Risk Theory

Examining the methods for measuring Market Risk

By

A NDREAS T HEODOULOU

Department of Mathematics
U NIVERSITY OF B RISTOL

BSc Mathematics with Statistics Project (20cp)

M AY 2018
 A CKNOWLEDGEMENT OF S OURCES

or all ideas taken from other sources (books, articles, internet), the source of

F the ideas is mentioned in the main text and fully referenced at the end of
the report. All material which is quoted essentially word-for-word from other
sources is given in quotation marks and referenced. Pictures and diagrams copied
from the internet or other sources are labelled with a reference to the web page,book,
article etc.
Signed: Andreas Theodoulou Dated:01/05/2018

SIGNED: A NDREAS T HEODOULOU....... .... .......................... ............... DATE:

01/05/2018 ..........................................

i
 TABLE OF C ONTENTS

Page

List of Tables iii

List of Figures iv

1 Introduction 1
1.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Project Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

2 Theory 5
2.1 Financial Asset Returns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.1.1 Log-returns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.1.2 Portfolio Returns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.1.3 Volatility clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.1.4 Leptokurtic distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.1.5 Asymmetry and leverage effect . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.1.6 Uncorrelated Daily returns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2 Risk Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2.1 Value at Risk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2.2 Expected Shortfall . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.2.3 Evaluation of VaR and ES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2.4 VaR and ES Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.3 Historical Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.4 Parametric Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.4.1 Volatility Forecasting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.4.2 Exponential Weighted Moving Average . . . . . . . . . . . . . . . . . . . . . 14
2.4.3 GARCH model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.4.4 GJR-GARCH . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.5 Backtesting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.5.1 Backtesting VaR - Unconditional Coverage testing . . . . . . . . . . . . . . . 17
2.5.2 Backtesting ES - Unconditional test . . . . . . . . . . . . . . . . . . . . . . . 18

ii
3 Data & Methodology 21
3.1 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.2 Programming . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.3 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.3.1 General Methodology and Hypotehses . . . . . . . . . . . . . . . . . . . . . . 23
3.3.2 VaR and ES Methods and Backtesting . . . . . . . . . . . . . . . . . . . . . . 25
3.4 Methods Ranking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

4 Results 27
4.1 Value at Risk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
4.1.1 95% Value at Risk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
4.1.2 99% Value at Risk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
4.2 Expected Shortfall . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
4.2.1 95% Expected Shortfall . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
4.2.2 99% Expected Shortfall . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
4.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

5 Conclusion 37
5.1 Recommendations for further research . . . . . . . . . . . . . . . . . . . . . . . . . . 38

A Derivation of parametric method equations 39

A.1 Value at Risk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
A.1.1 Normal Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
A.1.2 Student’s t Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
A.1.3 Expected Shortfall . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
A.1.4 Normal Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40

B Matlab Code 43
B.1 Matlab Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

Bibliography 55

L IST OF TABLES

TABLE Page

3.1 List of the Stock Indices, representing the portfolios to be tested . . . . . . . . . . . . . 22

iii
3.2 Sample statistics of the returns of our "portfolios" . . . . . . . . . . . . . . . . . . . . . . 24

4.1 Ranking statistics of 95%-VaR estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

4.2 Unconditional Coverage test LR results for 95%-VaR estimates during the Crisis period 29
4.3 Unconditional Coverage test LR results for 95%-VaR estimates during the Post-Crisis
period . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
4.4 Ranking statistics of 99%-VaR estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
4.5 Unconditional Coverage test LR results for 99%-VaR estimates during the Crisis period 30
4.6 Unconditional Coverage test LR results for 99%-VaR estimates during the Post-Crisis
period . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
4.7 Ranking statistics of 95%-ES estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
4.8 Unconditional test, Z-statistic for 95%-ES estimates during the Crisis period . . . . . 32
4.9 Unconditional test Z-statistic results for 95%-ES estimates during the Post-Crisis period 32
4.10 Ranking statistics of 99%-ES estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
4.11 Unconditional test Z-statistic for 99%-ES estimates during the Crisis period . . . . . . 33
4.12 Unconditional test Z-statistic results for 99%-ES estimates during the Post-Crisis period 34

L IST OF F IGURES

F IGURE Page

2.1 Autocorrelation of squared daily S&P 500 returns January 1, 2010-December 31, 2010
[4] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.2 Histogram of daily S&P 500 returns and the normal distribution January 1, 2010-
December 31, 2010 [4] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.3 Autocorrelation of daily S&P 500 returns January 1, 2010-December 31, 2010 [4] . . 8
2.4 Value at Risk from the normal distribution return probability distribution [4] . . . . . 9
2.5 Expected Shortfall in terms of the Probability density function of returns [12] . . . . . 10

3.1 S&P500 graph of daily prices. Source: Google . . . . . . . . . . . . . . . . . . . . . . . . 24

iv
 HAPTER
1
C
I NTRODUCTION

1.1 Background

As pointed out by Dowd, "Everything changes and changes can be good or bad for those affected
by them. Change therefore leads to risk, the prospect of gain or loss." [7, p.3] Consequently, this
holds true for any financial institution and in order to survive in the long-term they need to
deal with them. They must assess the risks they currently bear, decide on the risks they want
to bear, and change their exposures as the landscape changes so that they bear the risks they
want to. The risks a firm faces can be generally divided into the following: Business risks, Market
risks, credit risks, liquidity risks, operational risks, legal risks.[7, p.3-4] This project will focus on
Market risk1 and the measures currently used in the industry to interpret it.
The Value at risk (VaR) measure is probably the most important measure for quantifying
financial risk2 in the past 20 years. In the meantime, we have seen the rise in popularity for
Expected Shortfall (ES). Below, I aim to give a brief overview to the problem of measuring market
risk, its history and its importance.
Work on internal models by a number of major financial institutions had already started
in the late 1970s to help understand the bank’s exposure across the entire trading portfolio. In
particular, the first VaR report originated when the chairman of J.P. Morgan asked his staff to
give him, after the close of trading each day, a one-page report indicating risk and potential
losses over the next 24 hours across the bank's entire trading portfolio.[21, p.307] This problem
of course had by no means an easy solution as often a financial institution's portfolio depends
on hundreds, or even thousands, of market variables . In about 1990, his subordinates came up

1 Market Risk is the risk of losses owing to movements in the level or volatility of market prices
2 Financial risk is the risk which relate to possible losses owing to financial market activities such as loss occurred

as a result of interest-rate movements or defaults of financial obligations

1
CHAPTER 1. INTRODUCTION

with the following answer: "We are X percent certain that we will not lose more than V dollars
in time T", where the value V followed from the Value at Risk measure that was invented. The
measure was groundbreaking at the time as it essentially aggregated all the risks of a portfolio
into a single measure. [1, p183]
In fact, shortly after the fall of the world famous Barings bank at the time, the Basel committee
in 1998 gave the option to financial organisations to report their capital requirements using
the Value at Risk measure, which since then has been the popular approach for many financial
organisations. [1, p.266] [13] Even though, VaR has been heavily studied, and several models
were used in the financial world in the meantime, almost all of them seems that failed to predict
the severe financial crisis of the 2007. This seems to be one of the reasons the Basel III accord to
be introduced in 2019, has decided to shift its requirements from the Value at Risk measure to
the Expected Shortfall.[9] [5] Therefore, through this project I aim to examine the best methods
to estimate the VaR and ES measures. We will also evaluate the future shift from VaR to the ES
measure as the solution to capturing market risk and discuss whether VaR could still remain a
useful measure in the industry.

1.2 Project Overview

We will begin with the Chapter 2 of the project by introducing the theory around the material that
is going to applied later on through the project. We start by introducing the main characteristics
of the financial asset returns and defining our portfolio returns. We then introduce the two
risk measures to be studied, the Value at Risk and the Expected shortfall, mathematically and
qualitatively. From there on we discuss the challenge of estimating them and introduce the
methods to be studied, namely the Historical Method, the Parametric method, and the Volatility
forecasting methods used to aid the parametric method. Lastly, we review one Backtesting
method for each of the measures, namely the Unconditional Coverage test and the Unconditional
test for VaR and ES respectively.
In Chapter 3, we introduce the data to be used in applying the methods, explain the specifics
our methodology and our reasoning behind it. We first introduce the 10 portfolios which we will
try to forecast. Then we explain the programming used to aid the application of the methods in our
VaR and ES forecasts. Moreover, we explain our general methodology around these simulations
as well as specifically how each method is going to be applied, express the hypotheses we make
and what we expect to review from these simulations. Lastly, we introduce the method that will
be used for ranking the VaR and ES methods.
In Chapter 4, we introduce and discuss our results. These consist of the backtesting test
statistics of the 95% and 99% VaR and ES estimates and our ranking statistics used to evaluate
our models. We perform forecasts on two periods, one during the financial crisis of 2007-2008 and
the other from 2013 onwards, which we refer to as the Post Crisis period. We then discuss our

2
 1.2. PROJECT OVERVIEW

results for each situation. Concluding, we give an overall view on our results and conclude on the
best model given these results.

3
 HAPTER
2
C
T HEORY

2.1 Financial Asset Returns

Market risk is caused by movements in financial asset prices or equivalently asset returns.
Therefore, we begin by defining returns and then give an overview of the characteristics of typical
financial asset returns. These will provide part of the intuition based on which we will form our
models.

2.1.1 Log-returns

In this project we will follow log-returns, instead of simple returns, as this follows the usual
method in the literature. The main reasons for this is that they preserve better the lower bound
of zero on the prices and allow to easily calculate the compound return at the K-day horizon by
simply summing the daily returns.[4] From now on we will refer to log returns as just returns.

R t+1 = ln(S t+1 ) − ln(S t )

where R t+1 is the daily log return at time t+1, S t is the final price at which a security is
traded at time t, and t refers to a specific market day

2.1.2 Portfolio Returns

Consider a portfolio of n assets. Then the value of a portfolio at time t, VPF,t is the weighted
average of the asset prices, Rt, using the current holdings of each asset, N i as weights is
VPF,t = N
P
i =1 N i S i,t . The (log) return of the portfolio between day t+1 and day t is then defined as

R PF,t+1 = ln(VPF,t+1 ) − ln(VPF,t ) (2.1)

5
CHAPTER 2. THEORY

Note that we assume that the portfolio value on each day includes the cash from accrued
dividends and other asset distributions.

2.1.3 Volatility clustering

As first noted by Mandelbrot, volatility clustering is that "large changes tend to be followed by
large changes, of either sign, and small changes tend to be followed by small changes." The effect
can be seen more clearly in the Figure 2.1. This results in much stronger autocorrelations of the
squared returns than the returns, for shorter time lags. [21] As squared returns are used in the
calculations of volatility1 , the changing dependence of volatility over time can be captured by
statistical models, which will be introduced later.

Figure 2.1: Autocorrelation of squared daily S&P 500 returns January 1, 2010-December 31, 2010
[4]

2.1.4 Leptokurtic distributions

The distribution of daily returns tends to follow distributions that are more leptokurtic than the
standard Gaussian distribution. This means that the distributions tend to be more peaked around
zero and have more small positive and fewer small negative returns, also known as fatter tails.
Fatter tails mean a higher probability of large losses (and gains) than the normal distribution
would suggest. These characteristics can be shown at Figure 2.2 where a histogram of the daily
S&P 500 return data with the normal distribution is imposed. [4]
1 Please note that the volatility term is simply the term used in finance for standard deviation.

6
 2.2. RISK MEASURES

Figure 2.2: Histogram of daily S&P 500 returns and the normal distribution January 1, 2010-
December 31, 2010 [4]

2.1.5 Asymmetry and leverage effect

Asymmetry of magnitudes in upward and downward movements of asset returns is typical in

equity markets. In particular, the volatility response to a large positive return is considerably
smaller than that to a negative return of the same magnitude. This asymmetry is sometimes
referred to as a leverage effect.[21] The reasoning behind this characteristic is based on economic
principles, which are beyond the scope of this project. We will model the leverage effect in some
of our models.

2.1.6 Uncorrelated Daily returns

The autocorrelation of daily returns is generally around 0. In other words, returns are almost
impossible to predict using past data. Also, for short time daily conditional mean of returns is
very close to 0. Therefore, focusing just on volatility modelling of the returns distribution for
this project and assuming conditional mean is constant, seems to be a reasonable simplification.
Figure 2.3 shows the correlation of daily S&P 500 returns with returns lagged from 1 to 100 days.
[4]

2.2 Risk Measures

2.2.1 Value at Risk

VaR summarizes the worst loss over a target horizon that will not be exceeded with a given level
of confidence. More formally, VaR describes the quantile of the projected distribution of gains

7
CHAPTER 2. THEORY

Figure 2.3: Autocorrelation of daily S&P 500 returns January 1, 2010-December 31, 2010 [4]

and losses over the target horizon.[11] Mathematically, the q-%VaR for the 1-day ahead return,
satisfies the following equation:

q
P(R PF,t+1 < −V aR t+1 ) = q (2.2)

Note that VaR depends on the following 3 parameters, which from now on will be indicated as
follows:

q
• q, quantile: The probability to lose more than V aR t+1 in the specified period.

• t, time period: t is assumed to be the present day, so t+1 indicates the one-period VaR

• R PF,t+1 : The return distribution. This is to be calculated using information available in

period t. The specific methods will follow in the later chapters.

Figure 2.4 illustrates the V aR 0t+.01

1 from a normal distribution of returns. Multiplying the
V aR 0t+.01
1 by the total value of the portfolio will give you the worst loss over a day with a probability
of 99%.

2.2.2 Expected Shortfall

The Expected shortfall, also know as conditional Value at Risk (CVaR), is a risk measure that
accounts for the magnitude of large losses as well as their probability of occurring.[4, p.33] More
specifically it tells us the expected value of tomorrow loss, conditional on it being worse than
the VaR. The Expected Shortfall measure aggregates this information into a single number by

8
 2.2. RISK MEASURES

Figure 2.4: Value at Risk from the normal distribution return probability distribution [4]

computing the average of the tail outcomes weighted by their probabilities. Mathematically ES is
defined as follows:
q q
ES t+1 = −E t [R PF,t+1 | R PF,t+1 < −V aR t+1 ] (2.3)

where the negative signs in front of the expectation and the VaR are needed because the ES
and the VaR are defined as positive numbers.

In the figure 2.5 below, the shaded area represents the losses that exceed the VaR. Cumulative
probability in the shaded area is equal to 1-q. ES or CVaR, can be represented by the average of
the shaded area.

2.2.3 Evaluation of VaR and ES

Value at risk captures an important aspect of risk, namely how bad things can get with a
certain probability, q.[4] Furthermore, it is easily communicated and easily understood.VaR does,
however, have some drawbacks. Most important, extreme losses are ignored. The VaR number
only tells us that q% of the time we will get a return below the reported VaR number, but it says
nothing about what will happen in those q% worst cases. A second major criticism of VaR is that
the measure is not subadditive. Subadditivity holds that adding the risk of Asset A and the risk
of Asset B will not result in an overall risk that is greater than the sum of the two risks together.
This can undermine the principle of diversification, often important when managing assets. [12]
Lastly, the Value at Risk measure has developed a wide variety of applications through the years,

9
CHAPTER 2. THEORY

Figure 2.5: Expected Shortfall in terms of the Probability density function of returns [12]

including portfolio optimization. For the purpose of this project, we will only focus on its capability
of capturing the market risk.
Now Expected shortfall, satisfies both of the above VaR disadvantages, however it is not
perfect either. ES often requires a large number of observations to generate a reliable estimate,
and it is more sensitive to estimation errors than VaR [12]. Moreover, the reliability of ES depends
substantially on the accuracy of the tail model used. It should also be noted that ES is based on
an average loss beyond VaR, and thus it is not a measure of the most extreme potential loss.
Overall, it is important to remember that both measures are only an estimate. This means
that different samples will lead to different results. Therefore, the processes of calculating the
risk measures are as important as the number itself. [11] We will now review the processes of
deriving these risk measures, and more specifically the returns distribution needed to calculate
them.

2.2.4 VaR and ES Implementation

While VaR and ES are easy and intuitive concepts, their measurement are a challenging statistical
problem. The existing models for calculating VaR employ different methodologies, however all
follow a common general structure; 1) Mark-to-market the portfolio, 2) Estimate the distribution
of portfolio returns, 3) Compute the VaR of the portfolio. [14] The main challenge is due to the
difficulty of modelling the evolution of a portfolio containing hundreds of assets[19] [10].
The existing models can be classified into 3 broad categories [18]

10
 2.3. HISTORICAL SIMULATION

• Parametric models (e.g. RiskMetrics and GARCH)

• Nonparametric models (e.g. Historical Simulation and the Hybrid model)

• Semiparametric models (e.g. Extreme Value Theory and Conditional Autoregressive Value
at Risk)

In this project we will focus on parametric and nonparametric methods as one can extend to
the semi-parametric once the former are understood. We refer the interested reader to some of
the most popular semiparametric methods to [14]. Moreover, the number and types of approaches
to VaR estimation within parametric and nonparametric methods is growing exponentially [14],
therefore we will try to focus on the most popular ones. These will be the Historical Simulation
and the Risk Metrics, Garch and GJR models for the parametric method. Also, note that the
approaches considered only cover the univariate approaches, which consider portfolio-level risk,
whereas multivariate approaches, which consider individual asset level risk and aggregate them
into a portfolio-level risk, are out of the scope of this paper. The reader can refer to [4] Chapters 7
to 9 for the use of multivariate approaches. Lastly, we will only focus on estimating the next day
VaR. The next k-day VaR estimate can follow from there, and the methods are introduced in [4]
Chapter 8. In the following parts, estimated quantities will be denoted by a hat (ˆ).

2.3 Historical Simulation

Historical Simulation (HS) is one of the most used methods for VaR and ES estimation. This
approach drastically simplifies the procedure for computing the VaR and ES, since it doesn't
make any distributional assumption about portfolio returns. [14]
We will define the historical simulation following Christoffersen at [4, p,22]. Let today be day
t, consider a portfolio of n assets and consider the value of this portfolio and its corresponding
returns. Using today's portfolio holdings, one can consider its historical asset prices and compute
the history of the past m daily "pseudo" portfolio returns, call it {R PF,t+1−τ }τm=1 ,that would be true
if today's portfolio allocation had been the same in the past. The HS technique simply assumes
that the distribution of tomorrow's portfolio returns, {R PF,t+1−τ }τm=1 , is well approximated by
the empirical distribution of the past m observations, or simply the histogram of, {R PF,t+1−τ }τm=1 .
The q%-VaR is then calculated as 100qth percentile of the sequence of past portfolio returns.
Mathematically, we write

q
V
á aR t+1 = −P ercentil e({R PF,t+1−τ }τm=1 , 100q)

In practice, we simply sort the returns R PF in ascending order and choose the VaR to be the
number at the qth percentile of this list. As the VaR typically falls in between two observations,
linear interpolation can be used to calculate the exact number.

11
CHAPTER 2. THEORY

Similarly, the estimate of ES using HS can be calculated by taking the average of the returns
that were less that VaR in the specified time window.

1X τ
q
ES
à
t+1 = ri
τ i=1

Where τ is the number of times the returns were below VaR and r i the value of these returns.
The reasons for the popularity of the HS method are generally its convenience to implement
and it's model free nature. [4] [6] The former implies that the method makes no distributional
assumption about the returns, and the benefit of this is that model building approaches can be
misleading if the model is poor. However, from the model-free advantage follows one of its biggest
drawbacks, which is that it assumes that the distribution of portfolio returns doesn't change
within the window. One of the implications of this is that its performance may be affected by
the choice of sample length, m, mainly due to the volatility clustering periods which are difficult
to identify. [14] Also, this assumption means the HS is likely to result in VaR estimates with
high variance due to the discreteness of extreme returns. To see why, assume a rolling window of
180 days and that today's return is a large negative number. It is easy to predict that the VaR
estimate will jump upward, because of today's observation. [14]

2.4 Parametric Method

The parametric approach is put simply a model building approach. It aims to parameterize
and engulf in a single model, the characteristics of financial asset returns by using appropriate
statistical distributions and techniques. Once this is done, the calculations of VaR and ES can
follow from Equations 2.2 and 2.3, given the distribution of the model has an analytic solution.
As the complexity of a parametric model can always be furthered and in this particular topic
could cover a research paper in itself, for the scope of this project we will keep to the simpler
approaches and try to refer the reader where appropriate.
We start by considering the a generic model for individual asset returns, R t , following [4,
p11-12]. Based on the asset return characteristics reviewed in Chapter 2.1, the following generic
form is employed

R t+1 = µ t+1 + σ t+1 z t+1 , with z t+1 ∼ i.i.d D(0, 1)

The random variable z t+1 is an innovation term, which we assume is identically and in-
dependently distributed (i.i.d.) according to the distribution D(0,1), which has mean zero and
variance one. The conditional mean of the return, E t [R t+1 ], is thus µ t+1 , and the conditional
variance, E t [R t+1 − µ t+1 ]2 , is σ2t+1 . As mentioned in Subsection 2.1.6, for simplicity we will assume
conditional mean of returns to be 0. For longer horizons, the risk manager may want to estimate
a model for the conditional mean. Considering Equation 2.1, allowing for a dynamic variance

12
 2.4. PARAMETRIC METHOD

model, to be introduced later, and applying the assumption on conditional mean, we can now
define a generic model for the portfolio returns as follows,

R PF,t = σPF,t z t with z t+1 ∼ i.i.d D(0, 1) (2.4)

We will first consider the appropriate models for the innovation term and then review the
volatility models used to estimate the variance of the portfolio. The innovation term models will
be the Normal distribution and the Student’s t-distribution.
Although the Normal distribution does not satisfy the financial data characteristics as
explained earlier, it remains a popular method because of its simplicity and the quasi-maximum
likelihood GARCH result. [14] For more information on the result the reader can refer to [14,
p.17] and [4, p.75]. The use of the Student’s t-distribution is one of the methods used capture
the non-normality of the financial data. [21, p.144] However although it captures the fatter tails,
one should note that it still has limitations as it is less concentrated around the mean. For more
methods used to capture the non-normality of the financial asset returns the reader can refer to
[4] Chapter 6.
Under a Normal distribution of the innovation terms, N(0, 1), the VaR and ES estimations
can be calculated through the following equations.

q −1
V aR t+1 = −σ t+1 Φ (q) (2.5)
á 

q φ(Φ−1 (q)
ES t+1 = −σ (2.6)
à  t+1
1− q
where φ and Φ are the probability density function and cumulative density functions of N(0, 1)
respectively.

If the innovation terms belong to a Student’s t-distribution, with ν degrees of freedom, then
the VaR and ES are estimated as follows,

ν−2
r
q −1
V aR t+1 = −σ
á
t+1 t ν (q) (2.7)
ν


1 −1 2
g ν t−
ν (q) ν + (t ν (q)) ν−2
r
q
ES
à
t+1 = −σ t +1 (2.8)
ν−1 ν

1− q
where t ν and g ν is the cumulative density function and probability density function of the
Student’s t distribution respectively.

ν−2
Please note that the distribution is scaled by the ν
factor, or simply divided by its variance,
in the equations above in order to satisfy the unit variance assumption of the innovation’s
distribution. We introduce a full derivation of the above Equations 2.5, 2.6, 2.7 and 2.8 in the
Appendix A. Moreover, one of the methods of estimating the degrees of freedom for the Student’s

13
CHAPTER 2. THEORY

t-distribution with variance 1 and mean 0, is using the maximum likelihood method. The method
chooses the number of degrees of freedom that maximizes the log-likelihood function of the our
t-distribution over our chosen time frame. [4, p.129-130]
As part of the parametric method, we will now review volatility forecasting models.

2.4.1 Volatility Forecasting

The focus of Volatility Forecasting is to introduce the models to be used for forecasting tomorrow
's variance. Again, these will aim to capture some of the characteristics of the financial asset
returns and more specifically in this case the squared returns. For example, as mentioned in
Subsection 2.1.3, the varying autocorrelation of squared returns over time is a characteristic to be
considered. The models to be reviewed will be the Risk Metrics, the GARCH model and the GJR
model, through which we will follow [4] at Chapter 4. Please note that the models defined below
can be used for calculating both the variance of an individual asset and the portfolio as a whole.
For now, we will define their individual asset form. Their portfolio wide form follows by simply
replacing the individual asset variables with their equivalent portfolio wide ones. Moreover, these
models can be used to approach the volatility forecasting problem in a multivariate approach.
This essentially means that each asset can be modelled individually and are then aggregated into
a portfolio wide forecast, by considering their interdependence with methods like correlations
and copulas modelling. As for the purpose of risk measurement at a portfolio level the univariate
approaches are considered to be sufficient [4, p.68], we will only consider the univariate ones and
refer the interested reader to [4] Chapters 7 to 9 for the multivariate approaches.

2.4.2 Exponential Weighted Moving Average

The Exponential Weighted Moving Average (EWMA) model, also known as RiskMetrics model,
was first introduced by J.P Morgan’s for market risk management in 1996. It is a simpler approach
to volatility foracasting, where the weights on past squared returns decline exponentially as time
moves backwards. More specifically,
∞
σ2t+1 = (1 − λ) λτ−1 R 2t+1−τ
X
for 0 < λ < 1
τ=1

In this project we are going to use the following version of the equation, which can be found
by separating the first term of the sum in the above equation,

σ2t+1 = λσ2t + (1 − λ)R 2t (2.9)

RiskMetrics advantages consist of the following. First, it is broadly consistent in capturing

the volatility clustering. Second, the model is quite simple to implement as it contains only
one unknown parameters, namely λ. Moreover, for a large number of assets, the parameter
estimates have been found to be quite similar across assets, and it is therefore suggested to

14
 2.4. PARAMETRIC METHOD

simply set λ = 0.94 for every asset for daily variance forecasting. Lastly, little data is needed to
be stored as λτ−1 decays quickly to 0. On the other hand though, it does have drawbacks. For
example, it provides unrealistic longer-horizon forecasts and does not allow for the leverage effect
characteristic of returns. Its shortcomings led us to consider more sophisticated but still simple
models from the GARCH family.

2.4.3 GARCH model

The family of Generalized autoregressive conditional heteroskedasticity (GARCH) models were

introduced Bollerslev (1986), as an extension to the nobel awarded ARCH models suggested by
Engle. The model has been successfully applied to financial data since then. [14]. The GARCH(p,q)
model can defined as follows,

p q
σ2t+1 = ω + α i R 2t+1− i + β j σ2t+1− j
X X
i =1 j =1

where ω, α i and β i are weights to be estimated.

The difference between the generalized ARCH and the ARCH model lies in the addition of
the weighted summation over squared volatility term. This essentially means that the GARCH
model also allows squared volatility to depend on previous squared volatility, σ2t , on top of the
squared return. [2, p.145] For short-term volatility forecasting,the simplest model of the family,
namely GARCH(1,1), is often found to be sufficient.[4, p.78] Therefore for this project we consider
reasonable to only use the GARCH(1,1) model and from now on we will simply refer to it as the
GARCH model. Mathematically, it is defined as follows

σ2t+1 = ω + αR 2t + βσ2t with α + β < 1 (2.10)

Through this model periods of high volatility tend to be persistent. This is because |σ t+1 | has
a chance of being large if either |σ t+1 | is large or σ t is large. The same effect can be achieved
by ARCH models of higher order but lower-order GARCH models tend to capture it much more
persistently.[2, p.145] Moreover, notice here that the RiskMetrics model is a special version of the
GARCH models, if we set α = 1 − λ, β = λ so that α + β = 1 and ω = 0. However the models still
have a major difference. That is the RiskMetrics model does not account for the fact that long-run
average variance tends to be relatively stable over time. [4, p.71] This can be more explicitly seen
be defining the (long-run average) variance, σ to be,

σ2 = E[σ2t+1 ] = ω + αE[R 2t ] + βE[σ2t ]

= ω + ασ2 + βσ2 so that,
σ2 = ω/(1 − α − β)
One can now see that if α + β = 1, as it is true in the RiskMetrics model, then the long-run
variance is not well defined in that model. Moreover, the GARCH model does take into account

15
CHAPTER 2. THEORY

the long-run variance, as it implicitly relies on it. [4, p.71] By solving for ω in the above equation
and substituting into the the GARCH model then,

σ2t+1 = (1 − α − β)σ2 + αR 2t + βσ2 = σ2 + α(R 2t − σ2 ) + β(σ2t − σ2 )

Therefore, one can see that under the GARCH model also depends on the long-run average
variance, on top of the today's squared return and today's variance. Although this difference
might not be as important for the one-day horizon forecast, it does grows in importance for
longer-horizon forecasting. [4, p.71]
Lastly, the unknown coefficients of the model can be estimated through the maximum likeli-
hood estimation method. We are now going to review our last volatility forecasting model which
is an extension of the GARCH model.

2.4.4 GJR-GARCH

As we have seen in Subsection 2.1, the leverage effect is a characteristic we have not yet tried to
capture through our parametric models. The GJR-GARCH model introduced by Glosten et al. in
1993 as an extension of the GARCH model, aims to do this. One can notice that with a θ larger
than 0, the model defined as follows, will capture the leverage effect, [4, p.77]

σ2t+1 = ω + αR 2t + αθ 1 t R 2t + βσ2t (2.11)


1, if R t < 0.
for 1 t =
0, otherwise.

and where, ω, α, β and θ are weights to be estimated.

Now to compute the variance forecast in each of the GARCH models, we first need to estimate
the unknown parameters of our model. For this we will choose the Maximum likelihood Estimation
method, which finds the parameters, α, β, ω and θ , in each case accordingly, by choosing their
value that maximizes the log-likelihood function of the portfolio returns, as defined in Equation
2.4, over our chosen time frame. [4, p.73]
Overall, the family of GARCH models are widely used because of their ability to capture
important characteristics of the returns data and are flexible enough to accommodate specific
characteristics of individual assets. Various extensions to the GARCH model introduced exist,
which aim to capture even more characteristics of the returns data, as indicated by the GJR-
GARCH model. For further extensions the reader can review Chapter 4 Section 5 of [4]. On the
downside, the GARCH models requires a non-linear parameter estimation. Moreover, there are
different methods used to individually evaluate each volatility forecasting models. As one of the
scopes of the project is to evaluate the models as a whole for the measurement of VaR and ES, we
will not review these and the interested reader can refer to [4] Chapter 4 Section 6.

16
 2.5. BACKTESTING

2.5 Backtesting

Although some of the methods introduced for VaR and ES implementation may make sense in
theory, a considerate risk manager would want to evaluate them in practice. Therefore, we will
use backtesting on the methods introduced for this reason but also as a means of evaluating
the most suitable ones. For the Value at Risk measure we will use the Unconditional Coverage
testing,or Proportion of failuers test, as introduced in [4, p.301-304], whereas for the Expected
Shortfall measure we will use the "Test 2: testing ES after VaR" method as recently introduced in
[3].

2.5.1 Backtesting VaR - Unconditional Coverage testing

Consider the the time series of the past portfolio returns, {R PF,t+1−τ }τT=1 , and of the past VaR
q
aR t+1−τ }τT=1 . Then we can construct an indicator function, name it the "hit sequence",
foreacsts, {Vá
which returns 1 if the loss on the t+1 day was greater than the VaR forecast, or simply if VaR
was violated, and 0 if not.

q
1, if R PF,t+1 < −V aR t+1 .
It + 1 =
0, otherwise.

One can see that the hit sequence, { I t+1−τ }τT=1 , allows us to see when the past VaR violations
occurred. Moreover, recall that given the perfect model, the VaR measure promises that the
actual return will only be worse that the VaR forecast q100% of the time. In other words, the hit
sequence of violations should show 1 with a probability of q every next day, should be completely
unpredictable, and therefore distributed independently over time. It essentially provides us with
the general framework under which we can test our models, or an initial null hypothesis, which
can be put formally as follows,

H0 : I t+1 ∼ i.i.d. Bernoulli(p)

This is idea is usually the starting for most VaR backtesting methods. A more intuitive point
of interest at this point could be whether q is close to the fraction of violations obtained by the
VaR model, call it π. This is what the Unconditional Coverage test aims to examine, which can be
described as follows,
H0 : π = q
(2.12)
H1 : π 6= q
Let L(·) denote the likelihood function for an i.i.d. Bernoulli random variable, T0 and T1 be
the number of 0s and 1s in the sample of the hit sequence, and T the total days of the sequence.
Also, note that we can estimate π from π̂ = T1 /T. Then for the sample hit sequence L(π̂) =
(1 − π̂)T0 π̂T1 , and L(q) = (1 − q)T0 q T1 . We can now check the unconditional coverage hypothesis
using a likelihood ratio test,

17
CHAPTER 2. THEORY

LRUC = −2ln[L(q)/L(π̂)] ∼ χ21 (2.13)

Using the Likelihood ratio test, if our likelihood ratio value LRUC is larger than the critical
value of the significance level, from the χ21 distribution, then we reject our null hypothesis,
otherwise we do not reject the null hypothesis. Alternatively through the p-value approach, we
reject the null hypothesis if the p-value is less than the significance level value. The p-value is
calclulated by p − value = 1 − Fχ2 (LRUC ).
1

2.5.2 Backtesting ES - Unconditional test

Expected shortfall has traditionally been more limited in terms of the available backtesting
methods. This is mainly because of the following challenge. A VaR backtesting has two possibilities
for every day, either there is a VaR violation or not. On the other side, for ES the possibilities
every day are infinite, it may exceed VaR by 1%, 50% or 500%. Moreover, if the VaR was exceeded
by 30% on average and it was estimated to have been 20%, this error in estimation will be
much more significant for a thin-tailed distribution than for a fatter-tailed one. This means that
information about the distribution of returns for each day is needed. However, this problem
has been overcome recently when [3] introduced the "unconditional" test or "Test 2" where you
can get approximate test results without providing the distribution information.[16] This was
a very important result as it made the backtesting of ES much simpler. This paper was also
very important as a discovery in 2011 that ES is not elicitable, raised a debate over the risk
management community at the time, on whether ES can be actually backtested. This paper
comes to prove that it can be actually backtested.
Recall that the ES risk measure promises that whenever we violate the VaR , the expected
q
value of the violation will be equal to ES t+1 . Also, let F t+1 denote the real (unknowable) return
q
distribution and P t + 1 the model predictive distribution, used to estimate V aR t+1 . Moreover,
let ES P
t+1,q denote the value of ES when returns are distributed to P t+1 , with q and t as defined
earlier, and ES Ft+1,q the value of ES when returns are distributed to F t+1 . Then given the above,
we can formulate the following hypothesis,

[ q] [ q]
H0 : P t+1 = F t+1 , for all t
(2.14)
H1 : ES P F
t+1,q ≥ ES t+1,q

Now assuming H0 , and with I t as defined earlier, we can define ES as follows,

R PF,t+1 I t+1
· ¸
q
ES t+1 = −E
q
By expanding the expectation and rearranging the equation we can consider the following
test statistic,
T R
1 X PF,t+1 I t
Z2 (R PF,t+1 ) = +1 (2.15)
qT t=0 ES tq+1

18
 2.5. BACKTESTING

In other words our H0 and H1 could have been simply E[Z2 ] = 0 and E[Z2 ] < 0 respectively.
Moreover, the test statistic allows us to formulate our framework for testing our hypothesis
through the p-value approach as follows,

i
simulate independent R PF,t +1 ∼ P t for all t and for all i = 1, ..., M
³ ´
compute Z i = Z R PF,t
i
+1 (2.16)
M ³ ³ ´´
Z i < Z r PF,t
i
X
estimate p= +1 /M
i =1

where M is a suitably large number of scenarios. Given a preassigned significance level φ, the
test is rejected if p < φ.

19
 HAPTER
3
C
D ATA & M ETHODOLOGY

3.1 Data

In this section we will review the data used for the simulations, the details of our methodology in
general as well as for the specific methods, the programming used to run them, and introduce
our way for ranking the performance of the models. The data chosen are the stock indices listed
in Table 3.1, and each essentially represents a single portfolio. We have extracted our data from
Yahoo! Finance because of it is widely used and it’s easily accessible. Specifically, the choice of our
data are 10 stock indices of a pool of countries ranging from developed to developing economies.
This was firstly because of the simplicity that using a single index as a portfolio allows, but at the
same time a relatively representative sample of the movements of a country’s financial assets. As
indicated below each index consists of a relatively high number of stocks, and more importantly
these are the top stocks by market capitalisation1 in the country. For example, even for the Hang
Seng Index, which consists of 50 stocks, they represent about 58% of the market capitalisation of
Hong Kong Stock exchange. [Wikipedia] Secondly, we have deliberately chosen stock indices only
because of the linearity of the stock returns. For non-linear assets, like bonds and derivatives, our
methods do no apply and one has to resort to other ones like the delta-normal or delta-gamma
approach. For the scope of this project we have chosen to exclude these types of portfolios as
linear portfolios tend to be more common and their risk measures can be calculated by simply
separating the portfolio into its linear and the non-linear part and apply the appropriate methods.
Two sources suggested on calculating VaR and ES for non-linear portfolios are Chapter 11 and 12
[4] and [15].

1 Market Capitalisation is the value of a company that is traded on the stock market, calculated by multiplying

the total number of shares by the present share price.

21
CHAPTER 3. DATA & METHODOLOGY

Table 3.1: List of the Stock Indices, representing the portfolios to be tested
Stock Index No. of Stocks Region
Standard & Poor 500 (S&P 500) 500 USA
Financial Times Stock Exchange 100 (FTSE 100) 100 UK
Nihon Keizai Shimbun 225 (Nikkei 225) 225 Japan
EuroNext 100 100 Europe
DAX Performance Index (DAX) 30 Germany
Hang Seng Index (HSI) 50 Hong Kong
ASX 200 200 Australia
Standard & Poor Bombay Stock Exchange Sensitive Index (S&P BSE) 30 India
Bolsa de Valores do Estado de Sao Paulo Index (IBOVESPA) 60 Brazil
Indice de Precios y Cotizaciones (IPC) 35 Mexico

3.2 Programming

For our simulations, and any related numeric results posted in this Chapter we have chose
MATLAB because of its easy of use given the multiple available toolboxes and its widely used
environment. We have specifically used functions from the Risk Management Toolbox and the
Econometric Toolbox, as well as a few of their ideas in the relevant code constructions. Our codes
are listed in Appendix B.

We will now give an insight to the reader to the functions used from the above toolboxes.
For volatility foreacsting, we have first used the garch() and gjr() models, which were used to
create the EWMA, GARCH and GJR models. They allow to input the p and q parameters of the
models, the distribution of our innovation term, and specify the parameters if desirable. For the
EWMA we have therefore used the garch() function with specified parameters. Moreover, we have
used the estimate() function on these models, inserting the log returns. Given these inputs the
function estimates the coefficients of our GARCH family models using the maximum likelihood
estimation (mle) method. It is the method generally used for estimating the coefficients of GARCH
models and finds the value of coefficients that are most probable to fit the data. [4, p.73-74]
If the innovation term distribution is a t-distribution the function also estimates the degrees
of freedom of the distribution using the mle, given its stated mean and variance. Lastly, the
forecast() function was used to provide the forecast of our variance, with equations as described
in 2.1.3.

For the backtesting methods, we have used the varbacktest(), esbacktestbysim() and esback-
test() functions. The varbacktest() functions provides a report of the results various backtesting
methods, including the unconditional coverage test, or proportion of failures test. The esback-
testbysim() and esbacktest() produce again a report of various backtesting methods, including
our chosen one, with their difference being that the esbacktestbysim() allows us to choose the
parameters of the model which will be used for simulations as explained in Equation 2.16.

22
 3.3. METHODOLOGY

3.3 Methodology

3.3.1 General Methodology and Hypotehses

We have split the periods of our estimates into two periods, the "Crisis Period" ranging from 1st
of June 2006 to 1st of August 2009, and the "Post-Crisis Period" ranging from 1st of January 2013
to 1st of January 2018. We provide the sample statistics of the returns of each of our "portfolios"
at the Table 3.2, to give a broader view of how they differ between the two periods. The results in
this table and any table that follows are provided to two decimal places.
As expected, firstly we can notice a significance difference in the volatility, or standard
deviation values, of our returns between the Crisis and the Post-Crisis period. Also the mean
of around 0 is not surprising as discussed earlier in the characteristics of asset returns. The
kurtosis statistic essentially describes how leptokurtic or not a distribution is. The value of the
kurtosis of a standard Normal distribution is 3 and the greater it is the more it tends towards a
t-distribution. As we can see in our returns data, our kurtosis is generally significantly above
3. Moreover, the a positive skewnness implies that a distribution has a right tail that is longer,
or simply more data are concentrated towards the negative values of a distribution, and vice
versa. From there we could potentially argue that the generally positive skewnness of the Crisis
period relative to the Post Crisis, means more large negative returns and leads to the higher
standard deviation. This is roughly what the leverage effect argues, which recalling to section
2.1, it suggests that the volatility response to a large positive return is considerably smaller than
that to a negative return of the same mangitude.Therefore, these are reasons that lead us to test
the following models in the parametric method. The t-distribution to test whether it captures the
leptokurtic aspects of our returns data, the EWMA and GARCH models to check whether they
adapt to changes in volatility in the recent time period, or in technical terms volatility clustering,
and lastly the GJR which on top of that aims to also capture the leverage effect.
On top of the data for the forecasts time periods, our models need to be fed with data before
this in time. For example, the GARCH family of models need an amount of past data in order
to compute their coefficients. As suggested by [4, p.71], for the GARCH models one should use
at least 1,000 daily past observations. Given the year of trading days consists of roughly 250
days this equates to about past 4 years of data. We have chosen to stick to roughly 1,000 daily
past observations for our estimation periods for the following reason. During the Crisis period it
allows to feed the GARCH models with a period of relatively "calm" past data, and then ask it to
forecast "stormy" estimates, while for the Post-Crisis period we feed it with "calm" past data and
ask for "calm" forecasts. We illustrate this more explicitly in the Figure 3.1 for the S&P500 index,
which is a generally fair representative of our 10 portfolios. The reasoning behind this is that we
would like to test how much the GARCH models can stretch their capabilities in the first case,
and at the same time be an appropriate family of models under "normal" conditions.
Lastly, we have chose to test our risk measures at 95% and 99% quantiles. This allows us to

23
CHAPTER 3. DATA & METHODOLOGY

Table 3.2: Sample statistics of the returns of our "portfolios"

Period Index Mean Standard Deviation kurtosis skewness min max
S&P500 0 0.02 7.33 -0.10 -0.09 0.11
FTSE100 0 0.02 7.07 0.00 -0.09 0.09
Nikkei225 0 0.02 8.46 -0.32 -0.12 0.13
Euronext100 0 0.02 8.01 0.15 -0.09 0.10
DAX 0 0.02 8.41 0.32 -0.07 0.11
Crisis
HSI 0 0.03 6.41 0.15 -0.14 0.13
ASX200 0 0.02 5.26 -0.32 -0.09 0.06
S&P BSE 0 0.03 6.73 0.20 -0.12 0.16
IBOVESPA 0 0.03 6.84 0.07 -0.12 0.14
IPC 0 0.02 7.33 0.17 -0.07 0.10
S&P500 0 0.01 5.88 -0.43 -0.04 0.04
FTSE100 0 0.01 5.66 -0.20 -0.05 0.04
Nikkei225 0 0.01 7.62 -0.038 -0.08 0.07
Euronext100 0 0.01 6.66 -0.40 -0.07 0.04
DAX 0 0.01 5.51 -0.37 -0.07 0.05
Post-Crisis
HSI 0 0.01 5.47 -0.31 -0.06 0.04
ASX200 0 0.01 4.57 -0.30 -0.04 0.03
S&P BSE 0 0.01 5.95 -0.39 -0.06 0.04
IBOVESPA 0 0.01 4.96 -0.09 -0.09 0.06
IPC 0 0.01 4.91 -0.19 -0.05 0.04

test more thorouhgly the behaviour of the tail of the distribution and thus determine whether a
t-distribution does prove a better one in these cases. This might also help us to determine the
best distribution for the 97.5% required by the Fundamental Review of the Trading Book (FRTB)
regulation.[8]

Figure 3.1: S&P500 graph of daily prices. Source: Google

24
 3.4. METHODS RANKING

3.3.2 VaR and ES Methods and Backtesting

As stated earlier, the parametric methods and HS method introduced will be used to estimate
the one day VaR and ES. For the parametric method we will test both a Normal Distribution
and a Student’s t for the innovation term, and for the Volatility forecasting part of the method
we will test the EWMA, GARCH and GJR models. For the quantiles of VaR and ES we will test
the 95% and 99% quantiles. For the Historical Simulation method we will chose a time window
of 250 days as past data, which is the most popular time window as mentioned in [17]. For the
Volatility Forecasting models, we will choose a time window of 1000 days, for the estimation of
the model coefficients. These will be estimated by the mle method, and the estimation will be
updated on every day of the forecast. The volatility estimate will be calculated with the equations
introduced in Chapter 2. For the mean estimate in the parametric method, as mentioned earlier
we will simply assume it to be 0. Lastly, given these and the equations defined in Chapter 2 for
parametric and HS estimates, we will calculate our VaR and ES estimates. We will simulate
roughly 1800 forecasts for our 10 "portfolios", introduced in Table 3.1, in the time periods of
"Crisis" and "Post-Crisis" as defined earlier.
To test these methods we will use one backtesting method for each risk measure. For the
VaR we will use the Unconditional Coverage test with a 95% confidence interval and report
the Likelihood Ratio of our methods. We will then compare these to the critical values of a χ21
distribution, which for a 95% confidence interval is 3.841. For the ES we will use the Unconditional
test. We will simulate scenarios from a t-distribution with 15 degrees of freedom, for our t-
distribution parametric methods, as this was roughly the average of the degrees of freedom
calculated by the mle method using the estimate() function in MATLAB. These degrees of
freedom also fall in line with the values successfully tested by [3]. For the normal distribution
parametric method and for HS we will simply simulate the values from a N(0,1) distribution. The
Z-statistics of our backtest will be then reported which will be compared to the critical value of Z
for a 95% confidence level, which is -0.7.

3.4 Methods Ranking

To compare and rank the performance of our methods we will use two methods. Both of these
methods are based on the test statistics produced by our backtests. The first method will be
a measure of the frequency of best estimation, n ω over all our portfolios. In simple terms this
measure will count the times that a method performed best over the rest. Our indicator for best
performance will be the test statistics produced in each case relative to our critical values. In
the VaR case the smallest LR of a method will indicate the method performed best. Therefore
through this measure, we aim to capture which method performed best the most times. However,
the limitation that arises with this is that we do not have any information on how much "well"
did it perform, and we might also miss methods that on average performed well but not the best

25
CHAPTER 3. DATA & METHODOLOGY

in a single portfolio. Therefore, we will also consider a second measure, name it the average
performance measure. In the VaR case this measure will calculate the sum of the Likelihood
P
Ratios, LR, of a single model over all the portfolios. For the ES, this will be the sum of the
P
absolute value of the Z-statistic, | Z |, of a model over all portfolios. We have chosen the absolute
value of Z since the optimal Z lies in the value closest possible to 0, and therefore without the
absolute value our measure would not be representative of performance. Therefore, will allow us
to determine which method performed best on average.
As mentioned previously, apart from the best individual model, we would like to determine
which distribution over the Normal and t-distribution, represents best the asset returns distribu-
tion in the parametric method. For this we will make use of the two model ranking measures
introduced above and adjust them to this case. Therefore for the VaR we will sum the sum of the
LR ratios of a single normal parametric method over the portfolios, over the 3 normal distribution
PP
methods, LR and equivalently for the t-distribution methods. For the ES case we will do the
PP
same with the Z statistics and thus our measure will be, | Z |.
Lastly, we will also consider which methods satisfied our hypothesis testing, and consequently
were below the critical value over all. For the VaR case, as a stated earlier the critical value
P
of a 95% confidence interval test, for a single LR-statistic is 3.841, for the LR is 18.307 and
LR is 124.342. As summing a χ21 distributions n times is equal to a χ2n distribution, the
PP
for
LR are distributed to a χ210 and and χ2100 respectively, and therefore the critical
P PP
LR and
values are derived from there. For ES, we will use the critical value of Z-statistic is -0.7 for a 95%
P PP
confidence interval, where as for the for the | Z | and | Z |, we do not have any reasonable
critical value to use and will therefore just compare their values as explained above.

26
 HAPTER
4
C
R ESULTS

In this Chapter we will report our results from the simulations and discuss them, with the aim of
evaluating our methods. Our results are reported in the 12 tables below. We will first discuss the
results of VaR in each of the two time periods, Crisis and Post-Crisis, move on to the ES shortfall,
and lastly discuss the general picture. In each of the VaR and ES sections we firstly introduce
our Ranking statistics for the 95% and 99% quantiles and then for each period we report our
individual portfolios results for the 95% and 99% quantiles. The ranking statistics table include
our ranking measures as introduced in Section 3.4, where as the individual portfolios results
report the test statistics for each corresponding backtesting method as explained in Subsection
3.3.2. In the ranking statistics tables, we highlight with a grey area the methods that performed
best for each ranking measure, and write with red coloured numbers the methods that are
rejected by the null hypothesis.

4.1 Value at Risk

In this section we post the tables below with the ranking statistics results, in Tables 4.1 and 4.4
and the Unconditional Coverage results for the 95% and 99% VaR results and for the Crisis and
the Post Crisis period, Tables 4.2, 4.3, 4.5 and 4.6

4.1.1 95% Value at Risk

Reviewing our ranking statistics for the Crisis period, we can first determine that EWMA the best
model contender. It outperforms the other models in both the average performance measure and
the frequency of best estimation. The only other fair competitors are the GARCH and the GJR
t-distributed models which have performed relatively well by the frequency of best estimation

27
CHAPTER 4. RESULTS

Table 4.1: Ranking statistics of 95%-VaR estimates

Crisis Post Crisis
P PP P P PP P
LR LR n ω nω LR LR nω nω
HS 93.93 1 2.3 6
EWMA 14.99 3 3.28 4
Normal GARCH 72.38 176.45 0 4 44.11 294.29 0 4
GJR 89.08 1 246.91 0
EWMA 12.76 1 9.09 0
Student’s t GARCH 26.08 139.51 2 5 33.11 375.93 0 0
GJR 100.67 2 333.73 0

measure. The rest of the models seem to be quite far away in terms of the ranking statistics, with
the GARCH normal model in particular achieving no best estimation over our 10 portfolios. For
the Post Crisis period the picture changes slightly. HS becomes the best method according to
both of our ranking statistics. The EWMA remains a strong competitor, especially for the normal
distribution. The rest of the models remain far behind with the worst performing one being the
GJR, especially for the t-distribution.
For the distributions, we have a slightly better performance of the normal one over the t
during the Crisis period, for both of the ranking statistics. During the Post-Crisis period, the
normal distribution seems to be the clear winner, with the t-distribution achieving no best
estimation.
Lastly, in terms of our hypothesis tests, during the Crisis period all of our models but the
EWMA, are rejected at the 95% confidence interval test. More specifically, the EWMA is the
only model that is not rejected for every single of our portfolio, for both the normal and the
t-distributions. On the other hand no single method has been rejected for all of the portfolios. For
the Post Crisis period, the only significant changes is that the HS is not rejected in any of the
portfolios, why the GJR is rejected for all, for both of the distributions. Lastly, under our critical
value both the normal distribution and the t-distribution are rejected for both periods.

4.1.2 99% Value at Risk

Regarding the 99% Value at Risk estimation, our test statistics again rank the EWMA method
first for both the Crisis and the Post-Crisis period. For both periods, the normal EWMA model
performs best in the frequency of best estimation statistic, while the t-distributed one best for
the average performance. Its only fair competitor is the GARCH model, especially for the normal
distribution for both periods. For the Post-Crisis period HS becomes another good competitor,
performing relatively well in both of our ranking statistics. The worst performer over both periods
is the GJR model, which performed relatively worse in both of the test statistics for both of the
periods.
For the distributions, the change form the 95% VaR estmation is that the Normal becomes

28
 4.2. EXPECTED SHORTFALL

Table 4.2: Unconditional Coverage test LR results for 95%-VaR estimates during the Crisis period
Parametric
Portfolio
HS EWMA-N GARCH-N GJR-N EWMA-t GARCH-t GJR-t
S&P 500 18.07 3.75 7.78 3.10 3.75 9.75 1.10
FTSE 100 11.51 2.77 0.31 0.40 1.69 0.55 0.18
Nikkei 225 6.81 1.90 47.99 42.03 1.44 0.20 56.73
Euronext 100 20.14 4.15 6.50 3.12 3.48 6.50 3.12
DAX 15.22 1.44 1.44 3.79 1.44 2.41 3.79
HSI 4.24 0.31 0.13 2.54 0.31 0.05 3.28
ASX 200 7.61 0.08 0.08 12.80 0.11 0.22 14.70
S&P BSE 0.13 0.31 2.27 7.79 0.48 0.68 4.89
IBOVESPA 5.79 0.05 5.79 2.54 0.05 4.99 1.91
IPC 4.41 0.23 0.09 10.97 0.01 0.73 10.97

Table 4.3: Unconditional Coverage test LR results for 95%-VaR estimates during the Post-Crisis
period
Parametric
Portfolio
HS EWMA-N GARCH-N GJR-N EWMA-t GARCH-t GJR-t
S&P 500 0.84 0.84 12.06 88.36 4.06 11.03 111.16
FTSE 100 0.05 0.75 7.519 41.91 0.66 4.78 61.36
Nikkei 225 0.11 0.04 2.89 9.94 0.22 0.54 15.39
Euronext 100 0.00 0.41 6.53 42.91 1.07 6.53 59.41
DAX 0.01 0.04 3.20 19.85 0.20 2.29 44.27
HSI 0.23 0.12 1.04 8.21 0.12 0.39 7.39
ASX 200 0.72 0.48 2.69 6.90 1.51 2.26 8.52
S&P BSE 0.03 0.24 3.46 12.08 0.39 2.95 15.53
IBOVESPA 0.13 0.31 2.27 7.79 0.48 0.68 4.89
IPC 0.18 0.05 2.45 8.96 0.38 1.66 5.81

the best distribution in terms of both our test statics for the Crisis period, while it remains the
better one for the Post-Crisis. Especially, in terms of best frequency estimation it outperforms the
t-distribution by at least 3 times in both periods.
In terms of our hypothesis testing, the EWMA is again proven to be a reliable model, as it
is not rejected for any portfolio for both of the periods. On the other hand, for the Crisis period
we have no model rejected for all portfolios, while for the Post-Crisis the GJR for both of the
distributions is rejected over all the portfolios.

4.2 Expected Shortfall

For the Expected shortfall estimates, our relevant results are presented in the tables below. These
include the ranking statistics, in Tables 4.7 and 4.10 and the Unconditional test Z-statistics for
the 95% and 99% estimations for the Crisis and Post Crisis period, in Tables 4.8, 4.9, 4.11, and

29
CHAPTER 4. RESULTS

Table 4.4: Ranking statistics of 99%-VaR estimates

Crisis Post-Crisis
P PP P P PP P
LR LR nω nω LR LR nω nω
HS 64.66 1 18.22 2
EWMA 7.96 5 9.35 4
Normal GARCH 40.55 138.9 3 8 58.55 246.66 2 6
GJR 90.39 0 178.76 0
EWMA 7.13 0 6.6 0
Student’s t GARCH 49.06 152.46 0 1 123.42 306.25 2 2
GJR 96.27 1 176.23 0

Table 4.5: Unconditional Coverage test LR results for 99%-VaR estimates during the Crisis period
Parametric
Porftolio
HS EWMA-N GARCH-N GJR-N EWMA-t GARCH-t GJR-t
S&P 500 17.16 0.85 5.38 0.29 0.29 0.29 0.09
FTSE 100 4.45 1.01 0.42 10.92 1.01 2.89 10.92
Nikkei 225 4.42 0.36 11.12 11.12 0.36 11.12 11.12
Euronext 100 3.82 0.60 3.30 11.56 0.60 6.04 11.56
DAX 2.66 1.57 1.57 11.48 1.57 5.97 11.48
HSI 15.36 0.95 0.06 3.05 0.95 1.43 11.18
ASX 200 2.66 0.27 0.01 11.48 0.01 0.58 11.48
S&P BSE 1.51 1.84 5.82 7.87 1.83 7.87 5.82
IBOVESPA 5.66 0.49 1.43 11.18 0.49 1.43 11.18
IPC 6.96 0.02 11.44 11.44 0.02 11.44 11.44

Table 4.6: Unconditional Coverage test LR results for 99%-VaR estimates during the Post-Crisis
period
Parametric
Portfolio
HS EWMA-N GARCH-N GJR-N EWMA-t GARCH-t GJR-t
S&P 500 0.03 2.85 25.31 25.31 1.41 25.31 25.31
FTSE 100 1.37 0.83 14.00 25.41 0.42 18.31 25.41
Nikkei 225 4.08 0.22 2.74 24.74 0.15 7.68 24.74
Euronext 100 1.28 0.76 1.26 25.69 0.76 25.69 25.69
DAX 2.00 0.01 4.42 18.39 0.04 14.06 25.49
HSI 3.13 2.31 0.04 5.68 0.55 5.68 5.68
ASX 200 0.82 0.14 3.05 18.35 1.19 8.16 14.03
S&P BSE 3.11 0.22 0.01 13.49 0.04 7.72 13.49
IBOVESPA 1.51 1.84 5.82 7.87 1.84 7.87 5.82
IPC 0.89 0.17 1.90 13.83 0.20 2.94 10.57

30
 4.2. EXPECTED SHORTFALL

Table 4.7: Ranking statistics of 95%-ES estimates

Crisis Post-Crisis
P PP P P PP P
|Z| | Z | nω nω |Z| | Z | nω nω
HS 9.57 0 1.26 3
EWMA 1.77 2 0.45 7
Normal GARCH 2.85 9.67 1 3 2.56 8.27 0 7
GJR 5.05 0 5.26 0
EWMA 1.25 4 1.35 0
Student’s t GARCH 1.76 8.18 2 7 2.65 9.98 0 0
GJR 5.17 1 5.98 0

4.12.

4.2.1 95% Expected Shortfall

Looking at our test statistics, one can determine the EWMA is again the winner for both of the
periods. For the Crisis periods, it’s t-distribution model performs best and its competitors involve
mainly the GARCH model with the t model performing better. Regarding the Post-Crisis period,
the EWMA Normal model has the most frequencies of best estimation by far. The Historical
Simulation again proves to perform better during the Post-Crisis period, becoming the main
competitor to the EWMA normal model. The worst model for the Crisis period is the HS, with the
GJR normal model as a slight second. Both of the models achieve no best estimations. Regarding
the post-crisis period, the GJR model seems to be the overall worst model according to both
of the ranking statistics, while the GARCH model achieved was close next, achieving no best
estimations for both of its distributions.
For our distributions, the t-distribution seems to be the better one for the Crisis period. It
has a lower average performance value and more than double the number of best estimations.
During the post-crisis period, the normal distribution proves to be the better one, achieving 7
best estimations to the 0 of the normal one.
Regarding our hypothesis tests the only rejected method is the HS for the crisis period.
Specifically, it is also rejected for 7 out of the 10 portfolios, while the other methods are not
rejected for any of the portfolios. However, the GJR model in particular seems to have a large
number of positive values for both of the distributions, which indicates an overestimation of
the Expected Shortfall. As for the Post Crisis period, the only difference is that the HS method
performs much better and is not rejected, or otherwise proves reliable, for all of our portfolios.

4.2.2 99% Expected Shortfall

The 99% estimates of the Expected Shortfall estimates, generally have a similar picture with
the 95% ones. According to our ranking statistics, the EWMA outperforms any other model
in the Crisis period, being the only model with best estimates. The worst performing model is

31
CHAPTER 4. RESULTS

Table 4.8: Unconditional test, Z-statistic for 95%-ES estimates during the Crisis period
Parametric
Portfolio
HS EWMA-N GARCH-N GJR-N EWMA-t GARCH-t GJR-t
S&P 500 -1.46 -0.37 -0.60 -0.29 -0.29 -0.52 0.05
FTSE 100 -1.12 -0.31 -0.07 0.22 -0.20 -0.04 0.24
Nikkei 225 -1.25 -0.26 0.96 0.93 -0.11 0.16 1
Euronext 100 -1.30 -0.33 -0.37 0.42 -0.23 -0.31 0.44
DAX -1.10 -0.21 -0.16 0.41 -0.14 -0.17 0.45
HSI -0.67 0.10 0.08 0.33 0.16 0.06 0.39
ASX 200 -0.83 -0.08 -0.04 0.64 0.02 -0.03 0.67
S&P BSE –0.45 -0.03 0.16 0.85 0.03 0.15 0.92
IBOVESPA -0.77 -0.01 -0.39 0.35 0.03 -0.30 0.38
IPC -0.62 -0.07 0.02 0.61 0.04 -0.02 0.63

Table 4.9: Unconditional test Z-statistic results for 95%-ES estimates during the Post-Crisis
period
Parametric
Portfolio
HS EWMA-N GARCH-N GJR-N EWMA-t GARCH-t GJR-t
S&P 500 -0.10 0.05 0.45 0.92 0.24 0.48 0.98
FTSE 100 -0.11 -0.11 0.37 0.73 0.13 0.35 0.83
Nikkei 225 -0.11 -0.00 0.24 0.42 0.08 0.17 0.55
Euronext 100 -0.04 0.05 0.31 0.72 0.16 0.37 0.83
DAX -0.05 -0.02 0.23 0.55 0.11 0.28 0.75
HSI -0.01 0.00 0.11 0.35 0.08 0.14 0.39
ASX 200 -0.20 0.06 0.21 0.36 0.17 0.23 0.41
S&P BSE -0.07 0.05 0.21 0.45 0.13 0.25 0.53
IBOVESPA -0.04 0.10 0.22 0.37 0.14 0.17 0.34
IPC -0.15 -0.01 0.21 0.39 0.11 0.21 0.37

Table 4.10: Ranking statistics of 99%-ES estimates

Crisis Post Crisis
P PP P P PP P
|Z| | Z | nω nω |Z| | Z | nω nω
HS 17.06 0 4.24 2
EWMA 2.67 5 2.24 4
Normal GARCH 6.21 17.66 0 5 4.71 15.78 2 4
GJR 8.78 0 8.83 0
EWMA 2.07 5 10.84 2
Student’s t GARCH 6.84 18.27 0 5 7.75 27.48 0 2
GJR 9.36 0 8.89 0

32
 4.3. DISCUSSION

Table 4.11: Unconditional test Z-statistic for 99%-ES estimates during the Crisis period
Parametric
Portfolio
HS EWMA-N GARCH-N GJR-N EWMA-t GARCH-t GJR-t
S&P 500 -2.89 -0.33 -1.05 -0.12 0.01 -0.03 0.36
FTSE 100 -1.74 -0.42 0.31 1 -0.26 0.68 1
Nikkei 225 -2.29. -0.29 1 1 0.13 1 1
Euronext 100 -1.71 0.28 0.67 1 0.38 0.85 1
DAX -1.48 0.42 0.50 1 0.50 0.82 1
HSI -2.16 -0.32 0.15 0.66 -0.07 0.58 1
ASX 200 -1.63 -0.23 -0.07 1 0.04 0.36 1
S&P BSE -1.39 0.06 1 1 0.20 1 1
IBOVESPA -1.43 0.29 0.46 1 0.36 0.52 1
IPC -1.28 -0.03 1 1 0.12 1 1

the HS one, as it has the highest average performance value. As for the Post Crisis period, the
EWMA model is still the better model, with its best performing being the Normal one, while its
competitors seems to be the HS and the GARCH normal model. The rest of the models have no
best estimation. The worst performing model seems to be again the GJR, underperfoming in both
of the ranking statistics.
For the distributions, in the Crisis period the Normal distribution slightly outperforms the t
one, with its only edge being the slightly less average performance value. In the Post-Crisis period,
it seems to be more clearly the better distribution, outperforming the t distribution significantly
in both of the ranking statistics.
Regarding our hypothesis tests, for the Crisis period the HS proves again to be an unreliable
model, as it’s rejected for all of the portfolios. All other models except the GARCH Normal, are not
rejected by the critical value, however the GJR model seems to again overestimate the ES. For
the Post Crisis period the main difference concerns the improved performance of the HS, which
has only two rejects out of the ten, and no other model is being rejected over all the portfolios.

4.3 Discussion

Given the above results and methods of comparison, the EWMA proves to be the best method
quite clearly for both "calm" periods and "stormy" periods. For the Crisis period this was generally
followed by the GARCH model where as for the Post-Crisis by the HS model. The overall worst
performing model was the GJR model, which performed particularly bad in the Post-Crisis period.
This was followed by the HS method for the Crisis Period and with the GARCH model for the
Post Crisis.
Firstly, the outperformance of the EWMA might sensibly suprise. This is because compared
to the parametric methods it takes the least factors into account. Theoretically, it should not
adapt as quickly to volatility clustering compared to the GARCH model, as its coefficient is fixed,

33
CHAPTER 4. RESULTS

Table 4.12: Unconditional test Z-statistic results for 99%-ES estimates during the Post-Crisis
period
Parametric
Portfolio
HS EWMA-N GARCH-N GJR-N EWMA-t GARCH-t GJR-t
S&P 500 -0.10 -0.56 1 1 -0.11 1 1
FTSE 100 -0.56 -0.25 0.86 1 0.03 0.94 1
Nikkei 225 -0.72 -0.17 0.47 1 0.18 0.71 1
Euronext 100 -0.49 -0.26 0.34 1 -0.03 1 1
DAX 0.04 -0.02 0.54 0.93 0.22 0.87 1
HSI -0.63 -0.41 -0.02 0.62 0.02 0.65 0.69
ASX 200 -0.41 -0.06 0.48 0.93 0.38 0.72 0.87
S&P BSE -0.74 -0.11 0.03 0.84 0.21 0.67 0.87
IBOVESPA -0.35 0.35 0.58 0.66 0.42 0.68 0.65
IPC -0.49 -0.05 0.39 0.85 9.24 0.51 0.81

and it does not account for the leverage effect that the GJR model does. However, on a second
thought this raises concern over the estimation of the coefficients of the GARCH family of models.
Firstly, the mle method has different calculation variations and it is regarded by no means as an
optimal method for coefficient estimation. Secondly, another concern around the mle might be the
amount data it receives. Given that it needs more than 1,000 data to perform well, this means
that after all our coefficients might be slow to adapt. Therefore, it a large coefficient estimation
error using the mle might be a factor for the EWMA proving to be a better method. Moreover,
another limitation over the mle that we have is whether using the minimum suggested amount
to achieve our testing of "calm" and "stormy" periods, has proved costly.
On the positive side for the GARCH model, the fact that it has second best and better than
the HS method in the Crisis period, indicates that volatility clustering might indeed be worth
capturing. The HS method did not seem to be able to cope with the higher volatility of the Crisis
period and this might be mainly the reason for performing significantly worse during this period.
This agrees with the theory, as the method relies on a relatively large amount of past data and
therefore, it is slow to adapt to changing circumstances. Even using the least amount of past data
suggested by the Basel committee, the HS proved to perform worse to any other models tested.
Therefore, we would suggest that the Historical Simulation method is avoided during "stormy"
periods. On the bright side, under normal conditions Historical method seems to be a strong one.
This again agrees with the theory and it could have been expected by looking at the figure 3.1.
Regarding the distributions of the parametric methods, we have seen the normal distribution
to overall outperform the t-distribution, with the only case of the t-distribution performing better
being the 95% quantile for the Crisis period for both VaR and ES. This suggests that our data did
not have much fatter tails than a standard normal, and probably much less than a student’s t.
Moreover, our belief for the worse performance of the Student’s t in the Post Crisis period, is that
our returns tend to be more peaked around 0, as suggested by the min and max values in the

34
 4.3. DISCUSSION

Table 3.2. This is one of the limitations of using a Student’s t as described in the theory part of
the project, since it is less peaked around 0 than a normal distribution. Another reason, for the
worse performance of the Student’s t distribution might again lie around the mle, as it is used to
calculate the degrees of freedom of the distribution. Therefore, these led us to suggest the use of a
normal distribution for the innovation term in the parametric method, given our results and the
easier nature of use of the Normal distribution. However, at the same time we acknowledge that
the Normal distribution has been long empirically rejected as a model for the returns of financial
assets, and therefore suggest this topic, as a field for further research.
Concerning, the use of the ES over the VaR estimates, we have seen that the ES models have
performed much better given the backtests used. This might therefore make the use of ES over
VaR in the future a more reliable method of capturing market risk. On the other hand, we will
hold our limitations on this point as this has been based on a single backtesting method, which
has not been in the industry for too long.
Lastly, for these results, we have seen no significant difference in the performance of our
models over the 95% and 99%, and therefore we suggest that the same model is used for these
quantiles, as well as potentially any quantile in between them.

35
 HAPTER
5
C
C ONCLUSION

The aim of the project as introduced in the beginning was to determine the best methods for
computing VaR and ES. We also aimed on discussing how the two measures differ, and therefore
what the change in regulations from VaR to ES means.

Overall, regarding our models we have seen that going for more complex parameterisations,
like in the GJR model, has not paid out well. On the other hand the relatively simple EWMA has
proved to be consistent over almost all of our portfolios, for both of the time periods, and for both
of our risk measures. Therefore, given the simplicity of the model we suggest that this model
is used for stock portfolios. Lastly, the Student’s t distribution did not seem to prove competent
enough to capture the fatter tails of the asset returns characteristics, and therefore the more
easy to use Normal Distribution is suggested to be used.

Regarding the introduction of the ES over VaR for regulation purposes is expected to have
some effects on market risk calculations as explained. Firstly, as determined in Chapter 4
much more methods for ES were proved to be reliable under hypothesis testing compared to
VaR, and therefore we expect the ES to capture market risk generally more accurately over
different institutions. Moreover, theoretically as discussed in Chapter 2, the ES covers some of
the disadvantage of VaR like the subadditivity and the consideration of extreme losses. Therefore,
we expect it to be a more complete measure. However as explained, it still faces some challenges
itself. Also, the shift to it might pose a challenge to financial institutions in general, due to it
being less widely used for now. Concerning the future of Value at Risk, we expect it to remain
relevant and be used for internal purposes, due to its advantages, its wide use and understanding,
and its various applications developed, as discussed in Chapter 2.

37
CHAPTER 5. CONCLUSION

5.1 Recommendations for further research

As discussed in Chapter 4, we recommend further research on the best distribution fitting the
financial asset returns, as the Normal and Student’s t have proved inadequate in this project.
Moreover, we would suggest further research in the methods for backtesting ES. During our
research for this project we have noticed that backtesting methods on ES have been relatively
scarce and not as solid as for VaR. The method used in this project is currently seen as one of the
best in the literature, however given the short period it has been used, we would encourage that
it is tested further.
Lastly, since the volatility forecasting models with adaptable coefficients have not proved to be
adequate in this project, we encourage further research in volatility forecasting. More specifically,
through our research amongst others, we have came across the [20] article which illustrates a
integrated statistical model with deep neural networks that outperforms volatility forecasting
models such as GARCH. Although this is a field traditionally more relevant to Computer Science,
we encourage the keen risk manager, or statistician, to further research how fields like deep
neural networks and machine learning could be used to improve volatility forecasting, in a period
where they are becoming ever more important.

38
 APPENDIX
A
D ERIVATION OF PARAMETRIC METHOD EQUATIONS

A.1 Value at Risk

A.1.1 Normal Distribution

Let Φ denote the cumulative distribution function of the Standard normal distribution.
Recall the definition of the Cumulative distribution function, F, for a random variable X, over
a value x.

F X (x) = P(X < x)

We start with using the Equation 2.2, used to define Value at Risk.

q
P(R PF,t+1 < −V aR t+1 ) = q

By substituting our assumed model for the Returns of the portfolio distribution as defined in
equation 2.4 we have the following

q
P(σPF,t+1 z t+1 < −V aR t+1 ) = q

Suppose σ is estimated using a volatility forecasting model as introduced in Section 2.4.1,

then since a standard deviation is positive we have
à q !
−V aR t+1
P z t+1 < =q
σ
á PF,t+1

Assuming z t+1 is distributed to a N(0,1) distribution, we can now take the inverse cumulative
function on both sides of the equation and get our VaR equation 2.5 as follows,

39
APPENDIX A. DERIVATION OF PARAMETRIC METHOD EQUATIONS

q
−V aR t+1
= Φ−1 (q)
σ
á PF,t+1
⇐⇒
q −1
V aR t+1 = σ PF,t+1 Φ (q)
á á

A.1.2 Student’s t Distribution

For the Student’s VaR equation we begin from the following step of the Normal distribution
derivation.
à q !
−V aR t+1
P z t+1 < =q
σ
á PF,t+1

ν
Now recall that the variance of a Student’s t distribution with ν degrees of freedom is ν−2
.
Assuming that z t+1 is distributed to a Student’s t distribution
q with variance 1 and mean 0 as
q
ν−2
introduced in Equation 2.4, then it must be that z t ∼ t ν ν
or z t ν−ν 2 ∼ t ν . Multypling the
q
above equation by ν−ν 2 gives,
à q !
ν −V aR t+1 ν
r r
P z t+1 < )=q
ν−2 σ
b PF,t+1 ν−2

Taking the inverse of the Student’s t distribution and rearranging the equation gives the
desired VaR equation 2.7 as follows,

ν
r
q −1
V
á aR t+1 = −σ PF,t+1 t ν (q)
ν−2
á

A.1.3 Expected Shortfall

For the derivation of the Expected Shortfall parametric equations we will follow [2, p.45-46]

A.1.4 Normal Distribution

Assume that z t+1 is distributed to a N(0,1).

We start from the ES equation as defined in Equation 2.3

q q
ES t+1 = −E t [R PF,t+1 | R PF,t+1 < −V aR t+1 ]

Substituting equation 2.4 gives

q q
ES t+1 = −E t [σPF,t+1 z t+1 | σPF,t+1 z t+1 < −V aR t+1 ]
q
By taking the standard deviation out and using that V aR t+1 = Φ−1 q , we have

40
 A.1. VALUE AT RISK

q
ES t+1 = −σPF,t+1 E t [z t+1 | z t+1 < −Φ−1 (q)]

Using the equation of the conditional expectation we have,

1
Z ∞
q
ES t+1 = −σPF,t+1 zφ(z)
1− q Φ−1
⇐⇒
q 1
ES t+1 = −σPF,t+1 [−φ(z)]∞
Φ−1 ( q)
1− q
⇐⇒
q 1
ES t+1 = −σPF,t+1 φ(Φ−1 (q))
1− q
As derivation of the student’s distribution includes similar steps except for the integration
of a student’s t distribution with mean 0 and standard deviation 1, we will stop our derivations
here.

41
 APPENDIX
B
M ATLAB C ODE

B.1 Matlab Code

%% Value−at−Risk and Expected Shortfall Estimation and Backtesting

%
% This code estimates the value−at−risk (VaR) and Expected Shortfall(ES)
%using the following methods, and performs a VaR and ES backtesting
%analysis. The methods are:

% # Historical simulation
% # Exponential weighted moving average (EWMA) − Normal Distribution
% # GARCH(1,1) − Normal Distribution
% # GJR − Normal Distribution
% # Exponential weighted moving average (EWMA) − t−Distribution
% # GARCH(1,1) − t−Distribution
% # GJR − t−Distribution
%
% The estimation methods used, estimate the VaR at 95% and 99%
% confidence levels.

%% Start Function
function [] = VaRestimates(filename, estimationyear, estimationmonth,PortfolioID)

%filename − Data to be imported from excel, 1 column of dates and 1 column

43
APPENDIX B. MATLAB CODE

%of closing prices.

% estimation year − year on which the VaR and ES estimations begin.
%estimation month − month on which the VaR and ES estimations begin.

%% Load the Data and Define the Test Window

% Get daily close and date data

[~,~ , rawData ]= xlsread (filename,’’,’ ’ , ’ basic’ ) ;
date_cell = rawData (2:length(rawData),1);
close_cell = rawData (2:length(rawData),2);
date = zeros (1,length(date_cell));
close = zeros (1,length(close_cell));

% Convert the data to string format

for i =1: length (date)
date(i) = datenum(cell2mat(date_cell(i)));
close_i = cell2mat(close_cell( i ) ) ;
if iscellstr ( close_i )
close (i)= str2num (close_i);
else
close (i)= close_i ;
end
end
date = fliplr(date);
% Convert the date format to a correct one
DateReturns = datestr(x2mdate(date));
DateReturns = flipud(DateReturns);

Returns = zeros(length(DateReturns),1);
for i= 2:length(DateReturns)
Returns(i)= log(close(i)/ close(i−1));
end

%Define samplesize
SampleSize = length(Returns);

%%
% Define the estimation window as 250 trading days for HS and 1000 for

44
 B.1. MATLAB CODE

% parametric methods. For parametric methods the estimation window is used

% to calculate the coefficients of the GARCH family models.

%%
TestWindowStart = find(year(DateReturns)==estimationyear & month(DateReturns)==
estimationmonth,1);
TestWindow = TestWindowStart : SampleSize;
EstimationWindowSizeHS = 250;
EstimationWindowSizeGarch = 1000;

%%
% For a VaR confidence level of 95% and 99%, set the complement of the VaR level.
q = [0.05 0.01];

%% Compute the VaR Using the Historical Simulation Method

%Create the vector for the Historical Simulation VaR

VaR_Historical95 = zeros(length(TestWindow),1);
VaR_Historical99 = zeros(length(TestWindow),1);
ES_Historical95 = zeros(length(TestWindow),1);
ES_Historical99 = zeros(length(TestWindow),1);

%run iterations and save each of the length(TestWindow) Historical Simulation VaR
for t = TestWindow
i = t − TestWindowStart + 1;
EstimationWindow = t−EstimationWindowSizeHS:t;
X = Returns(EstimationWindow);
%VaR estimates
VaR_Historical95(i) = −quantile(X,q(1));
VaR_Historical99(i) = −quantile(X,q(2));
%ES estimates method
N = length(X);
k1 = ceil(N*(1−q(1)));
k2 = ceil(N*(1−q(2)));
z = sort(X);

if k1 < N
ES95 = ((k1 − N*(1−q(1)))*z(k1) + sum(z(k1 +1:N)))/(N*q(1));

45
APPENDIX B. MATLAB CODE

else
ES95 = z(k1);
end

if k2 < N
ES99 = ((k2 − N*(1−q(2)))*z(k2) + sum(z(k2+1:N)))/(N*q(2));
else
ES99 = z(k2);
end
%ES estimate
ES_Historical95(i) = ES95;
ES_Historical99(i) = ES99;

end

%% Compute VaR using the Exponential Weighted Moving Average(EWMA) model.

%Set lambda = 0.94

lambda = 0.94;
Zscore = norminv(q);

%Create EwmaN and GwmaT models

EwmaN = garch (’GARCHLags’,1,’ARCHLags’,1,’Distribution’,’Gaussian’,’Constant’,1e−322 ,’
ARCH’,(1−lambda)−(1e−14) ,’GARCH’,lambda−(1e−14));
EwmaT = garch (’GARCHLags’,1,’ARCHLags’,1,’Distribution’,’t’,’Constant’,1e−322,’ARCH’,(1−
lambda)−(1e−14),’GARCH’,lambda −(1e−14));

% Create vectors for GarchN and GarchT VaR estimates, sigma2, Degrees of
% Freedom.
VaR_EwmaN95 = zeros(length(TestWindow),1);
VaR_EwmaN99 = zeros(length(TestWindow),1);
VaR_EwmaT95 = zeros(length(TestWindow),1);
VaR_EwmaT99 = zeros(length(TestWindow),1);

ES_EwmaN95 = zeros(length(TestWindow),1);
ES_EwmaN99 = zeros(length(TestWindow),1);
ES_EwmaT95 = zeros(length(TestWindow),1);
ES_EwmaT99 = zeros(length(TestWindow),1);

46
 B.1. MATLAB CODE

Sigma2EwmaN = zeros(length(TestWindow),1);
Sigma2EwmaT = zeros(length(TestWindow),1);
DoFs1 = zeros(length(TestWindow),1);

% Estimate coefficients every day and compute GARCHN VaR estimates

for t= TestWindow
%create a variable that starts from 1
i = t − TestWindowStart + 1;

% Vector with relevant returns

EstimationWindow = t−EstimationWindowSizeGarch:t;
R_t = Returns(EstimationWindow);
% Estimate model coefficients
estEwmaN = estimate(EwmaN,R_t,’Display’,’off’);
estEwmaT = estimate(EwmaT,R_t, ’Display’,’off’);

% Forecast volatility
Sigma2EwmaN(i)= forecast(estEwmaN,1,’Y0’,R_t);
SigmaEwmaN = sqrt(Sigma2EwmaN(i));

Sigma2EwmaT(i) = forecast(estEwmaT,1,’Y0’,R_t);
SigmaEwmaT = sqrt(Sigma2EwmaT(i));

%degrees of freedom
DoF = estEwmaT.Distribution.DoF;
DoFs1(i) = DoF;
%VaR estimates
VaR_EwmaN95(i) = −Zscore(1)*SigmaEwmaN;
VaR_EwmaN99(i) = −Zscore(2)*SigmaEwmaN;
VaR_EwmaT95(i) = −Zscore(1)*SigmaEwmaT;
VaR_EwmaT99(i) = −Zscore(2)*SigmaEwmaT;

%ES estimates
ES_EwmaN95(i) = SigmaEwmaN*(normpdf(Zscore(1))/q(1));
ES_EwmaN99(i) = SigmaEwmaN*(normpdf(Zscore(2))/q(2));
ES_EwmaT95(i) = SigmaEwmaT*(tpdf(tinv(1−q(1),DoF),DoF)/(q(1))*((DoF + tinv(1−q(1),
DoF)^2)/(DoF−1))*sqrt((DoF−2)/DoF));

47
APPENDIX B. MATLAB CODE

ES_EwmaT99(i) = SigmaEwmaT*(tpdf(tinv(1−q(2),DoF),DoF)/(q(2))*((DoF + tinv(1−q(2),

DoF)^2)/(DoF−1))*sqrt((DoF−2)/DoF));

end

%% Compute VaR using the GARCH(1,1).

%Create GarchN and GarchT model

GarchN = garch(’GARCHLags’, 1, ’ARCHLags’, 1, ’Distribution’, ’Gaussian’ );
GarchT = garch(’GARCHLags’, 1, ’ARCHLags’, 1, ’Distribution’, ’t’ );

% Create vectors for GarchN and GarchT VaR estimates, sigma2, Degrees of
% Freedom
VaR_GarchN95 = zeros(length(TestWindow),1);
VaR_GarchN99 = zeros(length(TestWindow),1);
VaR_GarchT95 = zeros(length(TestWindow),1);
VaR_GarchT99 = zeros(length(TestWindow),1);

ES_GarchN95 = zeros(length(TestWindow),1);
ES_GarchN99 = zeros(length(TestWindow),1);
ES_GarchT95 = zeros(length(TestWindow),1);
ES_GarchT99 = zeros(length(TestWindow),1);

Sigma2GarchN = zeros(length(TestWindow),1);
Sigma2GarchT = zeros(length(TestWindow),1);
DoFs2 = zeros(length(TestWindow),1);

% Estimate coefficients every day and compute GARCHN VaR estimates

for t= TestWindow
%create a variable that starts from 1
i = t − TestWindowStart +1;
% Vector with relevant returns
EstimationWindow = t−EstimationWindowSizeGarch:t;
R_t = Returns(EstimationWindow);
% Estimate model coefficients
estGarchN = estimate(GarchN,R_t,’Display’,’off’);
estGarchT = estimate(GarchT,R_t, ’Display’,’off’);

48
 B.1. MATLAB CODE

% Forecast volatility and compute standard deviation

Sigma2GarchN(i)= forecast(estGarchN,1,’Y0’,R_t);
SigmaGarchN = sqrt(Sigma2GarchN(i));

Sigma2GarchT(i) = forecast(estGarchT,1,’Y0’,R_t);
SigmaGarchT = sqrt(Sigma2GarchT(i));

%store degrees of freedom

DoF = estGarchT.Distribution.DoF;
DoFs2(i) = DoF;

%VaR estimates
VaR_GarchN95(i) = −SigmaGarchN*Zscore(1);
VaR_GarchN99(i) = −SigmaGarchN*Zscore(2);
VaR_GarchT95(i) = −SigmaGarchT*(tinv(q(1),DoF)*sqrt((DoF−2)/DoF));
VaR_GarchT99(i) = −SigmaGarchT*(tinv(q(2),DoF)*sqrt((DoF−2)/DoF));

%ES estimates
ES_GarchN95(i) = SigmaGarchN*(normpdf(Zscore(1))/q(1));
ES_GarchN99(i) = SigmaGarchN*(normpdf(Zscore(2))/q(2));
ES_GarchT95(i) = SigmaGarchT*(tpdf(tinv(1−q(1),DoF),DoF)/(q(1))*((DoF + tinv(1−q(1),
DoF)^2)/(DoF−1))*sqrt((DoF−2)/DoF));
ES_GarchT99(i) = SigmaGarchT*(tpdf(tinv(1−q(2),DoF),DoF)/(q(2))*((DoF + tinv(1−q(2),
DoF)^2)/(DoF−1))*sqrt((DoF−2)/DoF));

end

%% Compute VaR using the GJR(1,1).

%Create GarchN and GarchT model

GjrN = gjr(’GARCHLags’,1,’ARCHLags’,1,’LeverageLags’,1,’Distribution’, ’Gaussian’ );
GjrT = gjr( ’ GARCHLags’,1,’ARCHLags’,1,’LeverageLags’,1,’Distribution’,’t’ );

% Create vectors for GarchN and GarchT VaR estimates, ES estimates, sigma2, Degrees of
% Freedom
VaR_GjrN95 = zeros(length(TestWindow),1);
VaR_GjrN99 = zeros(length(TestWindow),1);
VaR_GjrT95 = zeros(length(TestWindow),1);

49
APPENDIX B. MATLAB CODE

VaR_GjrT99 = zeros(length(TestWindow),1);

ES_GjrN95 = zeros(length(TestWindow),1);
ES_GjrN99 = zeros(length(TestWindow),1);
ES_GjrT95 = zeros(length(TestWindow),1);
ES_GjrT99 = zeros(length(TestWindow),1);

Sigma2GjrN = zeros(length(TestWindow),1);
Sigma2GjrT = zeros(length(TestWindow),1);
DoFs3 = zeros(length(TestWindow),1);

% Estimate coefficients every day and compute GARCHN VaR estimates

for t= TestWindow
%create a variable that starts from index no.1
i = t − TestWindowStart + 1;
% Vector with relevant returns
EstimationWindow = t−EstimationWindowSizeGarch:t;
R_t = Returns(EstimationWindow);
% Estimate model coefficients
estGjrN = estimate(GjrN,R_t,’Display’,’off’);
estGjrT = estimate(GjrT,R_t, ’Display’,’off ’ ) ;
% Forecast volatility
Sigma2GjrN(i)= forecast(estGjrN,1,’Y0’,R_t);
SigmaGjrN = sqrt(Sigma2GjrN(i));

Sigma2GjrT(i) = forecast(estGjrT,1,’Y0’,R_t);
SigmaGjrT = sqrt(Sigma2GjrT(i));

%degrees of freedom
DoF = estGarchT.Distribution.DoF;
DoFs3(i) = DoF;

%VaR estimates
VaR_GjrN95(i) = −Zscore(1)*SigmaGjrN;
VaR_GjrN99(i) = −Zscore(2)*SigmaGjrN;
VaR_GjrT95(i) = −Zscore(1)*SigmaGjrT;
VaR_GjrT99(i) = −Zscore(2)*SigmaGjrT;

50
 B.1. MATLAB CODE

%ES estimates
ES_GjrN95(i) = SigmaGjrN*(normpdf(Zscore(1))/q(1));
ES_GjrN99(i) = SigmaGjrN*(normpdf(Zscore(2))/q(2));
ES_GjrT95(i) = SigmaGjrT*(tpdf(tinv(1−q(1),DoF),DoF)/(q(1))*((DoF + tinv(1−q(1),DoF)^2)
/(DoF−1))*sqrt((DoF−2)/DoF));
ES_GjrT99(i) = SigmaGjrT*(tpdf(tinv(1−q(2),DoF),DoF)/(q(2))*((DoF + tinv(1−q(2),DoF)^2)
/(DoF−1))*sqrt((DoF−2)/DoF));

end

%% VaR Backtesting
%Set up the returns and dates vector for our the time window to be
%backtested.
ReturnsTest = Returns(TestWindow);
DatesTest = DateReturns(TestWindow);

%Name VaR data, models

VaRData = [VaR_Historical95 VaR_EwmaN95 VaR_EwmaT95 VaR_GarchN95 VaR_GarchT95
VaR_GjrN95 VaR_GjrT95 VaR_Historical99 VaR_EwmaN99 VaR_EwmaT99
VaR_GarchN99 VaR_GarchT99 VaR_GjrN99 ...
VaR_GjrT99];
Models = {’HS95’,’EwmaN95’,’EwmaT95’,’GarchN95’,’GarchT95’, ’GjrN95’, ’GjrT95’ ...
’ HS99’,’EwmaN99’,’EwmaT99’, ’GarchN99’,’GarchT99’, ’GjrN99’, ’GjrT99’};
VaRLevel = [0.95 0.95 0.95 0.95 0.95 0.95 0.95 0.99 0.99 0.99 0.99 0.99 0.99 0.99];

% Run the Unconditional Coverage backtest on our methods.

%Set up the backtesting results

vbt = varbacktest(ReturnsTest,VaRData,’PortfolioID’,PortfolioID,’VaRID’,Models,’VaRLevel’,
VaRLevel);
%Report the Unconditional coverage backtesting result
pof(vbt)

%% ES Backtesting

% Run the Unconditional backtest on our methods.

rng(’default’ ) ; % for reproducibility; the esbacktestbysim constructor runs a simulation

51
APPENDIX B. MATLAB CODE

%Name VaR data, ES data, and degrees of freedom.

ESData = [ES_Historical95 ES_EwmaN95 ES_EwmaT95 ES_GarchN95 ES_GarchT95

ES_GjrN95 ES_GjrT95 ES_Historical99 ES_EwmaN99 ES_EwmaT99 ES_GarchN99
ES_GarchT99 ES_GjrN99 ...
ES_GjrT99];

DoFs = [round(mean(DoFs1)), round(mean(DoFs2)), round(mean(DoFs3)),round(mean(

DoFs1)), round(mean(DoFs2)), round(mean(DoFs3))];

%set up the ES backtesting results

VaRDataT = [VaR_EwmaT95 VaR_GarchT95 VaR_GjrT95 VaR_EwmaT99 VaR_GarchT99
VaR_GjrT99];
ESDataT = [ES_EwmaT95 ES_GarchT95 ES_GjrT95 ES_EwmaT99 ES_GarchT99 ES_GjrT99
];
ModelsT = {’EwmaT95’, ’GarchT95’, ’GjrT95’, ’EwmaT99’, ’GarchT99’, ’GjrT99’};
VaRLevelT = [0.95 0.95 0.95 0.99 0.99 0.99];

VaRDataN = [VaR_EwmaN95 VaR_GarchN95 VaR_GjrN95 VaR_EwmaN99 VaR_GarchN99

VaR_GjrN99];
ESDataN = [ES_EwmaN95 ES_GarchN95 ES_GjrN95 ES_EwmaN99 ES_GarchN99
ES_GjrN99];
ModelsN = {’EwmaN95’, ’GarchN95’, ’GjrN95’, ’EwmaN99’, ’GarchN99’, ’GjrN99’};
VaRLevelN = [0.95 0.95 0.95 0.99 0.99 0.99];

VaRDataHS = [VaR_Historical95 VaR_Historical99];

ESDataHS = [ES_Historical95 ES_Historical99];
ModelsHS = {’HS95’,’HS99’};
VaRLevelHS = [0.95 0.99];

ebtsT = esbacktestbysim(ReturnsTest,VaRDataT,ESDataT,"t",...
’ DegreesOfFreedom’,15,...
’ Location’ ,0,...
’ Scale’ ,1,...
’ PortfolioID’ ,PortfolioID ,...
’ VaRID’,ModelsT,...

52
 B.1. MATLAB CODE

’ VaRLevel’,VaRLevelT);

ebtsN = esbacktestbysim(ReturnsTest,VaRDataN,ESDataN,"normal",...
’ Mean’,0,...
’ StandardDeviation’,1,...
’ PortfolioID’ ,PortfolioID ,...
’ VaRID’,ModelsN,...
’ VaRLevel’,VaRLevelN);

ebtsHS = esbacktest(ReturnsTest,VaRDataHS,ESDataHS,’PortfolioID’,PortfolioID,’VaRID’,
ModelsHS,’VaRLevel’,VaRLevelHS);
%Report the Unconditional test

unconditional(ebtsT)
unconditional(ebtsN)
unconditionalNormal(ebtsHS)

%% Compute Sample Statistics for the Returns

Mean = mean(Returns);
Standard_Deviation = std(Returns);
Kurtosis = kurtosis(Returns);
Skewness = skewness(Returns);
Min = min(Returns);
Max = max(Returns);

SampleStatistics = table(Mean,Standard_Deviation ,Kurtosis,Skewness,Min,Max)

end

53
 B IBLIOGRAPHY

[1] Risk Management and Financial Institutions, John C.Hull, 2012.

[2] P. E. A LEXANDER J. M C N EIL , R UDIGER F REY, Quantitative Risk Management, Princeton

University Press, 2005.

[3] B. S. C ARLO A CERBI, Backtesting Expected Shortfall, MSCI Inc., 2014.

[4] P. F. C HRISTOFFERSEN, Elements of Financial Risk Management, 2012.

[5] J. D ANIELSSON AND C. G. DE V RIES, Value-at-risk and extreme returns, first version from
1997, working paper, Financial Markets Group, London School of Economics, London.,
(2000).

[6] J. D ANIELSSON AND C. G. D E V RIES, Value-at-risk and extreme returns, Annales d’Economie
et de Statistique, (2000), pp. 239–270.

[7] K. D OWD, Beyond value at risk: the new science of risk management, John Wiley Sons, 1998.

[8] B. FOR I NTERNALTIONAL S ETTLEMENTS, Minimum capital requirements for market risk.

[9] , Revised framework for market risk capital requirements issued by the basel committee.

[10] J. H ULL, Options, futures and other derivatives, Prentice Hall, 2000.

[11] P. J ORION, Value at Risk: The New Benchmark for Managing Financial Risk, McGraw-Hill,
2007.

[12] D. K IDD, Value at risk and condition value at risk: A comparison, Investment Risk and
Performance Articles 2012.1, (2012).

[13] T. J. L INSMEIER AND N. D. P EARSON, Value at risk, Financial Analysts Journal vol. 56, no.
2, (2000).

[14] S. M ANGANELLI AND R. E NGLE, Value at risk models in finance, European Central Bank -
Working Papers, (2001).

[15] C. M ARSHALL AND M. S IEGEL, Value-at-risk: Implementing a risk measurement standard,

(1996).

55
BIBLIOGRAPHY

[16] M ATHWORKS®, Overview of expected shortfall.

[17] M. N. S. P. T. P. M C K INSEY C OMPANY, A MIT M EHTA, Managing market risk: Today and
tomorrow, McKinsey Working Papers on Risk, Number 32, (2012).

[18] S. M ITRA AND T. J I, Risk measures in quantitative finance, International Journal of Business
Continuity and Risk Management, 1 (2010), pp. 120–135.

[19] M. M USIELA AND M. R UTKOWSKI ., Martingale Methods in Financial Modelling, Springer,

2005.

[20] X. X. R UI L UO, W EINAN Z HANG AND J. WANG, A neural stochastic volatility model,
arXiv:1712.00504v1 30 Nov 2017, (2017).

[21] H. X. T ZE L EUNG L AI, Statistical Method for Financial Markets, 2007.

56