Examiners’ commentaries 2021

Examiners’ commentaries 2021

FN3142 Quantitative finance

Important note

This commentary reflects the examination and assessment arrangements for this course in the
academic year 2020–21. 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

Learning outcomes

At the end of this course and having completed the Essential reading and activities, you should:

be able to demonstrate mastery of econometric techniques required in order to analyse

issues in asset pricing and market finance

be able to demonstrate familiarity with recent empirical findings based on financial

econometric models

understand and have gained valuable insights into the functioning of financial markets

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

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 econometric techniques, such as time series analysis,
required to analyse empirical issues in finance. It provides applications in asset pricing, investment,
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 datasets and examples.

1
FN3142 Quantitative finance

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 reading:

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 (2004) Efficient Market Hypothesis and Forecasting,
International Journal of Forecasting, 20(1).
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, you are also expected to work through the learning activities and sample
examination questions provided in the subject guide. If you find it difficult to answer a given
learning activity, 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 last year’s and those of previous years,
i.e. candidates had to answer three out of four questions. The main difference compared to the
pre-2020 years, however, is that due to the Covid-19 situation, candidates completed the
examination online, over the course of 6 hours (there was an expected/recommended time of 3
hours). This way candidates had access to all usual resources, their subject guide, and essentially
any material available on the internet. This also offered an uncontrolled communication among
candidates, too. This was in stark contrast with previous formats: In previous years, a calculator

2
 Examiners’ commentaries 2021

was allowed to be used when answering questions and it had to comply in all respects with the
specification given in the Admission Notice; otherwise no additional help was allowed.

What the examiners are looking for

In a good answer to a quantitative question, you 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 your workings are shown. In a good answer to a qualitative question, you are expected to
produce an answer which presents appropriate concepts and empirical evidence. You are
furthermore expected to get your points across in a direct, structured, and concise fashion.

Due to the special environment, another important aspect of what examiners paid attention to was
that candidates had to answer the questions concisely and using their own words. Having access to
all the material mentioned above, many candidates ended up simply copy–pasting complete
sentences from the subject guide, using the same structure etc. These answers were heavily
discounted. Further, candidates (probably) spent more time on the examination than the suggested
3 hours, and ended up providing unnecessarily long answers to essay questions. These were again
heavily discounted by the examiners.

Steps to improvement
You must study the subject guide thoroughly. If you manage to work your way through all
questions and examples, you will then properly understand the course material. This means
also attempting all questions from the self-study sections and from past examination papers
all by yourself without ‘cheating’ by looking at the provided answers. After you have
completed the questions, you should carefully compare your answer with the provided
answer.
Once you have completed the subject guide you should read more widely on each topic.
Wider reading gives you a stronger appreciation of theory and empirical evidence, and will
enable you to take a more critical, analytical approach to qualitative examination questions.
Some key further readings are listed above.
Consult the Examiners’ commentaries.
Practise questions from past papers. As the examination draws closer, practise under time
pressure, remembering that you probably only have approximately 60 minutes per question
in which to write your answer.

Examination revision strategy

Many candidates are disappointed to find that their examination performance is poorer than they
expected. This may be due to a number of reasons, but one particular failing is ‘question
spotting’, that is, confining your examination preparation to a few questions and/or topics which
have come up in past papers for the course. This can have serious consequences.

We recognise that candidates might 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 available on the VLE. You should read
the syllabus carefully and ensure that you cover sufficient material in preparation for the

3
FN3142 Quantitative finance

examination. Examiners will vary the topics and questions from year to year and may well set
questions that have not appeared in past papers. Examination papers may legitimately include
questions on any topic in the syllabus. So, although past papers can be helpful during your 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. We strongly advise you not to adopt this strategy.

4
 Examiners’ commentaries 2021

Examiners’ commentaries 2021

FN3142 Quantitative finance

Important note

This commentary reflects the examination and assessment arrangements for this course in the
academic year 2020–21. 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 – Zone A

Candidates should answer THREE of the following FOUR questions.

Question 1

Denote the price process of Bitcoin by P , and consider (T + 1) consecutive

end-of-year price observations denoted by P0 , P1 , . . . , PT . Denote the simple net
Pt − Pt−1
return series by Rt ≡ and the logarithmic return series by
Pt−1
rt ≡ log Pt − log Pt−1 for t = 1, . . . , T . Assume that returns are independently and
identically distributed over time.

(a) Under the assumption that Bitcoin’s annual logarithmic returns are normally
distributed with a mean of µ and a variance of σ 2 , derive a formula for the
difference between the log expected simple gross return and the expected log
return:
log(E[1 + Rt ]) − E[rt ].
(15 marks)
T
1 X
b≡
(b) Consider the standard average estimator µ ri of the mean rate of
T i=1
return. Show that this estimator is unbiased in the sense that E[µ]
b = µ, and
derive its standard deviation, σ[µ],
b as a function of the model parameters.
(20 marks)
(c) Assume instead that you observe Bitcoin prices more frequently, in particular
N times per year (at equal distance from each other) for T years; for example,

5
FN3142 Quantitative finance

monthly observations would mean N = 12. Consider now the simple average
estimator of the annualised log return µ based on these N × T return
observations, and let us denote it by µ̃. Show that the standard deviation of
this estimator, σ[µ̃], does not depend on the frequency N .
(15 marks)
(d) Suppose that a researcher tells you that for a given choice of N and T she
estimated the annualised mean Bitcoin return to be µ̃ = 10%. Suppose also that
the annualised volatility of Bitcoin log returns is known to be σ = 45%. What
choice of T would imply that her estimate has a standard deviation of
σ[µ̃] = 1%, i.e. one-tenth the size of the mean estimate? Discuss whether it is
possible to have confidence in the rate of return of Bitcoin with reasonable
accuracy given the available historical data.
(20 marks)

Let us now assume that annual log Bitcoin returns have a non-zero autocorrelation
at lag 1 denoted by parameter ρ, but beyond lag 1 there is zero autocorrelation.

(e) Derive E[µ]

b and σ[µ]
b under this assumption, where µ
b refers to the estimator in
part (b).
(15 marks)
(f ) Discuss how a non-zero ρ affects your conclusion for part (d).
(15 marks)

Approaching the question

(a) The calculations are as follows:

      P 
Pt − Pt−1 Pt ln t 2
E[1 + Rt ] = E 1 + =E = E e Pt−1 = E[ert ] = eµ+σ /2 .
Pt−1 Pt−1

From here:
σ2
log E[1 + Rt ] = µ +
2
and hence:
σ2
log E[1 + Rt ] − E[r] = .
2

(b) One can write: " #

T T
1X 1X 1
E[b
µ] = E ri = E[ri ] = T µ = µ
T i=1 T i=1 T
and:
 2

T
!2 
1 X σ2
Var[b µ − µ)2 ] =
µ] = E[(b E (ri − µ)  =
T i=1
T

σ
µ) = √ .
σ(b
T

(c) The more frequent (for example, monthly) observations ri are now i.i.d. normally
distributed with mean µ/N and variance σ 2 /N , where µ and σ represent the annualised
parameters. Therefore, one can write:
" NT
# NT
1 X 1 X 1 µ
E[b
µ] = E N ri = N E[ri ] = N NT =µ
N T i=1 N T i=1 NT N

6
 Examiners’ commentaries 2021

and:
 2

T 
!2 
2
1 X µ 
= σ
Var[b µ − µ)2 ] =
µ] = E[(b E ri −
T i=1
N T

σ
µ) = √ .
σ(b
T
which indeed does not depend on N , only on T .
√
(d) With σ = 0.45, we have σ/ T = 0.01 if and only if T = 452 = 2,025. What this means is
that there is huge statistical uncertainty about whether the mean return estimated in the
sample period for which assets traded is close to the true mean or whether it is a fluke due
to sampling variation. Bitcoin in particular was invented in January 2009, and so presently
there is no statistical reason to believe its huge returns will continue in the long run.

(e) The estimator of the mean is unchanged. Regarding the effect of correlation on the standard
deviation:
−1
" T # T T
X X X
Var ri = Var[ri ] + 2 Cov[ri , ri+1 ] = T σ 2 + 2(T − 1)ρσ 2 = (T + 2(T − 1)ρ)σ 2
i=1 i=1 i=1

which implies: r
T −1 σ
σ(b
µ) = 1+2 ρ√ .
T T

(f) Based on the results of parts (d)–(e), and without making any calculations, it looks like that
depending on ρ, the variance of the estimator can be above or below of that in the i.i.d.
case. If there is short-term momentum in Bitcoin, then we could expect ρ to be positive and
then we would have even more uncertainty about its true mean. Conversely, if returns
display short-term reversals, then we would expect ρ to be negative and could be more
certain about the true mean.

Question 2

Consider the following AR(2) process:

zt = α0 + α1 zt−1 + α2 zt−2 + εt , (1)

where εt is a zero-mean white noise process with variance σ 2 , and assume

|α1 |, |α2 |, |α1 + α2 | < 1, which together ensure zt is covariance stationary.

(a) Calculate the conditional and unconditional means of zt , that is, Et−1 [zt ] and
E[zt ].
(15 marks)
(b) Let us now set α2 = 0. Calculate the conditional and unconditional variances of
zt , that is, Vart−1 [zt ] and Var[zt ].
(15 marks)
(c) Keeping α2 = 0, derive the autocovariance and autocorrelation functions of this
process for all lags as functions of the parameters α1 and σ.
(20 marks)

Suppose now that α2 6= 0, and let us denote the autocovariance at lag k by

γk ≡ Cov[zt , zt−k ].

7
FN3142 Quantitative finance

(d) Using equation (1), write down a recursive formula for γk , i.e. express γk as a
function of γk−1 , γk−2 and the model parameters.
(15 marks)
(e) Apply this recursive formula for k = 1 and k = 0, and explain how to solve for
the whole autocovariance function {γk }k≥0 . Note: No need to derive the exact
values!
Hint: think about what γ−1 and γ−2 mean.
(20 marks)
(f ) Can a linear transformation of the zt process be represented by an AR(1)
process? That is, do appropriate κ, δ0 , and δ1 constants exist such that the
process defined as yt ≡ zt + κzt−1 satisfies

yt = δ0 + δ1 yt−1 + t ,

where t is a zero-mean white noise process? Explain.

(15 marks)

Approaching the question

(a) The conditional expectation is simply:

Et−1 [zt ] = α0 + α1 Et−1 [zt−1 ] + α2 Et−1 [zt−2 ] + Et−1 [εt ] = α0 + α1 zt−1 + α2 zt−2 .

Taking unconditional expectations of (1), we obtain:

E[zt ] = α0 + α1 E[zt−1 ] + α2 E[zt−2 ] + E[εt ] = α0 + α1 E[zt ] + α2 E[zt ]

which implies:
α0
E[zt ] = .
1 − α1 − α2

(b) If α2 = 0, the conditional variance is simply given by

Vart−1 [zt ] = Vart−1 [εt ] = σ 2 .

Taking the unconditional variance of (1), we obtain:

Var[zt ] = α12 Var[zt−1 ] + Var[εt ]

which implies:
σ2
Var[zt ] = .
1 − α12

(c) Denote the autocovariance function for lag k by γk , and the autocorrelation function by ρk .
Then, using (1), we can write:

σ2
γ1 ≡ Cov[zt , zt−1 ] = Cov[α0 + α1 zt−1 + εt , zt−1 ] = α1 Var[zt−1 ] = α1 .
1 − α12

On the other hand, for all k ≥ 2 we have:

γk ≡ Cov[zt , zt−k ] = Cov[α0 + α1 zt−1 + εt , zt−k ] = α1 Cov[zt−1 , zt−k ] = α1 γk−1 .

From here, we obtain for all k ≥ 0 that:

σ2
γk = α1k and ρk = α1k .
1 − α12

8
 Examiners’ commentaries 2021

(d) We can repeat the above step in the general AR(2) case:
γk = Cov[zt , zt−k ] = Cov[α0 + α1 zt−1 + α2 zt−2 + εt , zt−k ]
= α1 Cov[zt−1 , zt−k ] + α2 Cov[zt−2 , zt−k ]
= α1 γk−1 + α2 γk−2 .

(e) Writing it down for k = 1, we obtain

γ1 = α1 γ0 + α2 γ−1 = α1 γ0 + α2 γ1
where we used that γ−1 = Cov[zt , zt+1 ] = γ1 . For k = 0, the same calculation gives:
γ0 = α1 γ−1 + α2 γ−2 = α1 γ1 + α2 γ2 .
Finally, for k = 2, it simply yields :
γ2 = α1 γ1 + α2 γ0 .
These are three equations and three unknowns. Solving this set of equations yields γ0 , γ1 ,
and γ2 . After having those, we can use the general recursive formula to obtain the rest of
the autocovariance function.

(f) Adding κzt−1 to both sides, (1) becomes:

yt = zt + κzt−1 = α0 + (α1 + κ)zt−1 + α2 zt−2 + εt
= α0 + (α1 + κ) (zt−1 + κzt−2 ) +(α2 − (α1 + κ)κ)zt−2 + εt (2)
| {z }
yt−1

so the right-hand side is linear in yt−1 iff α2 − (α1 + κ)κ = 0, i.e. κ solves:
0 = κ2 + α1 κ − α2 .
But this equation has no real roots as long as α12 + 4α2 < 0. Otherwise there are either 1 or
2, and hence one can find a suitable κ.

Question 3

Consider two unbiased forecasts f1,t and f2,t of the mean-zero quantity yt+h .
Denote the individual errors by ei,t+h ≡ yt+h − fi,t , i = 1, 2, the mean-square
forecast errors by σi2 ≡ E[e2i,t+h ] for i = 1, 2, and the correlation between the
individual errors by ρ ≡ Corr[e1,t+h , e2,t+h ].

Consider the linear forecast combination

fc,t ≡ λf1,t + (1 − λ)f2,t where λ ∈ [0, 1],
and denote the error associated with the combination by ec,t+h ≡ yt+h − fc,t .

(a) Compute the mean forecast error, E[ec,t+h ], and the mean-squared forecast
error, E[e2c,t+h ], of the combination forecast.
(20 marks)
(b) By minimising the mean-square-error loss of the combination show that the
optimal weights are functions of the accuracy of each prediction and the
correlation parameter, and given by
σ22 − ρσ1 σ2
λ∗ = . (2)
σ12 + σ22 − 2ρσ1 σ2

(20 marks)

9
FN3142 Quantitative finance

(c) Assuming the forecasts are equally accurate, i.e. σ1 = σ2 = σ, derive the
optimal weights and the mean-squared forecast error of the combined forecast.
(20 marks)
(d) Compute the range of the correlation parameter ρ for which the optimal
combined forecast displays diversification benefits, that is, its mean-squared
forecast error, E[e2c,t+h ], is lower than the mean-squared errors of the original
forecasts, σ12 and σ22 . Explain your results.
(20 marks)
(e) Imagine that you have generated a forecast with the optimal weights calculated
above, but then it turns out that the projection has poor performance in the
sense of having a low R2 in a regression of the forecasted variable on a constant
and the forecast. What would be the properties of a data-generating process
that would lead to such a result?
(20 marks)

Approaching the question

(a) We have:
E[ec,t+h ] = E[yt+h − fc,t ] = E[λyt+h + (1 − λ)yt+h − λf1,t − (1 − λ)f2,t ]
= E[λe1,t+h + (1 − λ)e2,t+h ]
=0
while:
E[e2c,t+h ] = Var[λe1,t+h + (1 − λ)e2,t+h ] = λ2 σ12 + (1 − λ)2 σ22 + 2λ(1 − λ)ρσ1 σ2 . (*)

(b) Differentiating (*) w.r.t. λ and rearranging, one obtains (2).

(c) Setting equal accuracy, λ∗ = 1/2. Plugging this optimal weight back into the objective, after
some tedious algebra the general solution is:
1 2
E[e2c,t ] = σ (1 + ρ).
2

(d) Given that |ρ| ≤ 1, we have that:

1 2
σ (1 + ρ) ≤ σ 2
2
for all ρ and equal if and only if ρ = 1.

(e) If a variable of interest exhibits little persistence then even an optimal forecast will not
generate a high R2 . For example, say:
Yt+1 = φYt + εt+1
where εt+1 ∼ N (0, 1) with φ ≈ 0. Then even an optimal forecast will not be good since Yt
does not exhibit much predictable variation.

Question 4

Assume that daily returns are conditionally normally distributed and given by
i.i.d.
rt+1 = µ + σt+1 νt+1 , νt+1 ∼ N (0, 1),
where the time increment between t and t + 1 is one day, µ is constant, and σt+1
denotes an estimate as of time t for the conditional standard deviation one day
ahead.

10
 Examiners’ commentaries 2021

(a) The 1-day Value-at-Risk at the critical level α, VaRα

t+1 , is defined as

P r[rt+1 ≤ VaRα
t+1 ] = α. (3)

Show that the exact formula for VaR at the α critical level and 1-day horizon is
given by
−1
VaRα t+1 = µ + σt+1 Φ (α), (4)

where Φ is the standard normal cumulative density function.

Note: The definition in equation (3) follows the notation of the subject guide.
Using an alternative definition based on losses leads to a different VaR
formula, which is also accepted if correct.
(20 marks)

(b) Describe the historical simulation, the RiskMetrics (also known as exponentially
weighted moving average), and the GARCH normal approaches to measuring
Value-at-Risk in your own words. Discuss their benefits and shortcomings.
(30 marks)
α
(c) The expected shortfall ESt+1 at the critical level α and 1-day horizon can be
defined as
α
ESt+1 = Et [rt+1 | rt+1 ≤ VaRαt+1 ].

Using the VaR formula from part (a), derive the following formula for the 1-day
expected shortfall at critical level α:
σt+1
α
ESt+1 =µ− ϕ(Φ−1 (α)),
α
where ϕ is the standard normal probability density function.
Hint: If z follows a standard normal distribution, z ∼ N (0, 1), we have
ϕ(A)
E[z | z ≤ A] = − .
Φ(A)
(25 marks)

(d) Discuss the benefits and shortcomings of using expected shortfall as a risk
measure instead of Value-at-Risk.
(25 marks)

Approaching the question

(a) We can write:

Varαt+1 − µ
 
rt+1 − µ
α = P r[rt+1 < VaRα
t+1 ] = P r <
σt+1 σt+1
α
VaRt+1 − µ
 
= P r zt+1 <
σt+1
α
VaRt+1 − µ
 
=Φ
σt+1

where Φ is the cumulative density function of a standard normal variable. This implies that:
−1
VaRα
t+1 = µ + σt+1 Φ (α).

11
FN3142 Quantitative finance

(b) Historical simulation in value at risk (VaR) analysis is a procedure for predicting the value
at risk by constructing/simulating the cumulative distribution function (cdf) of asset returns
over time. Unlike parametric VaR models, historical simulation does not assume a
particular distribution of the asset returns but instead assumes that the returns over the
past m periods were i.i.d. from some unknown distribution F – if this is true, then we can
use the empirical distribution of these returns to estimate F . It is extremely easy to
implement. However, there are a couple of shortcomings of historical simulation. Or, for
example, historical simulation applies equal weight to all returns of the whole period; this is
inconsistent with the diminishing predictability of data that are further away from the
present.
The simplest possible model based on the normal distribution is one that assumes that
returns are i.i.d. N (µ, σ 2 ). Then the last m observations are used to estimate µ and σ and
the forecast VaR is given as:
VaRα t+1 = µ
b+σbΦ−1 (α).
This method is very restrictive: it assumes i.i.d. normally distributed returns, making it
more restrictive than historical simulation which requires only the first of these assumptions.
The RiskMetrics model is a more sophisticated method that allows the conditional variance,
and possibly the conditional mean, to vary over time. One would be a GARCH-type of
model. Alternatively, RiskMetrics, or exponentially weighted moving average (EWMA)
assumes:
2
σt+1 = λσt2 + (1 − λ)rt2
where λ = 0.94 for daily and 0.97 for monthly returns. These parameters were found by JP
Morgan, the developers of RiskMetrics, to work well for a variety of asset returns at the
given frequency. Its advantage over GARCH-type models is that it does not require any
parameters to be estimated, at the cost of reduced flexibility and tractability – for example,
there is no finite long-run average volatility.

(c) From the properties of the normal distribution we know that:

ϕ(A)
E[z | z ≤ A] = −
Φ(A)
where ϕ is the density function of the standard normal. From its definition:
α
ESt+1 = Et [rt+1 | rt+1 ≤ VaRα t+1 ]

VaRαt+1 − µ
 
= µ + σt+1 Et zt+1 zt+1 ≤
σt+1
α
VaRt+1 − µ
 
ϕ
σt+1
= µ − σt+1  .
VaRαt+1 − µ
Φ
σt+1
Then using the result from part (a):
ϕ(Φ)−1 (α) σt+1
α
ESt+1 = µ − σt−1 =µ− ϕ(Φ−1 (α)).
Φ(Φ−1 (α)) α

(d) VaR is not a sub-additive risk measure, i.e. VaR of a portfolio can be higher than the sum of
VaRs of the individual assets in the portfolio. Therefore, it does not promote diversification.
The lack of sub-additivity implies that VaR is not a coherent measure of risk. This problem
is caused by the fact that VaR is a quantile on the distribution of profit and loss and not an
expectation. For a sub-additive risk measure such as expected shortfall, portfolio
diversification always leads to risk reduction. Moreover, expected shortfall is an alternative
to value at risk that is more sensitive to the shape of the loss distribution in the tail of the
distribution. This also means that ES requires a more precise estimate for the tail of the
distribution, making ES more complicated to calculate than VaR.

12
 Examiners’ commentaries 2021

Examiners’ commentaries 2021

FN3142 Quantitative finance

Important note

This commentary reflects the examination and assessment arrangements for this course in the
academic year 2020–21. 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 – Zone B

Candidates should answer THREE of the following FOUR questions.

Question 1

(a) What does serial correlation of a stochastic process mean? Explain in your own
words, and give an example of a process with zero serial correlation and an
example of a process with positive serial correlation.
(15 marks)
(b) Consider a sequence of coin toss gambles that you begin with a starting wealth
of W0 = 100 dollars. In each period, if the coin comes up heads you win 1
dollar, and if the coin comes up tails you lose 1 dollar. The probability of heads
(ph ) and tails (pt ) are equal, ph = pt = 1/2. Your wealth after playing the
gamble n ≥ 1 times is denoted by Wn . Your 1-period dollar return (also called
gains and losses) process is defined as Rn = Wn − Wn−1 for all n ≥ 1.
Suppose m coin tosses have already taken place, where m ≥ 0. What are your
k-period-ahead conditional expectations of the 1-period dollar return and of
your wealth?
What are the autocovariance functions of your wealth and dollar return
processes?
(20 marks)
(c) Imagine the equity price of Apple generated the same patterns as your wealth
with the coin tosses in part (b). Would the return process generated by
investing in Apple shares be consistent with the efficient market hypothesis?
(15 marks)

13
FN3142 Quantitative finance

(d) Suppose that for another (not fair) coin the probability of heads and tails are
ph = 0.75 and pt = 0.25, respectively. Compute the expected dollar return on
wealth k periods ahead, and the serial correlation of your dollar return process
for this new coin toss.
(20 marks)
(e) Suppose the equity price of Apple generated the same patterns as your wealth
with the unfair coin tosses in part (d). Would the return process on Apple
shares be consistent with the efficient market hypothesis now?
(15 marks)
(f ) Suppose that Apple announced this morning that its profit from last quarter has
dropped by 10% compared to the previous quarter due to a lower number of
iPhones sold. If Apple’s closing price today is up by 1% compared to yesterday,
can we conclude that this is evidence against the efficient market hypothesis?
(15 marks)

Approaching the question

(a) For a weak sense stationary process the autocorrelation function is defined as:

Cov(Xn , Xk ) E[(Xn − µ)(Xk − µ)]

R(k) = = .
Var(Xn ) σ2
A process displays serial correlation if it has non-zero autocorrelation for τ = n − k > 1.
Intuitively, a white noise process has no serial correlation. Each draw of a white noise
process is a draw from a random distribution, i.e. the observations in each period are
unrelated to each other. Or a series of coin tosses.

(b) What we want to compute is En−1 [Wn ], E[Wn ], En−1 [Rn ], E[Rn ]. In the general case with
ph = p, we can simply write:

Em [Rm+k ] = E[Rm+k ] = ph × 1 + pt × (−1) = p − (1 − p) = 2p − 1

since coin tosses are i.i.d. From here:

" k
#
X
Em [Wm+k ] = Em Wm + Rm+i = Wm + (2p − 1)k.
i=1

The autocovariance of the return process, again due to the i.i.d. nature, is 0 at all k ≥ 1:
γk = Cov[Rm , Rm+k ] = 0, whereas:
2
γ0 = Var[Rm ] = E[Rm ] − (E[Rm ])2 = 1 − (2p − 1)2 = 4p(1 − p).

To work out the autocovariance of the wealth, we write

"m #
X
Var[Wm ] = Var Ri = m Var[Ri ] = 4mp(1 − p)
i=1

and: " #
k
X
Cov[Wm , Wm+k ] = Cov Wm , Wm + Rm+i = Var[Wm ] = 4mp(1 − p).
i=1

Substituting in p = 1/2 gives the solutions to part (b): Em [Rm+k ] = E[Rm+k ] = 0,

Em [Wm+k ] = Wm , Var[Rm ] = 1, and Var[Wm ] = Cov[Wm , Wm+k ] = 1.

(c) Yes. The expected return is zero, and returns are serially uncorrelated. This is perfectly in
line with efficient markets.

14
 Examiners’ commentaries 2021

(d) See the steps above.

(e) Efficient markets imply the expected return can be non-zero but should be serially
uncorrelated unless we observe some market frictions. For example, in the presence of
transaction costs or microstructure noise, there can be low serial correlation that cannot be
traded away. Anyhow, this is not the case here – autocovariance was zero.

(f) No, for (at least) two reasons. First, the drop may be positive news if a larger fall in profits
was expected. Second, there may be other news released on the day that affect Apple’s
returns; for example, a consumer confidence or macro announcement may have coincided
with Apple’s profit announcement.

Question 2

Consider a process Yt that resembles a MA(1) process except for a small change:

Xt = (−1)t ut + δut−1 ,
i.i.d. 2
where ut ∼ N (0, σu ) and 0 < δ < 1 constant.

Reminder: (−1)t = 1 if t is an even number, and (−1)t = −1 if t is odd.

(a) Find Et [Xt+1 ], Et [Xt+2 ], and E[Xt ]. Pay attention to t being odd or even.
(20 marks)
(b) Find Var[Xt ].
(20 marks)
(c) Derive the autocovariance function and the autocorrelation function.
(20 marks)
(d) Explain what covariance stationarity means, and relate it to your findings in
parts (a), (b), and (c).
(20 marks)
(e) Can the Xt process be represented by an MA(1) process? That is, does a
i.i.d.
zt ∼ N (µ, σz2 ) series exist such that

Xt = zt + δzt−1

with appropriate constants µ, σ, and α? Explain.

(20 marks)

Approaching the question

(a) Since Et [ut+1 ] = 0, we can write:

Et [Xt+1 ] = Et [(−1)t+1 ut+1 + δut ] = (−1)t+1 Et [ut+1 ] + δut = δut .

Moreover, since Et [ut+s ] = 0 for all s ≥ 1, we can write:

Et [Xt+2 ] = (−1)t+2 Et [ut+2 ] + δ Et [ut+1 ] = 0.

Finally, since E[ut ] = 0 for all t, we can write:

E[Xt ] = E[(−1)t ut + δut−1 ] = (−1)t E[ut ] + δ E[ut−1 ] = 0.

15
FN3142 Quantitative finance

(b) We have:
Var[Xt ] = E[(Xt − E[Xt ])2 ] = E[Xt2 ]
so:

Var[Xt ] = E[Xt2 ] = E[((−1)t ut + δut−1 )2 ]

= E[((−1)t ut + δut−1 )((−1)t ut + δut−1 )]

= E[(−1)2t u2t + δ 2 u2t−1 + 2(−1)t δut ut−1 ]

= E[u2t + δ 2 u2t−1 + 2(−1)t δut ut−1 ].

Since Cov[ut , ut−1 ] = E[ut ut−1 ] = 0, the cross-product is zero, and we get:

Var[Xt ] = E[u2t + δ 2 u2t−1 ] = σu2 + δ 2 σu2 = (1 + δ 2 )σu2 .

(c) The autocovariance at lag 1 is given by:

γ1 = Cov[Xt , Xt−1 ] = E[(Xt − E[Xt ])(Xt−1 − E[Xt−1 ])]

= E[Xt Xt−1 ]

= E[((−1)t ut + δut−1 )((−1)t−1 ut−1 + δut−2 )]

= E[(−1)t δu2t−1 ]

= (−1)t−1 δσ 2
(
δσ 2 if t is odd
=
−δσ 2 if t is even

since all other cross-products have zero expectation. Similarly, the autocovariance at lag 2 is
given by:

γ2 = Cov[Xt , Xt−2 ] = E[(Xt − E[Xt ])(Xt−2 − E[Xt−2 ])]

= E[Xt Xt−2 ]

= E[((−1)t ut + δut−1 )((−1)t−2 ut−2 + δut−3 )]

=0

since all cross-products have zero expectation.

(d) A process is covariance stationary if (i) E[Xt ] does not depend on t; (ii) Var[Xt ] does not
depend on t; (iii) for all k and t then Cov[Xt , Xt−k ] is finite and depends on k only. This
process is not covariance stationary since Cov[Xt , Xt−k ] depends on t: it is δσ 2 if t is odd
and −δσ 2 if t is even. The unconditional means and variances of each of the Xt s are finite
and do not depend on t but that is not enough.

(e) No. A short response could be that an MA(1) process is covariance stationary, whereas the
Xt process above is not; however close to full marks were only shown when candidates
derived the terms properly.

Question 3

(a) For a given loss function L, an optimal forecast is obtained by minimising the
conditional expectation of the future loss, conditional on the information
available:
∗
Ybt+h|t ≡ arg min E [L(Yt+h , yb) | Ft ] .
y
b

16
 Examiners’ commentaries 2021

Given the quadratic loss function L(y, yb) = (y − yb)2 , show that the optimal
forecast is the conditional mean.
(15 marks)
(b) Consider another quadratic loss function, L0 (y, yb) = a + b(y − yb)2 . Derive the
optimal forecast in this case and discuss the effect of taking a linear
transformation of the mean-square-error loss function on the optimal forecast.
(15 marks)
(c) Describe in your own words how one can test forecast optimality with a
Mincer–Zarnowitz regression.
(20 marks)

Consider now a forecast Ybt∗ of a variable Yt+h . Imagine you have 100 observations of
Ybt∗ and Yt+h , run the regression

Yt+h = α + β Ybt∗ + εt+h ,

and obtain the following results:

Parameter Estimates:

α −0.50
β 3.55

Variance–Covariance Matrix of Parameter Estimates:

α β
α 0.30 0.20
β 0.20 0.55

(d) Can forecast optimality, in the Mincer–Zarnowitz sense, be tested with the
information provided above? Which statistical test would be implemented?
(20 marks)
(e) Can an economic forecast be ‘optimal’ but have poor forecasting power (in
terms of a low R2 from a regression of the forecasted variable on a constant and
the forecast)? If so, give an example, or explain in your own words why not.
(15 marks)
(f ) Can an economic forecast fail to be ‘optimal’ and still forecast well (in terms of
a high R2 from a regression of the forecasted variable on a constant and the
forecast)? If so, give an example, or explain in your own words why not.
(15 marks)

Approaching the question

(a) Substituting in the quadratic loss function, we have:

∗
Ybt+h|t ≡ arg min E[(Yt+h − yb)2 | Ft ].
y
b

Differentiating, the FOC is:

0 = −2 E[(Yt+h − yb) | Ft ]
which is equivalent to:
yb = E[Yt+h | Ft ].
Notice that this indeed corresponds to a minimum since the SOC is satisfied as the second
derivative is 2, positive.

17
FN3142 Quantitative finance

(b) Substituting in the quadratic loss function, we have:

∗
Ybt+h|t ≡ arg min E[a + b(Yt+h − yb)2 | Ft ].
y
b

Differentiating, the FOC is:

0 = −2b E[(Yt+h − yb) | Ft ]
which is equivalent to:
yb = E[Yt+h | Ft ]
as before. The additive constant a does not make a difference, however, b needs to be
positive because otherwise the second derivative would be 2b < 0, i.e. the FOC solution
would give a maximum, and indeed there would be no minimum.

(c) Mincer–Zarnowitz Test: forecast errors should be unforecastable on the basis of information
available at the time the forecast is made. A Mincer–Zarnowitz regression is one where the
dependent variable is regressed on the forecast:

Yt+1 = β0 + β1 Ybt+1|t + ut+1 .

Then one should test:

H0 : β0 = 0 and β1 = 1 vs. H1 : β0 = 0 ∪ β1 = 1.

The above regression can be generalised to include other elements of the time-t information
set. Such a regression is called an ‘augmented MZ’ regression.

(d) Yes, for the joint test of H0 : β0 = 0 and β1 = 1 we need to use the variance–covariance
matrix and the estimates, all given above. This is an F test, with parameters 2 (number of
restrictions) and 100 − (2 + 1) = 97, which takes the number of observations into account.
The F statistic ends up being 10.393, way above the threshold F statistic that is around 3,
so we would reject the joint null hypothesis. Or can be tested by a χ2 test.

(e) Yes. If a variable of interest exhibits little persistence, then even an optimal forecast will not
generate a high R2 . For example, if:

Yt+1 = φYt + εt+1

where εt+1 ∼ N (0, 1) with φ ≈ 0, then even an optimal forecast will not be good since Yt
does not exhibit much predictable variation.

(f) Yes. Suppose the variable of interest exhibits high persistence and we make a forecast that
is close to but not exactly equal to the optimal forecast. In this case the forecast will be
suboptimal but will still generate a high R2 . For example, if:

Yt+1 = 0.80Yt + εt+1

∗ ∗
the optimal forecast is Yt+1 = 0.80Yt , but the forecast Yt+1 = 0.75Yt will be suboptimal and
it will still be performing well.

Question 4

Assume that daily returns are conditionally normally distributed and given by
i.i.d.
rt+1 = µ + σt+1 νt+1 , νt+1 ∼ N (0, 1),

where the time increment between t and t + 1 is one day, µ is constant, and σt+1
denotes an estimate as of time t for the conditional standard deviation one day
ahead.

18
 Examiners’ commentaries 2021

(a) The 1-day Value-at-Risk at the critical level α, VaRα

t+1 , is defined as

P r[rt+1 ≤ VaRα
t+1 ] = α. (1)

Show that the exact formula for VaR at the α critical level and 1-day horizon is
given by
−1
VaRα t+1 = µ + σt+1 Φ (α), (2)

where Φ is the standard normal cumulative density function.

Note: The definition in equation (1) follows the notation of the subject guide.
Using an alternative definition based on losses leads to a different VaR
formula, which is also accepted if correct.
(20 marks)

(b) Describe the historical simulation, the RiskMetrics (also known as exponentially
weighted moving average), and the GARCH normal approaches to measuring
Value-at-Risk in your own words. Discuss their benefits and shortcomings.
(30 marks)
α
(c) The expected shortfall ESt+1 at the critical level α and 1-day horizon can be
defined as
α
ESt+1 = Et [rt+1 | rt+1 ≤ VaRαt+1 ].

Using the VaR formula from part (a), derive the following formula for the 1-day
expected shortfall at critical level α:
σt+1
α
ESt+1 =µ− ϕ(Φ−1 (α)),
α
where ϕ is the standard normal probability density function.
Hint: If z follows a standard normal distribution, z ∼ N (0, 1), we have
ϕ(A)
E[z | z ≤ A] = − .
Φ(A)
(25 marks)

(d) Discuss the benefits and shortcomings of using expected shortfall as a risk
measure instead of Value-at-Risk.
(25 marks)

Approaching the question

(a) We can write:

Varαt+1 − µ
 
rt+1 − µ
α = P r[rt+1 < VaRα
t+1 ] = P r <
σt+1 σt+1
α
VaRt+1 − µ
 
= P r zt+1 <
σt+1
α
VaRt+1 − µ
 
=Φ
σt+1

where Φ is the cumulative density function of a standard normal variable. This implies that:
−1
VaRα
t+1 = µ + σt+1 Φ (α).

19
FN3142 Quantitative finance

(b) Historical simulation in value at risk (VaR) analysis is a procedure for predicting the value
at risk by constructing/simulating the cumulative distribution function (cdf) of asset returns
over time. Unlike parametric VaR models, historical simulation does not assume a
particular distribution of the asset returns but instead assumes that the returns over the
past m periods were i.i.d. from some unknown distribution F – if this is true, then we can
use the empirical distribution of these returns to estimate F . It is extremely easy to
implement. However, there are a couple of shortcomings of historical simulation. Or, for
example, historical simulation applies equal weight to all returns of the whole period; this is
inconsistent with the diminishing predictability of data that are further away from the
present.
The simplest possible model based on the normal distribution is one that assumes that
returns are i.i.d. N (µ, σ 2 ). Then the last m observations are used to estimate µ and σ and
the forecast VaR is given as:
VaRα t+1 = µ
b+σbΦ−1 (α).
This method is very restrictive: it assumes i.i.d. normally distributed returns, making it
more restrictive than historical simulation which requires only the first of these assumptions.
The RiskMetrics model is a more sophisticated method that allows the conditional variance,
and possibly the conditional mean, to vary over time. One would be a GARCH-type of
model. Alternatively, RiskMetrics, or exponentially weighted moving average (EWMA)
assumes:
2
σt+1 = λσt2 + (1 − λ)rt2
where λ = 0.94 for daily and 0.97 for monthly returns. These parameters were found by JP
Morgan, the developers of RiskMetrics, to work well for a variety of asset returns at the
given frequency. Its advantage over GARCH-type models is that it does not require any
parameters to be estimated, at the cost of reduced flexibility and tractability – for example,
there is no finite long-run average volatility.

(c) From the properties of the normal distribution we know that:

ϕ(A)
E[z | z ≤ A] = −
Φ(A)
where ϕ is the density function of the standard normal. From its definition:
α
ESt+1 = Et [rt+1 | rt+1 ≤ VaRα t+1 ]

VaRαt+1 − µ
 
= µ + σt+1 Et zt+1 zt+1 ≤
σt+1
α
VaRt+1 − µ
 
ϕ
σt+1
= µ − σt+1  .
VaRαt+1 − µ
Φ
σt+1
Then using the result from part (a):
ϕ(Φ)−1 (α) σt+1
α
ESt+1 = µ − σt−1 =µ− ϕ(Φ−1 (α)).
Φ(Φ−1 (α)) α

(d) VaR is not a sub-additive risk measure, i.e. VaR of a portfolio can be higher than the sum of
VaRs of the individual assets in the portfolio. Therefore, it does not promote diversification.
The lack of sub-additivity implies that VaR is not a coherent measure of risk. This problem
is caused by the fact that VaR is a quantile on the distribution of profit and loss and not an
expectation. For a sub-additive risk measure such as expected shortfall, portfolio
diversification always leads to risk reduction. Moreover, expected shortfall is an alternative
to value at risk that is more sensitive to the shape of the loss distribution in the tail of the
distribution. This also means that ES requires a more precise estimate for the tail of the
distribution, making ES more complicated to calculate than VaR.

20