Examiners’ commentaries 2023

Examiners’ commentaries 2023

FN3142 Quantitative Finance

Important note

This commentary reflects the examination and assessment arrangements for this
course in the academic year 2022–2023. The format and structure of the
examination may change in future years, and any such changes will be publicised
on the virtual learning environment (VLE).
Information about the subject guide and the Essential reading references
Unless otherwise stated, all cross-references will be to the latest version of the
subject guide (2015). You should always attempt to use the most recent edition
of any Essential reading textbook, even if the commentary and/or online reading
list and/or subject guide refer to an earlier edition. If different editions of Essential
reading are listed, please check the VLE for reading supplements – if none are
available, please use the contents list and index of the new edition to find the
relevant section.

General remarks
Aims of the course
This course is aimed at candidates interested in obtaining a thorough grounding in
market finance and related empirical methods. It introduces the econometric
techniques, such as time-series analysis, required to analyse empirical issues in
finance. It provides applications in asset pricing, investments, risk analysis and
management, market microstructure and return forecasting. This course is
quantitative by nature. It aims, however, to investigate practical issues in the
forecasting of key financial market variables and makes use of a number of real-world
data sets and examples.

Learning outcomes
At the end of this course and having completed the Essential reading and Activities,
you should be able to:
• have mastered the econometric techniques required in order to analyse issues in
asset pricing and market finance,
• be familiar with recent empirical findings based on financial econometric models,
• have gained valuable insights into the functioning of financial markets,

1
FN3142 Quantitative Finance

• understand some of the practical issues in the forecasting of key financial market
variables, such as asset prices, risk and dependence.

Reading advice
The subject guide is designed to complement, not replace, the listed readings for
each chapter. Each chapter in the subject guide builds on the earlier chapters, as is
often the case with quantitative subjects. Chapters should therefore be studied in the
order in which they appear. Essential readings for this course come from:
• Christoffersen, P.F. Elements of Financial Risk Management. (Academic Press,
Oxford, 2011) second edition [ISBN 9780123744487].
• Diebold, F.X. Elements of Forecasting. (Thomson South-Western, Canada,
2006) fourth edition [ISBN 9780324323597].
Further readings:
• Bodie, Z., A. Kane and A.J. Marcus Investments. (McGraw-Hill Irwin, London,
2008) eighth edition [ISBN 9780071278287] and (2013) [ISBN 9780077861674].
• Brooks, C. Introductory Econometrics for Finance. (Cambridge University Press,
Cambridge, 2008) second edition [ISBN 9780521694681] and third edition [ISBN
9781107661455].
• Campbell, J.Y., A.W. Lo and A.C. Mackinlay. The Econometrics of Financial
Markets. (Princeton University Press, Princeton, NJ, 1997) [ISBN
9780691043012].
• Clements, M.P. Evaluating Econometric Forecasts of Economic and Financial
Variables. (Palgrave Texts in Econometrics, England, 2005) [ISBN
9781403941572].
• Elton, E.J., M.J. Gruber, S.J. Brown and W.N. Goetzmann. Modern Portfolio
Theory and Investment Analysis. (John Wiley & Sons, New York, 2010) eighth
edition [ISBN 9780470505847] and ninth edition [ISBN 9781118469941].
• Granger, C.W.J. and A. Timmerman. Efficient Market Hypothesis and
Forecasting, International Journal of Forecasting, 20(1) 2004.
• Hull, J.C. Options, Futures and Other Derivatives. (Pearson, 2011) eighth edition
[ISBN 9780273759072].
• McDonald, R.L. Derivatives Markets. (Pearson, 2012) third edition [ISBN
9780321847829].
• Taylor, S.J. Asset Price Dynamics, Volatility and Prediction. (Princeton University
Press, Oxford, 2007) [ISBN 9780691134796].
• Tsay, R.S. Analysis of Financial Time Series. (John Wiley & Sons, New York,
2010) third edition [ISBN 9780470414354].

Studying advice
In addition to reading, students are also expected to work through the learning
activities and sample examination questions provided in the subject guide. If they find

2
 Examiners’ commentaries 2023

it difficult to answer a given learning activity, they should go through the readings to
that learning activity again with a focus on resolving the issues at stake. It is important
to master the econometric techniques covered in the first part of the course before
moving onto more difficult topics.

Format of the examination

This year the format of the examination has been identical to those of previous years,
i.e., candidates had to answer three out of four questions. As it has been standard
since 2020, students completed the exam online, over the course of 4 hours (there
was an expected/recommended time of 3 hours). This way students had access to
all usual programs, their subject guide, and essentially any material available on the
internet, and in theory also offered an uncontrolled communication among students,
too. This was in stark contrast with previous formats. In previous years, a calculator
was allowed to be used when answering questions; otherwise no additional help was
allowed. While there were two Zones, A and B, most students ended up writing the
Zone B version (324 students vs 5 for Zone A).

What the examiners are looking for

In a good answer to a quantitative question, students must provide rigorous
derivations. Some quantitative questions may furthermore ask for a numerical
problem to be solved. With numerical questions, it is important that answers and
steps are carefully and clearly explained. Partial credit cannot be awarded if the final
numbers presented are wrong through errors of omission, calculation, etc., unless
the workings are shown. In a good answer to a qualitative question, students are
expected to produce an answer which presents appropriate concepts and empirical
evidence. They are furthermore expected to get their points across in a direct,
structured, and concise fashion.

Due to the special environment, an important aspect of what examiners paid attention
to was that students had to answer the questions concisely and using their own
words. Having access to all the material mentioned above, many students ended up
submitting answers that were very close to copy-pasting sentences, even
paragraphs, from the study guide, using the same structure as specific subsections,
etc. These answers were heavily discounted. Further, most students (probably) spent
more time on the exam than the suggested 3 hours, and ended up providing
unnecessarily long answers to essay questions. These were again heavily discounted
by the examiners.

Specific comments on the exam

In total we received 329 scripts; 5 students wrote Zone A and 324 wrote Zone B.
Students on average performed well on the exam, obtaining averages of 56.6 on
Zone A and 61.62 on Zone B; both higher that last year’s average score of 55.12 and
comparable to the previous year's average of 59.46. The standard deviations in the
two zones were 9.97 and 14.78, respectively; comparable to last year’s standard
deviation of 12.34. The slightly higher average is due to this year’s exam featuring an
almost identical number of 2nd grades as last year, but a significantly higher proportion

3
FN3142 Quantitative Finance

of 1st grades, meaning the left tail of the distribution is less pronounced than before,
and closer to the previous years’ distributions.

In Zone A, all students attempted Questions 1 and 2, 80% of them attempted

Question 3, and 20% Question 4. The average performance on Questions 1 and 2
was 67%, whereas Questions 3 and 4 proved to be harder, with averages of 30%
and 52%, respectively. In Zone B, the number of students attempting Questions 1-4
were 98.7%, 45.3%, 82.4%, and 69.4%, respectively, and the average performances
were 68%, 62%, 58% and 56%, respectively. The specific comments are discussed
separately for the individual exercises; see also more comments in the suggested
solutions file.

Examination revision strategy

Many candidates are disappointed to find that their examination performance is

poorer than they expected. This can be due to a number of different reasons and
the Examiners’ commentaries suggest ways of addressing common problems
and improving your performance. We want to draw your attention to one particular
failing – ‘question spotting’, that is, confining your examination preparation to a
few question topics which have come up in past papers for the course. This can
have very serious consequences.
We recognise that candidates may not cover all topics in the syllabus in the same
depth, but you need to be aware that examiners are free to set questions on any
aspect of the syllabus. This means that you need to study enough of the syllabus
to enable you to answer the required number of examination questions.
The syllabus can be found in the Course information sheet in the section of the
VLE dedicated to this course. You should read the syllabus very carefully and
ensure that you cover sufficient material in preparation for the examination.
Examiners will vary the topics and questions from year to year and may well set
questions that have not appeared in past papers – every topic on the syllabus is
a legitimate examination target. So although past papers can be helpful in
revision, you cannot assume that topics or specific questions that have come up
in past examinations will occur again.
If you rely on a question spotting strategy, it is likely you will find yourself
in difficulties when you sit the examination paper. We strongly advise you
not to adopt this strategy.

4
 Examiners’ commentaries 2023

Examiners’ commentaries 2023

FN3142 Quantitative Finance

Important note
This commentary reflects the examination and assessment arrangements for this
course in the academic year 2022–2023. The format and structure of the
examination may change in future years, and any such changes will be publicised
on the virtual learning environment (VLE).
Information about the subject guide and the Essential reading references
Unless otherwise stated, all cross-references will be to the latest version of the
subject guide (2015). You should always attempt to use the most recent edition
of any Essential reading textbook, even if the commentary and/or online reading
list and/or subject guide refer to an earlier edition. If different editions of Essential
reading are listed, please check the VLE for reading supplements – if none are
available, please use the contents list and index of the new edition to find the
relevant section.

Comments on specific questions

The assessment is an open-book take-home online assessment within a 4-hour
window. The requirements for this assessment remain the same as the closed
book exam, with an expected time/effort of 3 hours.
Candidates should answer THREE of the following FOUR questions.

Zone A

Question 1
Question text: see the exam questions & solutions file.
Reading for this question
Subject guide, Chapters 7 and 2.
Approaching the question
A question on market efficiency: asking for various definitions of market efficiency
and the classification of different types of information, and various applications of the
theory. Students performed well on this exercise – perhaps not surprisingly, as
market efficiency is one of the most emphasized component of the course material.
Again, the challenge in these largely essay-like questions was to explain things using
one’s own words. Therefore, it is advised that after reading the material of the subject
guide and performing exercises, students should pay attention to how they would
demonstrate their understanding in a structured and concise manner. Part c) and d)
tended to confuse students, so it is advised that they pay attention to the distinction
between efficiency and random walks. Finally, Part f) of the question asked for a
simple calculation of volatility of a portfolio – this must have surprised students

5
FN3142 Quantitative Finance

because the average performance was below 75% on this subquestion despite the
simplicity of the question.

Question 2
Question text: see the exam questions & solutions file.
Reading for this question
Subject guide, Chapters 1-3.
Approaching the question
A question to test sudents’ understanding of conditional and unconditional (non-
normal) distributions, and the relationship between covariance and independence. In
general students had no problems with pointing out that independence implies zero
covariance but not the other way around. However, formal arguments were often
flawed. Typically students had serious problems with the more abstract parts (e and
f), although this could have been the result of them leaving Question 2 to be the last
one and then running out of time.

Question 3
Question text: see the exam questions & solutions file.
Reading for this question
Subject guide, Chapters 8-9.
Approaching the question
A question with on the most important properties of price/return volatility, namely
volatility clustering, asymmetry, stationarity, and how these properties can be
modelled – with a small twist of asking students to think about how GARCH and
EWMA models can be applied for covariance modelling. There were subquestions
about technicalities and the understanding of how these models work, too. The
specific calculations were rather simple, following those in the subject guide, but
students tended not to pay attention to mathematical details, and thus ended up with
lower scores than they might have expected. The question also included short essay-
like subparts, which asked students to explain things using their own words – the
average performance on these subquestions was rather poor, suggesting that
students did not have a good understanding of the material and were relying more on
the provided texts. Finally, part (f) asked students to apply their knowledge in a
specific volatility model that the subject guide does not emphasize much – this again
showed lack of deep understanding and being surprised by unseen cases where one
needs to apply their knowledge.

Question 4
Question text: see the exam questions & solutions file.
Reading for this question
Subject guide, Chapters 13-14.
Approaching the question
A comprehensive question on Value-at-Risk: definition; calculations for individual
assets vs portfolios and different horizons; and the general understanding of risk
measures. The only student attempting the question had incorrect calculations
already for individual assets, and combining them to portfolios turned out to be even

6
 Examiners’ commentaries 2023

more challenging. It is important to highlight though that in last year’s exam a lot more
students performed badly on a very similar exercise, as they tended to use ad hoc
techniques as opposed to doing it in a more structural way. Students are encouraged
to spend more time on this material to understand the basic idea and techniques for
evaluating and comparing VaR and other risk measures.

Zone B
Question 1
Question text: see the exam questions & solutions file.
Reading for this question
Subject guide, Chapters 3-6.
Approaching the question
This is a standard exercise to test students’ understanding of basic concepts such as
expectations, covariances, variances, and basic properties like (covariance)
stationary, and apply them to time series. Students must demonstrate their ability to
work with stochastic processes, to identify stochastic and deterministic components
and apply their properties, to comment on autocovariance and autocorrelation, and
how these refer to forecasting future realizations of random variables and processes.
Students should study the examples provided in those chapters, especially in the end
of chapters, in details, to fully master the techniques. Typical mistakes were:
incomplete definition of covariance stationarity; not arguing well enough why the
autocovariance function is zero for all lags above 1; making statements like ‘if Y is
covariance stationary, it can be represented by an MA(1) process’, which surely does
not hold if autocovariance and autocorrelation are non-zero at a lag above 1; not
doing a full calculation in part d.

Question 2
Question text: see the exam questions & solutions file.
Reading for this question
Subject guide, Chapters 1-2.
Approaching the question
This is a standard exercise to test students’ understanding of basic concepts such as
conditional and unconditional distributions, independence, and first and second
moments. In particular, students had to demonstrate their ability to work with
conditional and unconditional normal distributions, and think about how distributions
behave when we have a continuum of random variables with conditionally
independent errors, to identify stochastic and deterministic components. Students
should study the examples provided in those chapters, especially in the end of
chapters, in details, to fully master the techniques. Standard mistakes were providing
incomplete arguments for part d) (e.g. stating things like ‘F must be equal to N^{-1}(1-
s)’), failing to calculate how these formulas behave in the limits in part e), and not
providing economic intuition.

Question 3
Question text: see the exam questions & solutions file.

7
FN3142 Quantitative Finance

Reading for this question

Subject guide, Chapters 12, 8-9, and 16, respectively.
Approaching the question
A question with two distinct parts. Subquestions a) and b) were about forecast
comparisons and the Diebold-Mariano test. The question included a short essay-like
subpart, which asked students to explain things using their own words, and then
asked them to perform an actual test with one part of the calculation being a little bit
more involved than previous practice exercises. The average performance on this
part was around 65%, but with a high standard deviation, suggesting that students
either had a good understanding of the material, or they were relying more on (i.e.
copy-pasting) the provided texts and solutions, leading to incorrect answers and
showing lack of deep understanding and being surprised by unseen cases.

Subquestions c) to f) tested students’ knowledge on the most important properties of

price/return volatility, namely volatility clustering, asymmetry, stationarity, and how
these properties can be modelled. There were subquestions about technicalities and
the understanding of how these models work, too. The specific calculations were
rather simple, following those in the subject guide, but students tended not to pay
attention to mathematical details, and thus ended up with lower scores than they
might have expected. These were particularly important for the exam, since having
an online open-book exam one can easily copy formulas and calculations from the
subject guide and other online and offline materials. Finally, many students had no
recollection of how trade duration can be modelled using GARCH-like methods,
suggesting a lack of high-level overview of the material of the course.

Question 4
Question text: see the exam questions & solutions file.
Reading for this question
Subject guide, Chapters 1-3 and 15
Approaching the question
Again a question with 2 distinct subparts. Parts a) and b) tested sudents’
understanding of conditional and unconditional (non-normal) distributions and the
relationship between covariance and independence. In general students had no
problems with pointing out that independence implies zero covariance but not the
other way around. However, the formal arguments were often flawed, and students
got confused between the 2 coin toss outcomes and variables X and Y.
Parts c)-e) tested students’ understanding of diurnality and asked them about
modelling approaches to this phenomenon. Again, providing a neither too short nor
too detailed summary of models using students’ own words as opposed to copying
from the subject guide seemed like a challenge to many. Finally, in part e), typical
problems were students not being able to solve for (local) minima of a cubic function,
and ignoring the fact that one coefficient was estimated to be insignificant.

8
 I. ZONE A

Question 1

(a) Roberts (1967) defines three types of information sets available to investors for (i) weak
form efficiency; (ii) semi-strong form efficiency; and (iii) strong form efficiency. Explain what
each of these mean.

[15 marks]

(b) To which information set, if any, do the following variables belong? Explain.

(i) Next year’s merger plans just approved by the company’s board of directors (and not
announced yet).

(ii) The inflation rate in the last quarter.

(iii) The price today of a call option written on a corporate bond with a 1-month maturity.

(iv) The Bank of England’s GDP growth forecast for next year.

[20 marks]

(c) Denote a series of daily closing prices for AMD as S = S1 , S2 , S3 , ..., Sn with correspond-
ing gross returns given by R = SS21 , SS23 , ..., SSn−1
n
. Does the series S have positive, negative,
or zero serial correlation according to (i) the random walk hypothesis; and (ii) the efficient
market hypothesis? Explain.

[15 marks]

(d) Does the series R, defined in part (c), have positive, negative, or zero serial correla-
tion according to (i) the random walk hypothesis; and (ii) the efficient market hypothesis?
Explain.

[15 marks]

(e) Stock prices tend to show consistently negative abnormal returns before Federal Open
Market Committee (FOMC) meetings, when the committee is expected to increase interest
rates, and they continue to decrease for a few hours after the announcement. Is this a
violation of the efficient market hypothesis? Explain.

[15 marks]

Page 1 of 36
(f ) Suppose that the daily volatility of the FTSE 100 stock index (measured in pounds
sterling) is 1.8% and the daily volatility of the dollar–sterling exchange rate is 0.9%. Suppose
further that the correlation between the FTSE 100 and the dollar–sterling exchange rate is
0.4. What is the volatility of the FTSE 100 when it is translated to U.S. dollars? Assume
that the dollar–sterling exchange rate is expressed as the number of U.S. dollars per pound
sterling.
Hint: When Z = XY , the percentage daily change in Z is approximately equal to the per-
centage daily change in X plus the percentage daily change in Y .

[20 marks]

Page 2 of 36
Solutions.

(a) Roberts (1967) defined three types of information set available to investors:

1. Weak Form Efficiency: The information set includes only the history of the prices or
returns.

2. Semi-Strong Form Efficiency: The information set includes all information known to
all market participants, i.e. it includes all publicly available information.

3. Strong Form Efficiency: The information set includes all information known to any
market participant, including all public and private information.

(b)

(i) Next year’s merger plans just approved by the company’s board of directors (and not
announced yet). Strong.

(ii) The inflation rate in the last quarter. Semistrong (if arguing for public non-price
information) or weak (TIPS price).

(iii) The price today of a call option written on a corporate bond with a 1-month maturity.
Weak (if publicly traded) or strong (if coming from bilateral trading).

(iv) The Bank of England’s GDP growth forecast for next year. Semistrong (if it has just
been announced) or strong (private information) otherwise.

(c) and (d) Students are expected to make a distinction between the random walk hypoth-
esis and the efficient market hypothesis. The random walk walk hypothesis says that log
prices are given by
Pt+1 = µ + Pt + εt+1 ,

where εt is a white noise process. This says that our best guess of tomorrow’s price is todays
price (plus a constant growth rate). Note: we include a constant growth rate term since
investors must require a positive positive return for holding stocks.

The efficient market hypothesis, on the other hand, say that the growth rate in the
price index, µ, can depend on any state variables that affect agents rational demand for risk
compensation. Thus, the efficient market hypothesis can be written as

Pt+1 = µ(Xt ) + Pt + εt+1 .

Page 3 of 36
So, the answer to (c) is that under both the efficient markets and the random walk hypoth-
esis, prices are positively serially correlated.

Answer to (d): Under the random walk hypothesis, returns have zero serial correlation

Pt+1 − Pt = Rt+1 = µ + εt+1 = constant + noise.

Under the efficient market hypothesis returns can have positive, negative, or zero serial
correlation depending on variation in µ(Xt )

Pt+1 − Pt = Rt+1 = µ(Xt ) + εt+1 = time-varying term + noise.

(e) The question for market efficiency is whether investors can earn abnormal risk-adjusted
profits. If the stock price decline occurs when only insiders are aware of the coming interest
rate hike, then it is a violation of strong-form, but not semistrong form, efficiency. If the
public already knows of the increase, then it is a violation of semistrong- form efficiency,
because the price should have already adjusted earlier. The continuing decrease after the
stock price means that markets have not fully adjusted to a public information, i.e., it is a
violation of semistrong-form efficiency.
Note that above the assumption was that (i) there is no uncertainty about the size of the
rate hike, and (ii) there is no Fed information effect, i.e., the rate decision does not contain
information about the future level of stock dividends, only about the relevant discount rates.

(f ) This is simply the definition of variance of the sum of two random components. Namely,
we can write

V ar[X + Y ] = V ar[X] + V ar[Y ] + 2Cov[X, Y ]

p p
= V ar[X] + V ar[Y ] + 2Corr[X, Y ] V ar[X] V ar[Y ]
= 0.0182 + 0.0092 + 2 × 0.4 × 0.018 × 0.009 = 0.0005346,
√
and from here the volatility we are looking for is 0.0005346 = 0.0231 = 2.31%. This is
the volatility of the FTSE expressed in dollars. Note that it is greater than the volatility of
the FTSE expressed in sterling. This is the impact of the positive correlation. When the
FTSE increases, the value of sterling measured in dollars also tends to increase. This creates
an even bigger increase in the value of FTSE measured in dollars. Similarly, when FTSE
decreases, the value of sterling measured in dollars also tends to decrease, creating an even
bigger decrease in the value of FTSE measured in dollars.

Page 4 of 36
Question 2

Consider two random variables, V1 and V2 . Suppose that there are three equally likely values
for V1 : −1, 0, and 1. Suppose also that if V1 = −1 or V1 = 1, then V2 = 1. If V1 = 0, then
V2 = 0.
(a) Calculate the expected value of the two variables.

[15 marks]

(b) Calculate the expected value of V1 conditional on V2 , and the unconditional variance of
the two variables.

[15 marks]

(c) Calculate the covariance and the correlation of the two variables.

[15 marks]

(d) Provide a general definition of two random variables being independent of each other.
Is V1 independent of V2 ? Explain and contrast your results with those of part (c).

[15 marks]

Assume now that a random variable X has a symmetric distribution around zero, i.e., its
probability density function satisfies fX (x) = fX (−x), and that for another random variable,
Y , we know that its expected value conditional on X is also symmetric: E[Y |X = x] =
E[Y |X = −x].
(e) Express the covariance of X and Y , and provide necessary and/or sufficient conditions
that it is zero.

[20 marks]

Another random variable, Z, satisfies that E[Z|X = x] = −E[Z|X = −x].

(f ) Express the covariance of X and Z, and provide necessary and/or sufficient conditions
that it is zero.

[20 marks]

Page 5 of 36
Solutions.

(a) We can simply write E[V1 ] = 31 × (−1) + 13 × 0 + 13 × 1 = 0 and E[V2 ] = 13 × 0 + 23 × 1 = 23 .

.

(b) If V2 = 1, we must have V1 = ±1 with equal (conditional) probabilities, hence E[V1 |V2 =
1] = 0. Moreover, if V2 = 0, we must have V1 = 0, and hence E[V1 |V2 = 0] = 0.

Regarding unconditional variances, we have V ar[V1 ] = 31 × (−1 − 0)2 + 13 × (0 − 0)2 + 13 ×

2 2
(1 − 0)2 = 23 , while V ar[V2 ] = 13 × (0 − 23 )2 + 23 × (1 − 23 )2 = 13 × 23 + 32 × 13 = 29 .

(c) By definition, Cov[V1 , V2 ] = E[V1 V2 ] − E[V1 ]E[V2 ]. From part (a) we know that E[V1 ] =
E[V2 ] = 0, so the second component is zero. To express E[V1 V2 ], we can either write out all
possibilities and their probabilities, to obtain E[V1 V2 ] = 31 ×(−1)×1+ 31 ×0×0+ 13 ×1×1 = 0,
or we can us the law of iterated expectations to write

E[V1 V2 ] = E[E[V1 |V2 ]×V2 ] = E[V1 |V2 = 1]×1×Pr[V2 = 1]+E[V1 |V2 = 0]×0×Pr[V2 = 0] = 0
(1)
due to our results given in part (b). This implies Cov[V1 , V2 ] = 0 and also Corr[V1 , V2 ] = 0.
This approach is actually useful to tackle parts (e) and (f).

(d) Independence means V1 ’s distribution is independent of V2 ’s realization (and vice versa).

In this case, there is clearly a dependence between V1 and V2 : V2 = V12 . If we observe the
value of V1 , we know the value of V2 . Also, a knowledge of the value of V2 will cause us to
change our probability distribution for V1 , even though the conditional expectation is the
same; for example, the conditional variance would already be different. However, because
E[V1 V2 ] = 0 and E[V1 ] = 0, the coefficient of correlation between V1 and V2 was zero. Zero
correlation does not imply independence, only the other way round.

(e) Based on the above example, we can again use the law of iterated expectations to make
use of the information given.
First note that since X has symmetric distribution around zero, its expected value should
be zero. Indeed,
Z ∞ Z 0 Z ∞
E[X] = xfX (x)dx = xfX (x)dx + xfX (x)dx
−∞ −∞ 0
Z ∞ Z ∞
= (−x)fX (−x)dx + xfX (x)dx
0 0
Z ∞ Z ∞
=− xfX (x)dx + xfX (x)dx = 0
0 0

Therefore, Cov[X, Y ] = E[XY ] − E[X]E[Y ] = E[XY ].

Page 6 of 36
 Further, we can write
Z ∞ Z ∞ Z ∞ Z ∞ 
E[XY ] = xyfX,Y (x, y)dxdy = yfY |x (y|x)dy xfX (x)dx
−∞ −∞ −∞ −∞
Z ∞
= [E[Y |X = x] xfX (x)dx,
−∞

where we simply used the conditional distribution of Y given X = x to rewrite the integral.
Decomposing the last expression for positive and negative x values then leads to the
following result:
Z ∞ Z 0 Z ∞
[E[Y |X = x] xfX (x)dx = E[Y |X = x]xfX (x)dx + E[Y |X = x]xfX (x)dx
−∞ −∞ 0
Z ∞ Z ∞
= E[Y |X = −x](−x)fX (−x)dx + E[Y |X = x]xfX (x)dx
0 0
Z ∞ Z ∞
=− E[Y |X = x]xfX (x)dx + E[Y |X = x]xfX (x)dx = 0,
0 0

where we used the same trick as in part (d). Therefore, under the given assumptions we
have Cov[X, Y ] = 0 for sure, no additional assumptions needed.

(f ) Since E[X] = 0, we have Cov[X, Z] = E[XZ] − E[X]E[Z] = E[XZ] exactly as in part

(e), so we only need to focus on E[XZ]. And, following the same steps as in part (e), we
can write
Z ∞
E[XZ] = [E[Z|X = x] xfX (x)dx. (2)
−∞

However, following the same decomposition, we obtain

Z ∞ Z 0 Z ∞
E[Z|X = x]xfX (x)dx = E[Z|X = x]xfX (x)dx + [E[Z|X = x]] xfX (x)dx
−∞ −∞ 0
Z ∞ Z ∞
= E[Z|X = −x](−x)fX (−x)dx + E[Z|X = x]xfX (x)dx
0 0
Z ∞ Z ∞
= E[Z|X = x]xfX (x)dx + E[Z|X = x]xfX (x)dx,
0 0
Z ∞
=2 E[Z|X = x]xfX (x)dx.
0
R∞
Thus, to obtain a zero correlation between X and Z, we must have 0 E[Z|x]xfX (x)dx = 0.
In particular, having E[Z|X = x] ≥ 0 for all x ≥ 0 means Cov[X, Z], Corr[X, Z] > 0.

Page 7 of 36
Question 3

We will be using a GARCH model for updating covariance estimates and forecasting the
future level of covariances. That is, instead of the usual GARCH(1,1) volatility model given
by
2
σt+1 = ωv + αv rt2 + βv σt2 , (3)

we will be working with the model

covt+1 = ωc + αc xt yt + βc covt , (4)

where xt and yt are two daily return series with zero means and covt+1 denotes the covariance
estimate for day t + 1 made at the end of day t.
(a) Explain the restrictions imposed on parameters ωv , αv , and βv to have an internally
consistent and stable GARCH volatility model (3).

[15 marks]

(b) Revisit the requirements listed in your answer in part (a), and explain which one of
them must be imposed on ωc , αc , and βc to have an internally consistent and stable GARCH
covariance model (4).

[10 marks]

Suppose the GARCH covariance model parameters are given by are given by ωc = −0.0002,
αc = 0.02, and βc = 0.95.
(c) Derive the long-run average covariance of the (4) model, and calculate it in this specific
numerical example.

[15 marks]

(d) Explain why we need to be careful when coupling the (4) model with arbitrary models
for variances. Hint: think about the restrictions your result in part (c) implies about the
processes x and y.

[20 marks]

(e) Describe a method you learned of during the course to model multivariate volatility, and
explain how it overcomes the issues raised in part (d).

[20 marks]

Page 8 of 36
(f ) Consider now an EWMA (exponentially weighted moving average) model for variances
and covariances jointly, described as follows:

2
σt+1,x = (1 − λ)x2t + λσt,x
2
, (5)
2
σt+1,y 2
= (1 − λ)yt2 + λσt,y , and (6)
covt+1 = (1 − λ)xt yt + λcovt , (7)

where λ ∈ (0, 1). Show that (i) this model is not covariance stationary, but (ii) it yields a
positive definite variance-covariance structure.

[20 marks]

Page 9 of 36
Solutions.

(a) GARCH models are used for the conditional variance, but without further restrictions
they can lead to a violation of covariance stationarity. The requirements for a GARCH(1,1)
process to be covariance stationary are the following:

• Condition 1: ωv > 0, αv , βv ≥ 0, for positive variance

• Condition 2: βv = 0 if αv = 0, for identification

• Condition 3: αv + βv < 1, for covariance stationarity

(b) We would still need to have condition 2 for identification, βc = 0 if αc = 0, and, as we will
show in part (d), condition 3 for covariance stationarity, αc + βc < 1. However, covariance
does not have to be positive, and thus we can allow for ωv ≤ 0. Note that it would still make
sense to keep the parts αv , βv ≥ 0, otherwise covariance would become oscillatory.

(c) As long as αc + βc < 1, which is satisfied here, we can follow the derivation as it is
normally done for a GARCH volatility model:

E[covt+1 ] = ωc + αc E[xt yt ] + βc E[covt ] = ωc + αc E[Et−1 [xt yt ]] + βc E[covt ]

= ωc + αc E[covt ] + βc E[covt ] = ωc + (αc + βc )E[covt ],

so if we are looking for a setting with E[covt ] = E[covt+1 ], we obtain

ωc
E[covt ] = E[covt+1 ] = . (8)
1 − (αc + βc )
−0.0002
Plugging in the given numerical coefficients we obtain E[covt ] = 1−(0.02+0.95)
= −0.02.

(d) One of the two main issues with variance-covariance models is that they need to sat-
isfy positive definiteness: Left unconstrained, some multivariate volatility models do not
always yield positive (semi-)definite conditional covariance matrix estimates. This is the
multivariate ‘equivalent’ of a univariate volatility model predicting a negative variance. This
is clearly undesirable. Imposing positive definiteness on some models leads to non-linear
constraints on the parameters of the models which can be difficult to impose practically.

Translating positive definiteness to our 2-variable case, with x and y, the covariance
matrix would fail to be positive definite ifqand only if the correlation coefficient was outside
2 2
the [−1, 1] range, i.e. if |cov(xt , yt )| > σx,t σy,t . In fact, the (4), just like the GARCH
model, would have a finite long-term mean to which it would mean revert, but it could easily
diverge from it. In the particular case of setting σx,t and σy,t constant, we would have to put

Page 10 of 36
an upper threshold on the covariance to ensure positive definiteness, which cannot happen
if daily returns are normally distributed and hence can take arbitrarily large values.

With multivariate models, requiring correlations between −1 and 1 would not be enough
to obtain positive definiteness. In fact, one would also have to ensure that pairwise corre-
lations are consistent with each other. E.g., having 2 variables that are perfectly negatively
correlated, it cannot be the case that one of them has a perfect positive while the other has
a perfect negative correlation with a 3rd variable.

(e) Here we are looking for students mentioning Bollerslev (1990)’s constant conditional
correlation (CCC) model. A (conditional) covariance can be decomposed into three parts
p
hijt = hit hjt ρijt

That is, the conditional covariance between rit and rjt is equal to the product of the two
conditional standard deviations and the conditional correlation between rit and rjt .
Bollerslev (1990) proposed assuming that the time variation we observe in conditional
covariances is driven entirely by time variation in conditional variances; i.e., the conditional
correlations are constant. Making this assumption greatly simplifies the specification and es-
timation of multivariate GARCH models. With this assumption, the conditional covariances
are obtained as
p
hijt = hit hjt ρ̄ij
√
where ρ̄ij is the unconditional correlation between the standardised residuals εit / hit and
p
εjt / hjt , and hit and hjt are obtained from some univariate GARCH model.
Bollerslev’s model is both parsimonious and ensures positive definiteness. However, there
exists some empirical evidence against the assumption that conditional correlations are con-
stant. Nevertheless, Bollerslev’s model is a reasonable benchmark against which to compare
any alternative multivariate GARCH model.
Actually, CCC also overcomes the general issue of parsimony: Models for time-varying
covariance matrices tend to grow very quickly with the number of variables being considered:
that the covariance matrix of a collection of k variables grows with k 2 = O (k 2 ), while
the number of data points only grows linearly with the number of variables (number of
observations = kT = O (k)). It is important that the number of free parameters to be
estimated be controlled somehow, and this often comes at the cost of flexibility.

(f ) Note that an EWMA model looks like a GARCH model with the additional restrictions
of λ = β = 1 − α and ω = 0.

Why does positive definiteness work in this case? Denoting the variance-covariance ma-

Page 11 of 36
trix by " #
2
σt,x covt
Ωt = 2
,
covt σt,y
we would need to have that
w> Ωt w ≥ 0

for all w 2 × 1 vectors. However, notice that the system of (5)-(7) can be written in the
matrix form of " #
2
xt xt y t
Ωt+1 = λ + (1 − λ) Ωt . (9)
xt yt yt2
But here " #
x2t xt yt
xt yt yt2
is positive definite, since
" #" #
x2t xt yt w1
[w1 , w2 ] = w12 x2t + 2w1 w2 xt yt + w22 yt2 = (w1 xt + w2 yt )2 ≥ 0.
xt yt yt2 w2

Therefore, if Ωt is positive definite, from (9) then Ωt+1 is also positive definite:
" # " #
w1 w1
[w1 , w2 ] Ωt+1 = λ (w1 xt + w2 yt )2 + (1 − λ) [w1 , w2 ] Ωt ≥ 0.
w2 w2

Page 12 of 36
Question 4

Suppose there are two one-year loans, each with current market value of $10 million, and
we want to study their (joint) payoff structure and risk. Assume that if one loan defaults,
it is certain that the other one will not default, and the three possible scenarios regarding
default/solvency have the following probabilities:

• Both loans remain solvent, with probability 84%.

• Loan A remains solvent and loan B defaults, with probability 8%.

• Loan B remains solvent and loan A defaults, with probability 8%.

Assume also that conditional on Loan A defaulting, it either suffers a 100% loss or a 20%
loss with equal probabilities. On the other hand, if Loan B defaults, all losses between 20%
and 100% are equally likely. Finally, if a loan does not default, its terminal payoff is $10.5
million.
(a) What are the 1% and 5% VaRs for each loan separately?

[15 marks]

(b) What are the 1% and 5% VaRs for a portfolio consisting of Loans A and B? Compare
them to the individual VaRs and discuss your findings.

[25 marks]

(c) Is VaR sub-additive in this example? Explain why the absence of sub-additivity may be
a concern for risk managers.

[10 marks]

(d) The expected shortfall ES α at the α level can be defined as

ES α = −Et [R|R < −V aRα ] ,

where R denotes the random variable describing the (dollar) profits/losses; i.e., it gives the
average loss conditional on suffering losses larger than the α critical level Value-at-Risk,
V aRα .
What are the 1% and 5% expected shortfalls for each loan?

[15 marks]

Page 13 of 36
(e) What are the 1% and 5% expected shortfalls for a portfolio consisting of Loans A and
B? Compare them to the individual ESs and comment on your results.

[20 marks]

(f ) Assume that daily returns on another portfolio are independently and identically nor-
mally distributed. A newly hired quantitative analyst has been asked by the portfolio man-
ager to calculate portfolio VaRs for 10-, 15-, 20-, and 25-day periods. The portfolio manager
takes a look at the analyst’s calculations displayed below (rounded to millions), and notices
that one of them is inconsistent with the others. Which one is inconsistent with the others
and why? What should be the correct number there?

A. VaR(10-day) = $316m

B. VaR(15-day) = $465m

C. VaR(20-day) = $537m

D. VaR(25-day) = $600m

[15 marks]

Page 14 of 36
Solutions.

(a) Each individual loan remains solvent with probability 92% and defaults with probability
8%. From here, the one-year 1% VaR on Loan A is $10m and the 5% is $2m, because the
unconditional probability of a loss greater than $2m, i.e. a loss of $10m, is 4%, whereas
suffering any loss has a probability of 8% in total. For Loan B, the uniform distribution
implies a 1% VaR of $9m and the 5% VaR is $5m — this is easiest to see with a graphical
illustration of the distribution of gains/losses (which are expected for full marks).

(b) For the portfolio of the two loans, the one-year 1% VaR is $5.8m:

• There is a 4% probability of Loan A defaulting and making a $10m loss while B would
make a gain of $0.5m, so the loss is $9.5m.

• There is a 4% probability of Loan A defaulting and making a $2m loss while B would
make a gain of $0.5m, so the loss is $1.5m.

• If B suffers a uniform loss, A pays off $0.5m for sure.

These imply the cdf of the portfolio has a jump of 4% at -$9.5m, a 4% jump at -$1.5m,
an overall increase of 8% in between, and then an 88% jump at the payoff of $1m. Thus,
after some algebra, the 1% VaR is -$9.5m and the 5% VaR is -$8.5m.

(c) In spite of diversification benefits—the second loan always pays off when the first fails,
i.e. always helps adding them together—portfolio VaR is higher for the 5% case. At 1%
we have 10 + 9 > 9.5, but at 5%, we have 2 + 5 < 8.5. This failure of sub-additivity can
lead to perverse incentives for portfolio managers: if they are evaluated based on VaR, they
might be reluctant to add additional assets to their portfolios for diversification or hedging
purposes, taking on more idiosyncratic risk.

(d) The expected shortfall is the average loss conditional of it being larger than the VaR.
The one-year 1% ES on Loan A is $10m and the 5% is

10 × 4% + 2 × 1%
= 8.4%.
5%

For Loan B, the one-year 1% ES is $9.5m due to the conditional uniform distribution in the
[9, 10] range, whereas the 5% is $7.5 due to the conditional uniform distribution in the [5, 10]
range.

(e) On the joint portofolio, the 1% ES is trivially $9.5m, whereas the 5% one is

9.5 × 4% + 8.5+9.5
2
× 1%
= 9.4m.
5%

Page 15 of 36
They both satisfy sub-additivity, as ES does in general.

(f ) The right answer is A. We need to, for example, calculate the VaR(1-day) from each
choice and compare them. It turns out that the VaR(1-day) corresponding to A is different
from those from others: A yields
√
V aR(1 − day) = V aR(10 − day)/ 10 = 316/3.16 = $100m,

whereas B, C, and D satisfy

V aR(15 − day) V aR(20 − day) V aR(25 − day)

V aR(1 − day) = √ = √ = √ = $120m;
15 20 25

the only differences come from rounding. The number from A is obviously incompatible with
the others.

Page 16 of 36
 II. ZONE B

Question 1

Consider the zero-mean M A(1) process Xt given by

Xt = ut + δut−1 ,

where ut is white noise, i.e.,

(
σu2 if j = 0
E[ut ut−j ] =
0 otherwise

Consider another white noise vt :

(
σv2 if j = 0
E[vt vt−j ] =
0 otherwise

such that v and u are independent for all leads and lags. Let an observed series Yt be defined
as
Yt = Xt + v t ,

and another one, Zt , defined as

Zt = ut vt .

(a) Define covariance stationarity.

[15 marks]

(b) Prove that the process Xt is covariance stationary, and calculate its autocovariances.

[20 marks]

(c) Prove that Yt is also covariance stationary and its auto-covariances are zero beyond one
lag.

[20 marks]

(d) Is it possible to represent Yt as an MA(1) process? In particular, assume that we want

to write
Yt = εt + θεt−1 , (10)

where εt is a zero mean white noise with variance σε2 . What are the required restrictions on
θ and σε2 so that Yt is indeed MA(1)?

Page 17 of 36
 [25 marks]

(e) Prove that Zt is also covariance stationary, and calculate its auto-covariances.

[20 marks]

Page 18 of 36
Solutions.

(a) A time series Wt is “covariance stationary” if

E[Wt ] = µ ∀t
V ar[Wt ] = σ 2 ∀t
Cov[Wt , Wt−j ] = γj ∀t, j,

that is, unconditional first and second moments are finite and constant through time (they
do not depend on time), and autocovariance depends only on the lag j.

(b) Standard calculations yield

E [Xt ] = E [ut + δut−1 ] = 0,

and hence

Cov [Xt , Xt−j ] = E [Xt Xt−j ]


2 2
 (1 + δ ) σu if j = 0

= E [(ut + δut−1 ) (ut−j + δut−j−1 )] = δσu2 if j = ±1

0 otherwise.


(c) Standard calculations yield

E [Yt ] = E [Xt ] + E [vt ] = 0

and

E [Yt Yt−j ] = E [(Xt + vt ) (Xt−j + vt−j )] = E [Xt Xt−j ] + E [vt vt−j ]

(
E [Xt Xt−j ] + σv2 if j = 0
=
E [Xt Xt−j ] otherwise,

2 2 2
 (1 + δ ) σu + σv if j = 0

= δσu2 if j = ±1

0 otherwise.


where we used part (b) in the last step.

It is imminent from here that (i) E [Yt ] does not depend on t; (ii) V ar [Yt ] does not depend
on t; and (iii) for all j and t Cov [Yt , Yt−j ] is finite and depends on j only. Therefore, Yt is
indeed covariance stationary, and its autocovariances are zero from lag two.

Page 19 of 36
(d) From (10) we obtain
E [Yt ] = E [εt + θεt−1 ] = 0

while

2 2
 (1 + θ ) σε if j = 0

Cov [Yt , Yt−j ] = E [Yt Yt−j ] = E [(εt + θεt−1 ) (εt−j + θεt−j−1 )] = θσε2 if j = ±1

0 otherwise.


In order to be consistent with the autocovariances calculated in part (d), we need to have

1 + θ2 σε2 = 1 + δ 2 σu2 + σv2

and
θσε2 = δσu2 .

The second implies

δσu2
σε2 =
,
θ
and substituting it back into the previous equation, we obtain the quadratic equation

(1 + δ 2 ) σu2 + σv2
θ2 − θ + 1 = 0,
δσu2

thus q
(1 + δ 2 ) σu2 + σv2 ± [(1 + δ 2 ) σu2 + σv2 ]2 − 4 (δσu2 )2
θ=
2δσu2
Notice that to insure σε2 remains positive, we must choose the root such that θ and δ have
the same sign.

(e) Standard calculations yield

E [Zt ] = E [ut vt ] = E [ut ] E [vt ] = 0,

where we used the independence of u and v. Further,

(
σu2 σv2 if j = 0
Cov [Zt , Zt−j ] = E [ut vt ut−j vt−j ] = E [ut ut−j ] E [vt vt−j ] =
0, otherwise,

simply from independence and the properties of u and v. Therefore, Zt is covariance sta-
tionary, and its autocovariance is zero for all j ≥ 1.

Page 20 of 36
Question 2

Vasicek’s loan portfolio model assumes that payoffs on individual loans over a fixed time
horizon T are given by the structure
√
Ui = aF + 1 − a2 Zi , (11)

where F and Zi , i = 1, ..., N , are pairwise independent standard normal random variables,
and a ∈ [0, 1) constant.
Suppose a loan is considered in default if its payoff is below a threshold U , and let us
denote the default probability of a single loan by p ∈ (0, 1).

(a) Express the pairwise correlation between the returns of 2 individual loans, Ui and Uj
with i 6= j, and denote it by ρ.

[15 marks]

(b) Express the relationship between U and p.

[15 marks]

(c) Express the probability of default of a single loan given the realization of the common
factor, F .
Hint: it might help expressing Zi from (11) and working with that form.

[20 marks]

Consider now a portfolio of many (a continuum of) identical loans. We want to calculate
the “worst case default rate,” denoted by W CDR(q), that is defined as the fraction of loans
defaulting that will not be exceeded with probability q.

(d) Building on your previous results, show that

√
N −1 (p) + ρN −1 (q)
 
W CDR(q) = N √ (12)
1−ρ

where N (.) and N −1 (.) are the standard cumulative normal distribution function and its
inverse, respectively.
Hint: With a continuum of identical loans, your answer to part (c) can be used as a proxy
for the fraction of loans defaulting, conditional on F .

[25 marks]

Page 21 of 36
(e) Calculate W CDR if ρ = 0, explain how it changes when ρ increases, and express its
limit as ρ → 1. Provide intuition for these results.

[25 marks]

Page 22 of 36
Solutions.

(a) From (11), it is imminent that

√
E[Ui ] = aE[F ] + 1 − a2 E[Zi ] = 0, (13)

whereas
√ √
Cov[Ui , Uj ] = E[Ui Uj ] − E[Ui ]E[Uj ] = E[(aF + 1 − a2 Zi )(aF + 1 − a2 Zj )] (14)
(
a2 E[F 2 ] + (1 − a2 )E[Zi2 ] if i = j
= (15)
a2 E[F 2 ] otherwise
(
a2 V ar[F ] + (1 − a2 )V ar[Zi ] if i = j
= (16)
a2 V ar[F ] otherwise
(
1 if i = j
= (17)
a2 otherwise,

which also yields that as long as i 6= j,

Cov[Ui , Uj ]
ρ ≡ Corr[Ui , Uj ] = p = a2 , (18)
V ar[Ui ]V ar[Uj ]

where we repeatedly used the assumptions of independence and jointly standard Gaussian
distributions.

(b) Note that by definition

p = P r[i is in default] = P r[Ui < U ] = N (U ), (19)

where in the last step we used that due to its definition and the assumptions on F and Zi ,
Ui ∼ N (0, 1). From here we can also write U = N −1 (p).

(c) Suppose we know F ; then from above, we have P r[i is in default|F ] = P r[Ui < U |F ].
However, given the definition of Ui , (11), we can write

Ui − aF
Zi = √ ∼ N (0, 1), (20)
1 − a2

so, after some algebra, we obtain

 
Ui − aF U − aF
P r[Ui < U |F ] = P r √ <√ |F (21)
1 − a2 1 − a2
 −1 √ 
N (p) − ρF
   
U − aF U − aF
= P r Zi < √ =N √ =N √ . (22)
1 − a2 1 − a2 1−ρ

Page 23 of 36
Note that this is very similar to the expression in (12).

(d) For a large portfolio of loans with the same probability of default, where the correlation
for each pair of loans is ρ, the above expression equation provides a good estimate of the
percentage of loans defaulting by time T conditional on F . Let’s call it default rate.
As F decreases, the default rate increases. So we can write down the connection between
the default rate and the level of F . In particular, since F has a standard normal distribution,
the probability that F will be less than N −1 s is s. There is therefore a probability of s that
the default rate will be greater than
√
N −1 (p) − ρN −1 (s)
 
N √ . (23)
1−ρ

The default rate that we are q% certain will not be exceeded is then obtained by substituting
s = 1 − q into the preceding expression. Since N −1 (s) = −N −1 (1 − s), we obtain equation
(12).
(e) When ρ = 0, we obtain

W CDR(q) = N N −1 (p) = p.

This makes sense as when the loans are uncorrelated, exactly a portion of p will default,
and hence forecasting this default rate of p has nothing to do with the certainty we want to
assign to our forecast.
As ρ increases, W CDR(q) increases—increasing correlation means that the worst case
scenarios are getting worse and worse, since loans will be defaulting at the same time.
Finally, as ρ → 1, the denominator of the fraction inside the parentheses goes to zero,
i.e., W CDR(q) → 1. Indeed, with perfectly correlated loans, if one of them fails all fail.
Eventually this is the case when we we would plot a pdf of outcomes that jumps from 0 to
1, so if we want to prepare for the worst q% of outcomes, we must be prepared for all loans
defaulting.

Page 24 of 36
Question 3

(a) Describe the Diebold-Mariano test for comparing two forecasts.

[20 marks]

(a) (b)
(b) Suppose two analysts have provided you with a forecast each, ŷt and ŷt , and you
have the following information:

1000 
X 2  2  1000 
X 2   2 2
(a) (b) (a) (b)
et − et = 100 and et − et = 2000, (24)
t=1 t=1

(a) (b)
where et and et are the forecasting errors from the two models. Assuming that the data are
serially independent, describe how you would use this information to test whether forecast
a is significantly better/worse than forecast b. Conduct the test and interpret the result.
Hint: recall that the 95% confidence value for a N(0,1) random variable is 1.96.

[25 marks]

Suppose the parameters in a GARCH (1,1) model

σt2 = ω + αrt−1
2 2
+ βσt−1 (25)

are given by ω = 0.000003, α = 0.03, and β = 0.94.

(c) Suppose yesterday’s volatility is 2% and the realized return is 1%. What is your forecast
of the current volatility based on this GARCH model?

[5 marks]

(d) Derive the long-run average volatility of a GARCH model and calculate it in this specific
numerical example.

[15 marks]

(e) Explain how to derive the formula to obtain 10-day ahead volatility estimates in GARCH
models, and calculate it for this particular model.

[15 marks]

Page 25 of 36
(f ) During the course it was discussed how models “technically similar” to the GARCH
model of volatility can be used to study other aspects of trading data. Briefly describe one
such topic and the modelling technique.

[20 marks]

Page 26 of 36
Solutions.

(a) A Diebold-Mariano test compares the accuracy of two forecasts. Assume that there are
two forecasts of Yt , Ŷta and Ŷtb . The test works by comparing the expected loss incurred
when using the two forecasts:
   
la,t = L Yt , Ŷta and lb,t = L Yt , Ŷtb ,

where the loss function L can be any valid loss function. Then we define dt ≡ la,t − lb,t . The
null hypothesis is that the two forecasts are equally good, and so we test

H0 : E[dt ] = 0 vs Ha : E[dt ] 6= 0.

If Forecast A is better than Forecast B it will have lower expected loss and thus the mean
of dt will be significantly less than zero.
During the course the most frequently used approach was comparing the forecasts by
comparing the difference in the squared errors from the two forecasts:
 2  2
(a) (b)
dt = et − et , (26)

as in part (b) of this exam question. One could also think about the utility or profits made
when following the two forecasts, thus using economic measures of goodness-of-fit. In this
case one would write:
   
la,t = −U Yt , Ŷta and lb,t = −U Yt , Ŷtb ,

where U is some utility function. We use a minus sign in front of the utility function to
make la,t and lb,t “losses” rather than “gains”. The difference in the squared errors measure
is expected from students for a full mark.
Since the Diebold-Mariano test is simply a test that the mean of a random variable is
equal to zero; one can use a simple t-test that E[dt ] = 0, using the fact that

d¯
DM = q   → N (0, 1) , T → ∞,
V̂ d¯

where the sample mean is computed in the usual way:

T
¯ 1X
d ≡ Ê [dt ] = dt
T t=1

If the data are serially independent, then the dt series will be serially uncorrelated, and so

Page 27 of 36
the denominator is simply
1
V̂ d¯ = V̂ [dt ]
 
T
where
T
1X
2
V̂ [dt ] = dt − d¯
T t=1
and thus the DM test statistic is the usual t-statistic:
√
T · d¯
DM = q (27)
V̂ [dt ]

Students are expected to describe the test statistic written in the last equation, together
with the definitions, although the exact derivation is not necessary.

(b) A Diebold-Mariano test can be used here. This test will tell us whether forecast a has a
different expected squared error than forecast b. To compute the DM test statistic, we first
use the first equation given in (24), which yields

1000   1000
1 X  (a) 2  (b) 2 1 X
d¯ = et − et = dt = 0.1, (28)
1000 t=1 1000 t=1

where dt is defined as in (26). At the same time, combining the two equations in (24) yields

1000  2 1000
1 X  (a) 2  (b) 2 ¯ 1 X
2
V̂ [dt ] = et − et −d = dt − d¯ = 1.99,
1000 t=1 1000 t=1

2
where we made use of the identity dt − d¯ = d2t − 2dt d¯ + d¯2 . Therefore, since d¯ = 0.1 and
V̂ [dt ] = 1.99, the write the DM test statistic as:
√ √
T · d¯ 1000 · 0.1
DM = q = √ = 2.24
V̂ [dt ] 1.99

The test statistic is higher than the critical threshold of 1.96, and so we reject the null
hypothesis that the two forecasts are equally accurate. Since d¯ > 0, the squared forecast
errors of forecast a are larger, i.e., b is a better forecast.

(c) We simply write

E σt2 = 0.000003 + 0.03 · 0.012 + 0.94 · 0.022 = 0.000382

 

and the volatility forecast is 1.95%.

Page 28 of 36
(d) Taking expectations, we obtain E[σt2 ] = ω + αE[rt−1
2 2
] + βE[σt−1 ], and from here

ω 0.000003 0.000003
VL = = = = 0.0001
1−α−β 1 − 0.03 − 0.94 1 − 0.03 − 0.94

and thus the the long-run average volatility is 0.01 = 1%.

(e) The variance rate estimated at the end of day t−1 for day t, when a GARCH(1,1) model
is used, is given by
σt2 = (1 − α − β)VL + αrt−12 2
+ βσt−1 ,

hence
2
σt2 − VL = α rt−1 2



− VL + β σt−1 − VL .

Suppose we are at date t For day t + n in the future, conditional on the information available
at t + n − 1, we write

2 2 2
σt+n − VL = α(rt+n−1 − VL ) + β(σt+n−1 − VL ).

2 2
Let us now go back to t + n − 2. If Et+n−2 [rt+n−1 ] = 0, the expected value of rt+n−1 is σt+n−1 :
 2  2
Et+n−2 rt+n−1 = σt+n−1 . Hence,
 2   2  2
Et+n−2 σt+n − VL = α(Et+n−2 rt+n−1 − VL ) + β(σt+n−1 − VL )
2
= (α + β) (σt+n−1 − VL ).

Using this equation repeatedly yields

− VL = (α + β)n−1 (σt2 − VL ).
 2   2 
Et σt+n − VL = Et+n−n σt+n

i.e.
2
= VL + (α + β)n−1 (σt+1
 2  2
σt,t+n = Et σt+n − VL ).

This equation forecasts the volatility on day t + n using the information available at the end
of day t.
With the current numerical example, we obtain

= VL + (α + β)9 (σt+1
 2  2
Et σt+10 − VL ) = 0.000314385.

Taking squareroot, the volatility forecast is 1.773%. (Slightly different answers can come
if the “current” volatility calculated in part (a) is used as σt2 instead of σt+1
2
, but both are
accepted.)

(f ) Trade duration can be modelled similarly to conditional variance. Trade duration is

defined simply as time between two trades, usually measured in seconds. The ACD (“au-

Page 29 of 36
toregressive conditional duration”) model proposed by Engle and Russell (1998) for modelling
durations is similar in spirit to the GARCH model. In certain situations the similarity be-
tween the ACD and the GARCH models even allows certain results for GARCH models to
be applied to ACD models. One possible model for durations is:

xj = sj ψj εj , where
εj ∼ iid(1),
X13
log sj = βi Di,j ,
i=1
xj
x∗j =
sj
ψj = ω + βψj−1 + αx∗j−1 , ω > 0, α, β ≥ 0

That is, realised durations are decomposed into three parts: a deterministic diurnal com-
ponent (sj ), an innovation term with conditional mean 1 (εj ) and a time-varying conditional
mean term (ψj ), which is modelled as an ACD(1,1) process. If we look at de-seasonalized
durations we get:
xj
x∗j = = ψj εj , hence
sj
Ej−1 x∗j = Ej−1 [ψj εj ] = ψj Ej−1 [εj ] = 1
 

since εj is iid with conditional mean 1.

There is a similarity between the GARCH and the ACD models regarding also parameter
restrictions and moments. The conditions for covariance stationarity, identification and
positivity of the duration series are identical for the ACD(1,1) as for the GARCH(1,1), if
they are written as
ψj+1 = ω + βψj + αx∗j ,

and
2
σt+1 = ω + βσt2 + αε2t ,

respectively:
• Condition 1: ω > 0 (or ω ≥ 0 for ACD), α, β ≥ 0, for positivity.

• Condition 2: β = 0 if α = 0, for identification.

• Condition 3: α + β < 1, for covariance stationarity.

If α + β < 1, then
ω
E [ψj ] = E [Ej−1 [xj ]] = E [xj ] = ,
1−α−β

Page 30 of 36
and hence the (unconditional) average duration implied by the ACD(1,1) is the same formula
as that used to obtain the unconditional variance of the residuals from a GARCH(1,1) model.

Both the ACD model and the parameter restrictions are needed for a complete answer.

Some extra students might also bring up: Finally, the GARCH likelihood can be written
as
n−1 1P n
2
+ α (rt−1 − µt−1 )2


LGARCH = − log (2π) − log ω + βσt−1
2 2 t=2
1P n (rt − µt )2
−
2 t=2 ω + βσt−1
2
+ α (rt−1 − µt−1 )2

Suppose we set µt = 0 for all t and rt = x∗j ; then we can show that (i) ψj performs the
p

same function as σj2 in a GARCH(1,1) model, and (ii) the GARCH likelihood is simply a
linear transformation of the ACD likelihood. Formally,

σj2 = Ej−1 (rj − µj )2 = Ej−1 rj2

   

= Ej−1 x∗j = Ej−1 [ψj εj ] = ψj ,

 

so σj2 and ψj have the same interpretation: they are the conditional mean of the squared
dependent variable.

Moreover, in the LGARCH formula, the RHS becomes

n−1 1P n
2


LGARCH = − log (2π) − log ω + βσt−1 + αxt−1
2 2 t=2
∗
1P n xt n−1 1
− 2
=− log (2π) + LACD.
2 t=2 ω + βσt−1 + αxt−1 2 2

Therefore, the GARCH likelihood is simply a linear transformation of the ACD likelihood. In
addition to being an interesting link between the two models, it allowed Engle and Russell
(1998) to obtain theoretical results on the ACD model from previously known results on
the GARCH model. It also has the practical benefit that we may use GARCH software to
estimate ACD models: we simply impose a zero mean and estimate a GARCH(1,1) model
on the square root of the duration series. The resulting parameter estimates will be those
that maximize the ACD log-likelihood, and the forecast variance series from the GARCH
software will instead be the time series of expected durations from the ACD model.

Page 31 of 36
Question 4

(a) Imagine the following gamble. First, flip a fair coin to determine the amount of your bet:
if heads, you bet $1, if tails you bet $2. Second, flip again: if heads, you win the amount of
your bet, if tails, you lose it. For example, if you flip heads and then tails, you lose $1; if
you flip tails and then heads you win $2.) Let X be the amount you bet, and let Y be your
net winnings (negative if you lost).
Show that the covariance between X and Y is zero.

[20 marks]

(b) In the game described in part (a), show that X and Y are not independent.

[20 marks]

(c) Describe the diurnal dummy model for returns, and show how one can test for the
importance of seasonality/diurnality.

[25 marks]

(d) Explain the linear and higher order polinomial trend models for capturing patterns in
intra-daily returns.

[15 marks]

(e) Estimating a cubic trend model on IBM returns we obtained the following parameter
estimates:

Coeff Std error

Constant 0.1067 0.0469
HOURS −0.0153 6.1543 × 10−3
HOURS2 8.4096 × 10−4 3.1064 × 10−4
HOURS3 −8.3457 × 10−6 5.5186 × 10−6

Derive analytically the time of day that yields the lowest expected returns, according to
the cubic trend model.

[20 marks]

Page 32 of 36
Solutions.

(a) One solution is to explicitly calculate the covariance the following way: since X = 1 or
2 with equal (1/2) probabilities, we have E[X] = 1.5. At the same time, Y takes the values
−1, 1, −2 and 2 with equal (all 1/4) probabilities, so we obtain E[Y ] = 0. Finally, XY takes
the values of −1, 1, −4 and 4 with equal probabilities. Therefore, E[XY ] = 0 again. From
here, by the definition of covariance, we obtain Cov[X, Y ] = E[XY ] − E[X]E[Y ] = 0.

An alternative approach is the following: This is a game conditional on the outcome of

two fair coin tosses. Therefore, to show zero covariance, it is enough to show a fair coin
toss is serially uncorrelated. To do this, denote the first toss event by A, with outcomes
AH and AT , and the second toss B, with outcomes BH and BT . If these events are serially
uncorrelated, the occurrence of one event does not impact the likelihood of the other event,
i.e., they are independent:
P (B A) = P (B)

and this implies

P (A ∩ B) = P (A) × P (B A) = P (A) × P (B)

Draw a binomial tree for the path of the coin tosses. Note then that

1 1 1
P (AH ∩ BH ) = P (AT ∩ BH ) = P (AH ∩ BT ) = P (AT ∩ BT ) = × = .
2 2 4
Hence, a fair coin toss in serially uncorrelated because the event tree implies independence.

Finally, there is a way to ”combine” these two approaches that also highlights the general
forces going on here. Note that by definition, Y ≡ X · S, where S = −1 or 1, with equal
probabilities. This variable essentially denotes the sign, that is, whether one wins or loses
money. Since X depends on the first coin toss and S depends on the second coin toss, they
are independent. From here, we can write

Cov[X, Y ] = Cov[X, XS] = E[X 2 S] − E[X]E[XS] = E[X 2 ]E[S] − E[X]E[X]E[S]

= (E[X 2 ] − E 2 [X])E[S] = V ar[X]E[S] = 0,

where the second equality uses the definition of covariance, the third equality uses that X
and S are independent, the fifth equality uses the definition of variance, and the last one
uses that S has a zero mean. This approach also shows that the specific values of X do not
affect the outcome of the covariance of X and Y .

(b) However, the outcomes of the X and Y are not independent. If you know Y then you
know X—e.g., if Y = −2 then you know X has to be 2. Explicitly, the probability that
Y = −2 is 41 and the probability that X = 2 is 12 , but the probability that they occur jointly

Page 33 of 36
is 14 , which does not equal the 2 probabilities multiplied together, which would be 81 . Indeed,
there is no event in the game with a probability of 18 .

(c) Say we have a variable called T IM Et , which contains the time that trade t took place
(say 9:30, 14:23, etc.). Then we could define diurnal dummy variables for time intervals, e.g.
as (
1 T IM Et ≤ 10 : 00
D1,t =
0 otherwise
and (
1 10 : 00 < T IM Et ≤ 11 : 00
D2,t =
0 otherwise
etc. The pure diurnal dummy model for returns is then
n
X
rt = βi Di,t + εt .
i=1

where n is the number of time intervals we chopped the day up to. This model will capture
any pattern in returns over the trade day. (Since we have included a dummy for all possible
times of the day, we cannot also include a constant term in the above regression; doing
so would lead to perfect multicollinearity of the regressors and the regression would fail.
Alternatively, we could include a constant term and then use only n − 1 of the possible n
dummy variables.) Note that we expect intra-daily returns to be heteroskedastic, and so it
is important to use robust standard errors, such as Newey-West standard errors.

We test for the importance of seasonality/diurnality by running the regression and then
testing whether all coefficients are equal. In the above example we would test:

H0 : β1 = β2 = ... = βn vs Ha : at least two βi 6= βj .

This test can be simply conducted using a Chi-square test with n − 1 degrees of freedom.

(d) For example, we could specify a linear trend:

rt = β0 + β1 · HOU RSt + εt ,

where HOU RSt counts the number of hours between midnight and the time of return; or
with M IN St , which counts the number of minutes between midnight and the time of the
return (e.g., 570 for a 9:30 return, 573.28 for a 9:33:17 return, etc.). This model allows
expected returns to be increasing or decreasing (or constant) in a linear fashion throughout
the trade day.

Page 34 of 36
 We could alternatively use a quadratic or even higher order trend model:

rt = β0 + β1 · HOU RSt + β2 · HOU RSt2 + ... + βk · HOU RStk + εt ,

which allows expected returns to take more elaborate shapes throughout the trade day—
e.g., a quadratic model would allow for U or inverted U shape. In high-frequency data the
quadratic trend is often used.

(e) The cubic trend model for expected returns is

E [rt ] = 0.1067 + (−0.0153) · HOU RSt + 8.4096 × 10−4 · HOU RSt2 − 8.3457 × 10−6 · HOU RSt3 .

Actually, if one calculates the t-stats associated with the coefficient of the cubic term, it
turns out to be -1.51, i.e., insignificant, and one way to proceed is to ignore this term (the
others have t-stats above 2, see the below table).

Coeff Std error t-stat

Constant 0.1067 0.0469 2.28
HOURS −0.0153 6.1543 × 10−3 −2.49
HOURS2 8.4096 × 10−4 3.1064 × 10−4 2.71
HOURS3 −8.3457 × 10−6 5.5186 × 10−6 −1.51

If one proceeds like this, we get

E [rt ] = 0.1067 + (−0.0153) · HOU RSt + 8.4096 × 10−4 · HOU RSt2 .

The minimum of this quadratic function must satisfy the first order condition

0 = −0.0153 + 2 × 8.4096 × 10−4 · HOU RSt

i.e.
0.0153
HOU RSt = = 9.09,
2 × 8.4096 × 10−4
and it is indeed a minimum as the seccond derivative is 2 × 8.4096 × 10−4 > 0. Thus, the
lowest expected returns of the day are at 9.09 hours after midnight, i.e., at 9:06am.

Alternatively, one could still take the cubic term into account and use the full cubic
model; then the FOC is

0 = −0.0153 + 2 × 8.4096 × 10−4 · HOU RSt − 3 × 8.3457 × 10−6 · HOU RSt2

Page 35 of 36
which has roots
q
−2 × 8.4096 × 10 −4
± [2 × 8.4096 × 10−4 ]2 − 4 × 3 × 8.3457 × 10−6 × 0.0153
HOU RSt =
−2 × 3 × 8.3457 × 10−6
= 10.84 and 56.33.

Clearly this second one is outside the [0, 24] range, and we need to confirm that in that range
10.84 is indeed a minimum—the SOC should be positive and the values at the boundaries
should all be higher than at this value. Since these are all satisfied, we can conclude that
the minimum is achieved 10.84 hours after midnight, i.e., at 10:51am.

Page 36 of 36