trading. Journal of Financial
Markets, 16(4), 2013.

[7] Han-Ching Huang, Yong-Chern Su, and Yi-Chun Liu. The

performance of imbalance-based trading strategy on
tender offer announcement day. Investment Management and Financial Innovations,
11(2):38-46, 2014.

67
[7] Van Kervel, V. and Menkveld, A. J. High-frequency trading around large
institutional orders. SSRN Working Paper, 2016

[8] Charles M.C. Lee and Mark J. Ready. Inferring trade direction from
intraday data. The Journal of Finance, 46:733-747, 1991.

[9] Jaeram Lee, Geul Lee, and Doojin Ryu. The difference in the intraday
return-volume relationships of spot and futures: a quantile regression
approach. Economics: The Open-Access, Open-Assessment E-Journal, 13:1–37, 2019.

[10] Rahul Ravi and Yuqing Sha. Autocorrelated Order-Imbalance and Price Mo-
mentum in the Stock Market. International Journal of Economics and Finance,
6(10):39-54, 2014.

[11] Koenker, Roger, and Gilbert Bassett. "Regression

Quantiles." Econometrica 46, no. 1 (1978): 33-50. Accessed February 6, 2020.

68
 SUMMARY

The technological develpment and the innovations in the financial

markets made possible to collect and study the information descending
from the market activity in the microstructure.
The information obtained can be extremely useful in order to explain
the price movements in the short-term.
The first empirical studies highlight the relevance of trading volume and
successively the attention shifted to the orderflow. The orderflow is a
more advanced concept respect to the trading volume because it
attributes an additional feature to each trade executed making possible
to identify where a trade is buyer-initiated or seller-initiated.
Consequently, an investor able to get and exploit this information may
have an advantage respect the other partecipants who do not, at least in
the short-term. The partecipants who, for definition, exploit and trade
relying on these signals are the High Frequency Trader (called ‘HFT’).
They trade at extreme speed against the other investors and aim to gain
small profits countless times.
In order to support the thesis that the market activity is a driver of
market returns, we investigated with statistical tools the relationships
between returns, orderflow and trading volume for four time intervals
(30 seconds, 1 minute, 2 minutes, 5 minutes) and we were able to
develop a trading strategy based on signals deriving from these
variables.
In the first chapter it is illustrated the market framework, the evolution
during this last decades and the way how investors trade.
The second chapter describes with the market activity from a more
technical view. There are described the methods use to assign the sign
to every trade, the first model that riassumes the relations between

69
trading volume, returns with information flow that comes to the
market. In this part some studies has been mentioned in order to
understand how previous authors deal with these concepts and what
they found. Moreover, we describes the most common high frequency
trading strategies and analyzed some market inefficiencies.
The third chapter is the main part of this thesis. We studied the
distributions of the three variables taken into account, the
autocorrelation function of the orderflow and the correlation between
returns and orderflow. The first tool used to get a description of these
relationships is a multiple linear regression that shows that the effects
of the independent variables (orderflow and trading volume) are
significant. To examine more deeply the relation between returns and
orderflow we availed of quantile regression and found that positive
returns are more influenced by orderflow than negative returns.
In the light of these findings a profitable trading strategy has been
developed exploiting the orderflow and trading volume statistical
features and the relationship with returns. This strategy considers these
variable and takes in account the returns of one period ahead.
After having shown that choosing specific thresholds of orderflow and
volume can be profitable in the sample considered, another rule is
introduced in the strategy deriving from the relation between orderflow
and returns. Until the orderflow variable remains positive the trade will
not be closed in order to exploiting the buying momentum until it ends.
This improves the results of the strategy and shows how the
information deriving from orderflow and trading volume can be used
in a proper way.
Starting from the definition of the limit order book It is a centralized
database of pending limit orders for a specified price or better and with
a given size. It is a transparent system for the market partecipants and

70
the matching algorithm for limit orders with market orders works with
a price time priority algorithm: it works with matching priority for price,
time and lastly visibility.
A buy limit order with the highest price represents that buyer is willing
to pay more than the other traders and he gets the priority of execution
when a sell market order will arrive. Consequently, a market order is
executed on the best proposal, best bid or best ask (in the US market
the first levels are guaranteed by NBBO, National Best Bid Offer).
After that, it follows the FIFO principle: First in First out.
The limit order that is submitted earlier will be executed before the later
arrival.
Therefore the queue of limit orders follows this chronological rule.
In most markets, there is the possibility to put hidden orders, that
stands on the book without being displayed. They are the last limit
orders to be executed because the precedence is given to the visible
orders.
The ordebook events can be mainly identified in three types:
submission of limit orders, cancellation of limit order, execution of
market orders.
The empirical studies in the financial literature focus their attention on
studying the price impact of the orderbook events. The outstanding
limit orders, called Market Depth, concept related to the market
liquidity, is associated with the magnitude of price changes. In a thin
orderbook, where the market depth is low given by the small size and
the low number of limit orders, it is easier to cause a bigger price change
given to the fact that there is less need of market orders, respect a thick
orderbook, to buy or sell all the first level proposals and than to move
the price.

71
The price dynamics are also conditioned by the arrival, the size and the
direction of the market orders. The funcioning can be summarized by
the following three images. (Figure 2.1, Figure 2.2, Figure 2.3)

72
 Before, starting with the empirical studies with the regressions, I analyze
the statistical features of the three variables considered for the work:
market return, orderflow and trading volume. The feature are illustrated
in the following tables. (Table 2.1, Table 2.2, Table 2.3, Table 2.4)

30 seconds time interval

Returns Orderflow Volume

Mean 0.000 0.02 75

Median 0.000 0.00 55
Standard deviation 0.0002 20.60 73.48
Skewness 0.21 0.09 3.85
Kurtosis 23.52 14.65 32.7
1𝑠𝑡 Percentile -0.000675 -58 7
5𝑡ℎ Percentile -0.000365 -31 14
25𝑡ℎ Percentile -0.000121 -9 32
75𝑡ℎ Percentile 0.000121 9 94
95𝑡ℎ Percentile 0.000363 31 204
99𝑡ℎ Percentile 0.000659 59 367

1 minute time interval

Returns Orderflow Volume

Mean 0.000 0.06 98

Median 0.000 0.00 115
Standard deviation 0.0004 29.22 111.92
Skewness 0.7 0.05 3.22
Kurtosis 24.87 10.42 22.5
1𝑠𝑡 Percentile -0.000954 -81 21
5𝑡ℎ Percentile -0.000517 -45 35
25𝑡ℎ Percentile -0.000165 -14 70
75𝑡ℎ Percentile 0.000164 16 188
95𝑡ℎ Percentile 0.000511 45 390
99𝑡ℎ Percentile 0.000943 83 649

73
 2 minutes time interval
Returns Orderflow Volume

Mean 0.000 -0.11 302

Median 0.000 0.00 239
Standard deviation 0.0005 41.65 234
Skewness 0.8 0.11 2.79
Kurtosis 21.52 7.4 18.2
1𝑠𝑡 Percentile -0.00133 -112 54
5𝑡ℎ Percentile -0.00072 -65 82
25𝑡ℎ Percentile -0.00024 -21 152
75𝑡ℎ Percentile 0.00024 21 377
95𝑡ℎ Percentile 0.00072 65 741
99𝑡ℎ Percentile 0.00130 116 1193

5 minutes time interval

Returns Orderflow Volume

Mean 0.000 -0.31 759

Median 0.000 0.00 624
Standard deviation 0.0008 66.36 511
Skewness 0.48 0.02 2.29
Kurtosis 14.90 5.8 13.3
1𝑠𝑡 Percentile -0.00217 -181 166
5𝑡ℎ Percentile -0.00115 -105 233
25𝑡ℎ Percentile -0.00036 -36 416
75𝑡ℎ Percentile 0.00010 35 957
95𝑡ℎ Percentile 0.00034 105 1712
99𝑡ℎ Percentile 0.00213 187 2664

After having described the characteristics of these variables we exploit

a multiple linear regression in order to found the relationships between
trading volume, orderflow with the market return.
The outcomes show a positive relation meaning that these variables
have an explanatory power on the retunrs.

74
 Multiple Linear Regression
30 seconds 1 minute 2 minutes 5 minutes
Intercept -0.00016 -0.00080 -0.00170 -0.00430
𝐵1 0.00069 0.00066 0.00042 0.00023
𝐵2 0.00038 0.00011 0.00007 0.00004
𝑅2 0.367 0.349 0.316 0.248

Table 3.7 shows another interesting finding is that the 𝑅 2 of the

regression is bigger if we move through smaller time intervals.
After having ensure the positive relation between these variables we
investigated these effects with a more deep examination exploiting the
quantile regression. The results show a clear situation. The coefficients
for the orderflow are bigger in the positive percentile of the returns
distributions. We attribute this effect to the fact that if the investors are
willing to buy the Futures contract, they want to increase their long
exposure in the market causing an upturn in the prices. On the contrary,
the sell of a Futures contract can be done to hedge a portfolio with long
exposure on the underlying. The effect in this case is much weaker on
the price, as shown in the results of quantile regression.
The table 3.9, table 3.10, table 3.11 and table 3.12 summary the
outcomes of the regression.

75
 30 seconds time intervals
Quantiles Intercept Orderflow
0.01 (0.001) -0.055 (0.000) 0.000624
0.05 (0.000) -0.025 (0.000) 0.000620
0.10 (0.000) -0.016 (0.000) 0.000672
0.25 (0.000) -0.07 (0.000) 0.000678
0.50 (0.000) 0.000 (0.000) 0.000676
0.75 (0.000) 0.007 (0.000) 0.000676
0.90 (0.000) 0.016 (0.000) 0.000667
0.95 (0.000) 0.025 (0.000) 0.000659
0.99 (0.001) 0.55 (0.000) 0.000648

1 minute time intervals

Quantiles Intercept Orderflow
0.01 (0.002) -0.082 (0.000) 0.000556
0.05 (0.001) -0.040 (0.000) 0.000657
0.10 (0.000) -0.027 (0.000) 0.000659
0.25 (0.000) -0.012 (0.000) 0.000664
0.50 (0.000) 0.000 (0.000) 0.000669
0.75 (0.000) 0.012 (0.000) 0.000669
0.90 (0.000) 0.027 (0.000) 0.000684
0.95 (0.000) 0.040 (0.000) 0.000679
0.99 (0.002) 0.08 (0.000) 0.000696

2 minutes time intervals

Quantiles Intercept Orderflow
0.01 (0.004) -0.113 (0.000) 0.000556
0.05 (0.001) -0.058 (0.000) 0.000657
0.10 (0.001) -0.039 (0.000) 0.000659
0.25 (0.000) -0.017 (0.000) 0.000664
0.50 (0.000) 0.000 (0.000) 0.000669
0.75 (0.000) 0.017 (0.000) 0.000669
0.90 (0.001) 0.038 (0.000) 0.000684
0.95 (0.001) 0.056 (0.000) 0.000679
0.99 (0.003) 0.114 (0.000) 0.000696

76
 5 minutes time intervals
Quantiles Intercept Orderflow
0.01 (0.012) -0.194 (0.000) 0.000439
0.05 (0.003) -0.092 (0.000) 0.000580
0.10 (0.001) -0.063 (0.000) 0.000548
0.25 (0.001) -0.027 (0.000) 0.000553
0.50 (0.001) 0.001 (0.000) 0.000564
0.75 (0.001) 0.028 (0.000) 0.000552
0.90 (0.001) 0.061 (0.000) 0.000560
0.95 (0.003) 0.090 (0.000) 0.000532
0.99 (0.009) 0.188 (0.000) 0.000501

The goal of the previous statistical studies was to highlight the features of both
the orderflow and the trading volume, together with their short-term
explanatory power for the analyzed returns.
This trading strategy is based on two particular aspects emerged in the
precedent paragraph: the orderflow has a positive and significant first lag in its
autocorrelation function. The quantile regression on orderflow and returns
illustrates that the effect of the orderflow is stronger for the positive quantile
of returns’ distribtuion. Moreover, with the use of the multiple regression it is
shown that trading volume has a positive relation with returns.
In the light of these outcomes, the strategy will only open long positions due to
the fact that the coefficients of orderflow for positive returns are higher,
meaning that the effects are stronger. The trading volume is included as variable
of the strategy because we choose to identify the situations when the market
activity is intense and with a positive buying pressure. The strategy will perform
only on 1 time intervals because the statistical results show clearer effects than
results based on the other time intervals. There are some underlying
assumptions to take in account for the evaluation of the strategy: no transaction
costs are taken in account, the bid-ask spread is always 1 tick and there is no
delay in the data received.

77
The trigger to open a trade has been choosen considering the orderflow and
trading volume distributions in order to identify situations with strong buying
pressure and intense trading activity.
A market buy order will be sent only if the orderflow and trading volume in the
1 minute time interval are bigger than specific thresholds. The position will be
opened at the beginning of the next time interval and closed at the end.
It was conducted an analysis of 17 percentiles of both distributions, moving by
5% from 10% to 90%.
To choose the most performing parameters it was conducted an investigation
for all the combinations of percentiles.
The most performing parameters are the 85𝑡ℎ percentile of the orderflow
distribution and the 85𝑡ℎ percentile of the trading volume distribution, equal
respectively to 24 and 242.
As expected, when the orderflow is relatively high and positive and trading
volume assumes a big value respect to its distribution, the 1 minute ahead return
is often positive.
This situation meaning that the majority of investors are willing to open long
positions driving the prices up.
This strategy yields a positive returns of 4.1% in the 62 days of sample, it has
an average win ratio of 69% and an average risk-reward ratio of 1.18.
Considering the results deriving from the quantile regression on
contemporaneuos orderflow and returns, the trading strategy is improved
introducing the trailing stop. The same parameters are used as threshold for the
trades and the open positions will not be closed at the end of the time interval,
but it will be held until the orderflow variable is positive in order to exploit the
positive momentum.
The strategy opens less trades than the strategy without trailing stop but the
win-ratio is approximately the same.

78
The total return increase by a 1.1% for the sample considered and the driver is
the improvement of the risk-reward ratio. Now, the strategy obtains bigger
returns than the previous backtest and the size of stop losses are the same.
The results are illustrated in the table 3.13 and in the figure 3.16 as follow:

Trading Strategy Results

Return Win-Ratio Risk:Reward Nr. Trades
4.12% 68.5% 1.18 496
5.27% 69.7% 1.37 462

79