Lund Institute of Technology

Master Thesis in Risk Management and Safety Engineering

VaR methods for linear instruments

Birgir Viðarsson
Title: VaR methods for linear instruments
Author: Birgir Viðarsson
Instructor: Dr. Freyr Harðarson. Risk Management, Landsbanki
Examiner: Dr. Henrik Johansson. Department of Fire Safety Engineering
and Systems Safety, Lund University

Report nr.: 5268

ISSN: 1402-3504
ISRN: LUTVDG/TVBB5268SE
Keywords
VaR, Volatility, Covariance, GARCH, EWMA, Univariate, Multivariate,
Market risk, Financial time series
Abstract
In this thesis various Value-at-Risk models are compared and evaluated to-
wards nding the optimal model for the bank's trading book. The focus is
on linear instruments (stocks, indicies and currency) and real market data,
both domestic and foreign, is used for the calculations. I nd that the
GARCH(1,1) based model outperforms other models in volatility estimation
and should thus be a wise choice when volatility estimation is needed, but
GARCH based methods become rather complex for a multivariate covari-
ance estimation. Therefore a mixture of simpler models such as EWMA is
needed for making rational estimates.

c Copyright: Brandteknik och Riskhantering, Lunds tekniska högskola,

Lunds universitet, Lund 2008.

Brandteknik och Riskhantering Department of Fire Safety

Lunds tekniska högskola Engineering and Systems Safety
Lunds universitet Lund University
Box 118 P.O. Box 118
221 00 Lund SE-221 00 Lund
Sweden

[email protected] [email protected]
http://www.brand.lth.se http://www.brand.lth.se/english
Telefon: 046 - 222 73 60 Telephone: +46 46 222 73 60
Telefax: 046 - 222 46 12 Fax: +46 46 222 46 12
Thanks:
Dr. Freyr Harðarson, Dr. Henrik Johansson and the employees at
Risk Management department in Landsbanki, Reykjavík
Summary
In the last decades the growth in nancial markets has been enormous.
Global trading is easier then ever and business with all kinds of nancial
instrument such as stocks and bonds and their derivatives is common. Usu-
ally the aim with all trades is to save or gain money one way or another,
but with all trades there exist as well the risk that the trade will not be
protable and result in a nancial loss. This risk that leads to nancial loss
can be regarded as nancial risk. One of the main category of nancial risk
is market risk which is the risk that market uctuations will lead to nan-
cial loss. For nancial institutions who have large amount of their assets in
nancial instrument (traded on markets), market risk can have great impact
on their performance and therefore essential to quantify.

In this thesis the goal is to quantify market risk and for that cause use the
term Value-at-Risk (VaR), which is commonly used among nancial institu-
tions. The term VaR is dened as the amount X that you are α% certain
of not losing more than the following N days. More general, VaR gives a
kind of worst case scenario at preferred level of probability (α) and time pe-
riod (N -days). There are various ways and techniques for calculating VaR,
all with their pros and cons, and generally depended on presumptions. In
this thesis the focus is on linear instruments, such as equity and currency,
and for that sake the main categorize of calculating VaR are parametric and
non-parametric approaches. VaR is also used in regulatory terms. The Basel
Committee, which is a international banking supervisor, uses VaR to stip-
ulate the minimum amount of regulatory capital that nancial institution
must have available at all times. This is done to prevent nancial crisis and
possible bankruptcy due to unforeseen market movements, credit defaults or
any other risk faced by the nancial institution. Both parametric and non-
parametric methods are used to calculate VaR for real market data, with the
goal of nding the method that suits the bank's trading book the best while
fullling regulations set by the Basel Committee.

Criteria's are set up for judging the best model where the main conclusions
are that a mixture of both parametric and non-parametric methods should be
used. GARCH volatility estimate is found to be the best to describe market
volatility conditions, but they become very complex as the dimensionality
increases. Therefore I recommend that a mixture of both parametric and
non-parametric methods with GARCH and/or EWMA volatility estimates
should be used. I also nd that a student's t-distribution ts the data better
then the commonly used normal distribution. But student's t-distribution
with GARCH volatility estimates are sensitive and therefore my recommen-
dation is that normal distribution should be used as well as some alternative
distribution's.
Contents

1 Introduction 1
1.1 Aim and purpose . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 What is VaR? . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.4 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

2 Financial time series modeling 6

2.1 Return analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.2 Portfolio . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.3 Volatility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.4 Estimating volatility . . . . . . . . . . . . . . . . . . . . . . . 8
2.4.1 Historical volatility . . . . . . . . . . . . . . . . . . . . 8
2.4.2 EWMA model . . . . . . . . . . . . . . . . . . . . . . 10
2.4.3 GARCH models . . . . . . . . . . . . . . . . . . . . . 11
2.5 Covariance and correlation . . . . . . . . . . . . . . . . . . . . 14
2.6 Estimating covariance and correlation . . . . . . . . . . . . . 15
2.6.1 Historical covariance and correlation . . . . . . . . . . 15
2.6.2 EWMA covariance . . . . . . . . . . . . . . . . . . . . 15
2.6.3 GARCH covariance . . . . . . . . . . . . . . . . . . . . 15

3 Parametric methods 19
3.1 Univariate parametric methods . . . . . . . . . . . . . . . . . 19
3.2 Multivariate parametric methods . . . . . . . . . . . . . . . . 20

4 Non-parametric methods 23
4.1 Basic historical simulation . . . . . . . . . . . . . . . . . . . . 23
4.2 Weighted historical simulation . . . . . . . . . . . . . . . . . . 25
4.2.1 Age weighted historical simulation (Age-WHS) . . . . 25
4.2.2 Volatility weighted historical simulation (VWHS) . . . 25
4.2.3 Filtered historical simulation (FHS) . . . . . . . . . . 26

iv
 v

5 Analysis 27
5.1 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
5.1.1 Backtesting . . . . . . . . . . . . . . . . . . . . . . . . 27
5.1.2 Basel zones . . . . . . . . . . . . . . . . . . . . . . . . 28
5.1.3 The basic frequency test . . . . . . . . . . . . . . . . . 29
5.2 Univariate case . . . . . . . . . . . . . . . . . . . . . . . . . . 31
5.2.1 Parametric methods . . . . . . . . . . . . . . . . . . . 31
5.2.2 Non-parametric methods . . . . . . . . . . . . . . . . . 40
5.3 Multivariate case . . . . . . . . . . . . . . . . . . . . . . . . . 44
5.3.1 Parametric methods . . . . . . . . . . . . . . . . . . . 44
5.3.2 Non-parametric methods . . . . . . . . . . . . . . . . . 48

6 Conclusion 50
6.1 Criticism on VaR . . . . . . . . . . . . . . . . . . . . . . . . . 53
6.2 Further analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 54

A Pictures 55
A.1 Univariate case . . . . . . . . . . . . . . . . . . . . . . . . . . 55
A.1.1 Parametric methods . . . . . . . . . . . . . . . . . . . 55
A.1.2 Non-parametric methods . . . . . . . . . . . . . . . . . 59
A.2 Multivariate case . . . . . . . . . . . . . . . . . . . . . . . . . 60
A.2.1 Parametric methods . . . . . . . . . . . . . . . . . . . 60
A.2.2 Non-parametric methods . . . . . . . . . . . . . . . . . 61
Chapter 1

Introduction

As pointed out by Kaplan and Garrick (1981) quantitative risk can be

dened as combination of scenario, probability and a consequence. It is
necessary to evaluate all these parts to get a relative and rational estimate
of risk. More general in risk assessment and risk management we want to
answer the question; What can go wrong, how likely is it and what will
be the consequences. The main elds of risk assessment can be separated
into 5 categories; Safety risk, Health risk, Ecological and Environmental risk,
Public welfare risk and Financial risk (Kolluru, 1995).

Financial risk can be thought of as any risk concerned with nancial loss
due to some random changes in underlying risk factors (stock, currency,
derivatives, interest rate etc.) and can furthermore be categorized into three
main level of concerns; Market risk (due to movements in market factors),
Credit risk (the risk that a person or an organization will not fulll his/her
obligations) and Operational risk (risk of loss because of systematic failures)
(Dowd, 1998). Methods for measuring and evaluating nancial risk are many
and depend on what is of interest to examine. Just for market risk there are
numerous of ways. For options and derivatives examining `the greeks' might
by a good choice, stress testing might by good for worst case scenario anal-
ysis while Value-at-Risk (VaR) could give a universal risk measure for the
exposed market risk. In this thesis I will concentrate on VaR for measuring
market risk1 .

In the last 50 years there has been enormous growth in trading worldwide,
for example the New York Stock Exchange has grown from $4 million in
1961 to $1.6 trillion2 in 2005. This growth and the massive increase of
new instruments (all kinds of derivatives, swaps, CDO's (collateralized debt
obligation) and so on) has invoked more need for good risk management in
1
This was proposed by Landsbanki bank
2
One trillion equals million millions (1012 )
2 INTRODUCTION

the nancial sector. Managing risk improves the value of the company and
can avoid major nancial disaster such as Orange County (1994), Barings
Bank (1995), Enron (2001), WorldCom (2002) and Sociètè Générale (2008)
which have all been related to poor risk management.

1.1 Aim and purpose

The aim of the thesis is to evaluate dierent methods for calculating VaR
for bank's trading book, where the goal is to minimize VaR while fullling
regulatory requirements. The focus will be on linear instruments (such as
equity and currency) and actual market data will be used for the evaluation,
both domestic and foreign (Icelandic and Swedish).

VaR estimation can be categorized into three main categories;

parametric methods, non-parametric methods and Monte Carlo simulation
methods. Both parametric and non-parametric approaches will be used to
calculate VaR for the market data and compared with critical judgment to-
wards obtaining the optimal VaR method for the bank's trading book. Monte
Carlo methods are mostly used on non-linear nancial instruments and will
therefore not be the topic of this thesis.

1.2 What is VaR?

You are responsible for managing your company's foreign ex-

change positions. Your boss, or your boss's boss, has been read-
ing about derivatives losses suered by other companies, and
wants to know just how much market risk the company is tak-
ing. What do you say?
You could start by listing and describing the company's positions,
but this isn't likely to be helpful unless there are only a handful.
Even then, it helps only if your superiors understand all of the
positions and instruments, and the risks inherent in each. Or you
could talk about portfolio's sensitivities, i.e., how much the value
of the portfolio changes when various underlying market rate or
prices change, or perhaps options delta and gammas. Even if
you are condent of your ability to explain these in English, you
still have no natural way to net the risk of your short position
in Deutsche marks against the long positions in Dutch guilders.
(It makes sense to do this because gains or losses on the short
position in marks will be oset almost perfectly by gains or losses
on the long position in guilders.) You could simply assure your
superiors that you never speculate but rather use derivatives only
1.2 WHAT IS VAR? 3

to hedge, but they understand that this statement is vacuous.

They know that the word `hedge' is so ill-dened and exible
that virtually any transaction can be characterized as a hedge.
So what do you say? (Linsmeier and Pearson, 1996)

Value at Risk (VaR) is an attempt to give a relatively simple measure of

nancial risk (not only market risk) with a single number answering the
question `how bad can things get?' (Dowd, 2005). VaR could thus be a fair
attempt to answer the question in the example above. The statement we
wan't to make with VaR is:

We are α percent certain to lose not more than X much money

in the following N days.

The amount X is a loss due to market movements and could be for a sin-
gle asset or a portfolio (see section 2.2). The amount X is function of two
variables, the condence level α and time period, usually given in N days.
The calculation of VaR is thus based on the probability of changes in asset
or the portfolio value over the next N days. To be more mathematical VaR
is the quantile corresponding to the (1 - α) of the return distribution, so if
we set p = 1 − α and call qp the p-th quantile of α, then VaR can be dened
as; VaRα% = −qp for specied condence level and period. Meaning that we
can be α certain to not lose more than VaR in that period (see gure 1.1)

Figure 1.1: VaR at α% condence level for some imaginary prot/loss series, as-
suming normality

Let's imagine that the x-scale represents the prot/loss distribution of some
imaginary company (and the scale is perhaps in million euros). Say that
4 INTRODUCTION

a company is interested in knowing their market risk status and especially

what the 1 − α worst case would be. Then VaRα% would give a estimate of
the loss that would not be exceeded α percent of the time. Likewise, losses
larger then VaRα% happen only (1 - α)% of the time. This VaRα% is equal
to the boundaries between the blue and white area on gure 1.1.

The reason for VaR's popularity, as a denition of risk, is its simplicity in

interpretation. It is relatively easy to understand, has the unit of the mea-
sure (i.e. euro, SEK, ISK)3 , it is probabilistic (concerned with probability), it
can be used for any type of positions (bond, stock, currency, derivatives etc.)
and portfolios (meaning that it will aggregate many sub-positions into one
measure) and it is holistic (meaning that takes into account all underlying
risk factors) (Dowd, 2005). VaR is also used in regulatory terms. Financial
institutions are required to have some minimum regulatory capital available
at all times for safety reasons and with the Basel accord4 , nancial insti-
tution were allowed to base this minimum capital partly on their own VaR
estimates. This is known as the internal approach, see section 5.1.1.

1.3 Data

In the thesis I will examine four stocks, two indices, two portfolios and one
currency pair. Portfolio 1 will be made of equal shares (25%) in all four
stocks, 25% LAIS, 25% MARL, 25% NDA and 25% ERIC and Portfolio 2
will consist stocks in the following ratios; 50% LAIS, 30% MARL, 10% NDA
and 10% ERIC, see 1.1.

Name Ticker Description

Landsbankinn LAIS Icelandic bank
Marel MARL Icelandic food processing company
OMX Iceland 15 ISXI15 Index consisting of 15 Icelandic companies
Nordea NDA Swedish bank
Ericsson ERIC Swedish tele & datacommunication company
OMX OMX Index consisting of 30 Swedish companies
Portfolio 1 - Equal share portfolio
Portfolio 2 - 50% LAIS, 30% MARL, 10% NDA and 10% ERIC
USD/ISK USDISK USA dollars to Icelandic króna
Table 1.1: The time series used for calculation

In the thesis I will though use percentage for the sake of comparison, see section 5.1
3

Banking supervision accords (recommendations on banking laws and regulations),

4

Basel I and Basel II issued by the Basel Committee on Banking Supervision (BCBS), see
Basel Committee (2006).
1.4 OVERVIEW 5

Various time periods will be used for modeling to try to capture the most
ecient model. Time periods are given in table 1.2.
Name From date To date
Short 1. July 2006 30. June 2008
Long 28. June 2004 30. June 2008
Extra Long 1. July 1998 30. June 2008
Table 1.2: The time series used for calculation

where the following time periods are assigned to the time series, table 1.3.

Name Ticker Description

Landsbankinn LAIS Short and Long
Marel MARL Short and Long
OMX Iceland 15 ICEXI15 Short and Long
Nordea NDA Short and Long
Ericsson ERIC Short and Long
OMX OMX Short and Long
Portfolio 1 - Long
Portfolio 2 - Long
USD/ISK USDISK Extra Long
Table 1.3: The time periods used for each time series

All data is provided through the Bloomberg Terminal5 . It should be noted

that all data used are `hypothetical' outcomes, meaning that it is assumed
that no dividends are paid out, no trades are done and weights in portfolios
are assumed to be xed at all times. The `window size' (sample size) used to
estimate each days VaR estimate is 500 day's (meaning that each days VaR
estimate is calculated from the previous 500 days) which are the requirement
set by the Basel Committee (2006).

1.4 Overview

In chapter 2 I will give an overview of the main topics of nancial time

series modeling, introduce how data is treated, present various methods
for estimating volatility and covariances. Chapter 3 describes parametric
approaches for calculating VaR and chapter 4 decribes non-parametric ap-
proaches for calculating VaR. Analysis, criteria's and results are presented
in chapter 5 and nally chapter 6 gives conclusions, discussions and ideas for
further analysis.
5
A computer system by Bloomberg that provides information on nancial market data.
Chapter 2

Financial time series modeling

Financial time series modeling is the task of building a model used to pre-
dict, evaluate and/or forecast the performance of nancial instruments. The
models often use a mixture of theories from economics, engineering, statistics
and business administration to obtain information. The main characteris-
tics with nancial data is uncertainty and sudden movements and therefore
the central aim of the modeling is to be able to explain and predict these
characteristics. Two known facts with nancial time series is that low and
high uctuations tend to come in periods (resulting in periods of high and
low returns) and the fact that their probability distributions usually have
fatter tails than normal time series, meaning that extreme cases (high and
low returns) are more likely than normal distribution describes.

2.1 Return analysis

For modeling nancial data it is common to work with prot/loss data (P/L)
and return series rt . Prot/Loss data tells how much you have gained (prot)
or lost (loss) over a time period between time t − 1 and t and can be written
as:

P/Lt = Pt − Pt−1
where Pt represents the value of the asset at time t and Pt−1 the value at
time t − 1. Return series tells you how much in percentage you gained or
lost between time t − 1 and t. There are two ways of representing the return
series, arithmetic (2.1) or geometric (2.2)

Pt − Pt−1
rt = (2.1)
Pt−1
Pt
rt = ln( ) (2.2)
Pt−1
2.2 PORTFOLIO 7

The former is more common when working with non-parametric methods

and the latter with parametric methods. It is also common to work with the
deviation from the mean of the series at each time t, called error terms or
residuals t and represented as:

t = rt − r̄
where r̄ is the average return of the series (often when dealing with daily
returns the mean will be low and therefore the approximation r̄ → 0 is often
made, which leads to t ≈ rt ).

2.2 Portfolio

In nance a portfolio is a mix or a collection of assets, k ≥ 2. The idea with

a portfolio is often to build up a more stable ownership and spreading the
risk (`not putting all the eggs in the same basket'). This is called to diversify.
Each asset in the portfolio is given weight depending on it's market-value
M V , i.e.

M Vi
wi = Pk (2.3)
i=1 M Vi
The total return of the portfolio rp can then be calculated as:
 T  
w1 r1

wTr =  w2 r2
   
rp = .. .. (2.4)
  
. .
  
  
wk rk
Variance of the portfolio can be obtained as:

T  
σ12 σ1 σ2 · · · σ1 σk
 
w1 w1
..
w w w2   σ2 σ1 σ22 . w2
 
TΣ
   
σp2 (2.5)

= = .. ..
 
  . .

.  .. .. .
 
  
wk σk σ1 . . . σk2 wk

where Σ is variance-covariance matrix.

8 FINANCIAL TIME SERIES MODELING

2.3 Volatility

The volatility σ of a variable is dened as the standard deviation of the re-

turns rt per unit of time t, when the returns are expressed using continuous
compounding1 . Usually the unit of time is one day so that the volatility is
expressed as a standard deviations of the continuous compounded return per
day (Hull, 2007).

By examining a long return series such as in gure 2.1 it is clear that variance,
and therefore volatility, varies with time. Volatility represents risk and since
volatility is changing that implies that market risk is changing. ( ) ?

Figure 2.1: Return series for OMX

This is one of the main characteristics with nancial volatility, called

heteroscedasticity, which is the behavior of having time varying periods of
low and high volatilities. We therefore want to take this characteristics into
account when estimating volatility.

2.4 Estimating volatility

2.4.1 Historical volatility

The most obvious choice of estimating volatility is the historical (equal

weighted) volatility dened as (here I talk about volatility although equa-
tion gives the variance σ 2 , the volatility
√ is obtained of course by taking the
square-root of the variance, σ = σ 2 , this also applies for other equation in
this chapter):

N
1 X
σt2 = (rt−i+1 − r̄)2 (2.6)
N −1
i=1

1
Continuous compounding means that the growth and loss of the variable (asset) is
expressed continuously, not just the prot or loss from the time when bought till the time
when sold.
2.4 ESTIMATING VOLATILITY 9

where N is the number of days used in the estimate. As can be seen in gure
2.2 the size of N has an eect on how the volatility estimate will look like.
For a small N the estimate is usually more responsive and jumps a lot, but
becomes more stable as N gets larger.

Figure 2.2: Volatility Estimates for MARL, shows how dierent time periods eect
historical volatility estimates

By making the approximation that average of daily return is close to zero

and it makes insignicant dierence to the estimate, and that it makes in-
signicant dierence to divide with N instead of N − 1 when dealing with
long time series, equation 2.6 can be modied as:
N
1 X 2
σt2 = rt−i+1 (2.7)
N
i=1

The problem with this method is that it treats all observations equally, mean-
ing that all observations in the estimate will have the same weight. That
is old and new (in time) observations are treated equally and have as much
contribution to the estimate. This means that if a shock on the market had
occurred in the past it will have as much impact on the estimate as any
other day until it falls out of the sample space. This could cause a overesti-
mate while the shock is still in the sample space and a sudden jump when it
10 FINANCIAL TIME SERIES MODELING

falls out of the sample. This overestimate is known as `ghost eects'. This
problem and the fact that volatility tends to vary with time in nancial time
series has led to development of weighted volatility estimates in the form:

N
X
σt2 = 2
αi rt−i+1 (2.8)
i=1

where αi are the weights which are assigned to each return rt . The weights,
0 < αi < 1, decline as i gets larger and sum up to 1.

2.4.2 EWMA model

One of the most known weighting models is the EWMA model (exponentially
weighting moving average) where the weights decrease exponentially as we
move back in time;

N
X
σt2 ≈ (1 − λ) λi−1 rt−i+1
2
(2.9)
i=1

Here λ is dened as λ = αi+1 /αi and thus a constant between 0 and 1.

The choice of λ is depending of the behavior desired. Low λ values gives
more responsive and rougher behavior in the volatility estimate and high λ
values gives less responsive and smoother volatility estimates, see gure 2.3.
RiskMetrics2 suggests that λ = 0, 94 should be used for equity and λ = 0.97
for foreign exchange (FX), such as currency trade. The parameter could also
be optimized using traditional maximum log-likelihood estimate. With little
modication equation (2.9) can be written as:

σt2 ≈ λσt−1
2
+ (1 − λ)rt2 (2.10)

which is a simple updating formula, where the only variable needed for esti-
mating the volatility at time t is the most recent return rt (the return after
the market closes) and volatility σt−1 (the estimated volatility from the day
before).

The advantages with the EWMA model is that it only relies on one param-
eter, λ, tends to produce much less ghost eects than the historical equal
weighted model and a very little data needs to be stored. The main disad-
vantages with the EWMA model is that it takes λ to be constant and can
therefore be unresponsive to market conditions (Dowd, 2005).
2
J.P. Morgan Guaranty Trust Company, see Morgan (1996)
2.4 ESTIMATING VOLATILITY 11

Figure 2.3: Volatility Estimates for MARL, shows how dierent volatility estimates
react to a shock on market

2.4.3 GARCH models

GARCH (generalized autoregressive conditional heteroscedasticity) models
proposed by Bollerslev (1986), which was an extension of Engle's (1982)
ARCH models, give a solution to this kind of problem. GARCH models can
show volatility clustering and leptokurtosis (fatter tails than normal tails)
which are two of the most important facts with nancial time series. The
GARCH(p, q) model depends on q past volatilities and p last error terms and
has the following representation:
(
t = σt zt
(2.11)
σt2 = ω + qi=1 αi 2t−i + pi=1 βi σt−i
2
P P

where the residuals t are dened as before as t = rt − r̄. The parameters

must be non-negative and fulll;
q
X p
X
αi + βi < 1
i i

The GARCH(p, q) model diers with the choice o distribution governing

12 FINANCIAL TIME SERIES MODELING

the residuals. The most common choice is normal distribution, but could
also be for example t-distributed (which produces even fatter tails). A max-
imum likelihood is ideal for obtaining parameter values (α, β and ω ).

To obtain the optimal number of parameters k and p tests such as deviance

statistics test can be used:

D = 2{`2 (M2 ) − `1 (M1 )} > cα (2.12)

where `i (Mi ) is the maximum log-likelihood parameter for model i. The
maximum log-likelihood of the model with fewer parameters M1 should be
subtracted from the higher number of parameters M2 and compared with cα
which is the (1 − α) quantile of the χ2 distribution. Model M1 is rejected if
D > cα in favor of M2 .

GARCH(1,1)
The GARCH(1, 1) model is the most popular GARCH model. The reason
is because of its simplicity (depends only on the `last' volatility and return),
seems to t most nancial data fairly well (higher values of k and p usually
don't give signicantly better result) and from the principle of parsimony
(to choose as simple model as possible to t the data). The model is given
as:

σt2 = ω + α2t−1 + βσt−1

2
(2.13)
where the parameters must as before be non-negative and fulll α + β < 1.
If the governing distribution is assumed to be normal (most common) then
the maximum log-likelihood function is obtained as:

n X 2 n
n X
` = − ln(2π) − ln(σi2 ) − i
(2.14)
2
i=1
2σi2 i=1
n n
X X 2i
' − ln(σi2 ) −
i=1 i=1
2σi2

since the constant doesn't matter in the maximization. The GARCH(1,1)

model depends on the same variables as the EWMA model (volatility and
return) but now there are three parameters instead of one. Figure 2.3 shows
a comparison between GARCH and EWMA model. From the gure it is
clear that the GARCH model adopts the market condition better than the
EWMA model. The EWMA model can be regarded as a special case of
GARCH(1,1) with ω = 0, α = 1 − λ, and β = λ. The main advantages
with GARCH models is how they can accommodate heteroscedasticity and
fat tails and how responsive they are to market condition (see gure 2.3).
2.4 ESTIMATING VOLATILITY 13

The main disadvantages is that the GARCH models are more complex then
the other volatility estimates.

Many modication have been developed to the GARCH(1, 1) model such

as A-GARCH, E-GARCH, GJR-GARCH, I-GARCH, V-GARCH and many
more3 . Some of those models can explain `leverage eect' which is the behav-
ior of not responding the same way to negative and positive returns (good
and bad news on market), i.e. are asymmetric not symmetric as the GARCH
model. Here I will introduce two of them.

GJR-GARCH: Glosten, Jagannathan & Runkle

The GJR-GARCH model (also known as Threshold-GARCH or T-GARCH)
was proposed by Glosten, Jagannathan and Runkle (1993) is similar to the
GARCH(1,1) model but also exhibits the term St−1 to capture the leverage
eect.

σt2 = ω + βσt−1
2
+ (α + γSt−1 )2t−1 (2.15)
where St−1 = 0 if t−1 ≥ 0 and St−1 = 1 if t−1 < 0, so it doesn't react
the same to positive and negative returns. The GJR-GARCH model has
the same parameter restriction as GARCH(1,1), that is α and β must be
non-negative (γ can be negative) and fulll:

α+β <1

E-GARCH: Exponential GARCH

The E-GARCH model by Nelson (1991) is a little bit dierent from the
GARCH(1,1) and GJR-GARCH, but also captures the leverage eect by
letting the volatility estimate depend on the sign of the lagged residual.

 
|t−1 | t−1
ln(σt2 ) = ω + β ln(σt−1
2
)+α − E{zt−1 } + γ (2.16)
σt−1 σt−1

One of the main advantages with the E-GARCH model is that it has no
parameter restrictions as the GARCH model. Figure 2.4 shows a compari-
son between the GARCH models introduced, GARCH(1,1), E-GARCH and
GJR-GARCH. By looking at gure 2.4 the leverage eect can be examined,
GARCH(1,1) rises more than the other two when large positive shocks occur
and rises less when a short period of negative returns occurs.

3
see for example Hansen and Lunde (2005) who make a comparison of volatility fore-
casting with many GARCH based methods
14 FINANCIAL TIME SERIES MODELING

Figure 2.4: Volatility Estimates for MARL, shows how dierently the GARCH
models behave.

2.5 Covariance and correlation

In statistics the covariance gives an estimate of how much two variables

change together and is only used when two or more variables are of concern
(bivariate or multivariate). For two variables x and y the covariance can be
dened as:

cov(x, y) = E[xy] − E[x]E[y] (2.17)

A strictly related term is the correlation between two variables which gives
a measure of how the two variables move together. The correlation is always
between −1 and +1 and can be dened as

cov(x, y)
corr(x, y) = (2.18)
σx σy

where −1 means that instruments x and y move totally against each other, 0
means that nothing can be said about their movement together and 1 means
that they move totally the same.
2.6 ESTIMATING COVARIANCE AND CORRELATION 15

2.6 Estimating covariance and correlation

2.6.1 Historical covariance and correlation

Estimation of covariance and correlation is parallel to estimation of volatility.
The historical correlation estimate can be written as:
Pn
xt−i yt−i
corr(x, y)t = qP i=1 P (2.19)
n 2 n 2
i=1 xt−i i=1 yt−i

and then the covariance could be obtained with:

cov(x, y) = σx σy corr(x, y) (2.20)

2.6.2 EWMA covariance

EWMA covariance can be estimated as:

cov(x, y)t = λcov(x, y)t−1 + (1 − λ)xt−1 yt−1 (2.21)

and the correlation can be obtained with equation 2.18. As before RiskMet-
rics suggests that λ = 0, 94 should be used for equity and λ = 0.97 for FX.
To ensure that the matrix is positive denite or semi-positive denite4 it is
important to use the same λ value for all parameters in the matrix.

2.6.3 GARCH covariance

Since the GARCH based volatility model seem to be powerful tool for the
estimation of volatility, it seem to be obvious idea to generate a multivariate
version of GARCH estimation. The problem is that multivariate GARCH
models are computationally complex and the number of parameters to be
estimated grow rapidly as the number of assets in the portfolio grow. Mul-
tivariate GARCH models, often called VECH can be written as:

Ht = W + A(t−1Tt−1) + B(Ht−1) (2.22)

where Ht is the vector of volatilities, W is a vector of ω coecients and
A and B are matrix for α and β coecients, respectively. For a portfolio
consisting of only two assets the multivariate VECH becomes:

4
Positive denite or semi-positive denite means in this case that the portfolio volatility
σp has to be larger or equal to zero, i.e. σp2 = wT Σw ≥ 0 for allw.
16 FINANCIAL TIME SERIES MODELING

2 21,t−1
      
σ1,t ω1 α11 α12 α13
 σ1,t σ2,t  =  ω2  +  α21 α22 α23   1,t−1 2,t−1  (2.23)
2
σ2,t ω3 α31 α32 α33 22,t−1
2
  
β11 β12 β13 σ1,t−1
+  β21 β22 β23   σ1,t−1 σ2,t−1 
β31 β32 β33 2
σ2,t−1

Here the total number of parameters to be estimated are 21 and with a 3

asset portfolio the number of parameters needed to be estimated becomes
H
78. Furthermore this formulation doesn't ensure t to be positive denite
?
( ).

Because of this complexity there has been developed several simplied GARCH
based covariance models such as diagonal VECH (DVECH) proposed by
Bollerslev, Engle and Woolridge (1988) where the A and B are assumed to
be diagonal and the total number of parameters becomes 3(k(k+1)/2) where
k is the number of asset in the portfolio, the Constant Conditional Corre-
lation model (CCC) proposed by Bollerslev (1990) where total number of
parameters becomes k(k + 5)/2 and Dynamic Conditional Correlation model
(DCC) proposed by Engle (2002) where total number of parameters becomes
(k + 1)(k + 4)/2. Table 2.1 gives an comparison of parameters needed to be
estimated in multivariate GARCH models.

Number of assets, k VEC DVEC CCC DCC

2 21 9 7 9
3 78 18 12 14
4 210 30 18 20
Table 2.1: Number of parameters to be estimated in multivariate GARCH models

Recent researches such as Sheppard (2003) and Sigurdarson (2007) have

shown the advantages with the CCC and DCC model, both because of their
`simplicity' (compared to the other) and behavior. I will therefore focus on
those two and here I will give a short introduction to them.

Constant conditional correlation, CCC

The model assumes k assets which all are conditionally normal distributed.
The covariance matrix, Σt , is dened as:

Σt = DtRDt (2.24)
2.6 ESTIMATING COVARIANCE AND CORRELATION 17

D
where t is the k × k diagonal volatility matrix, estimated from univariate
GARCH(1,1) process (one at a time as in section 2.4.3)
2
 
σ1,t 0 0
Dt =  0
..
. 0 

(2.25)
0 0 2
σk,t
and R is the correlation matrix. The maximum log-likelihood function for
the multivariate case when assuming normality can be written as:

T
` = −
1X
2
(k log(2π) + log(|Σt |) + rtΣ−1
0
t rt ) (2.26)
t=1
T
= −
1X
2
(k log(2π) + 2log(| Dt|) + log(|R|) + tΣ−1
0
t t )
t=1
T
Dt|) + zt R−1zt)
0
X
' − (2log(| (2.27)
t=1

since the constants don't matter in the maximization. Where zt ∼ N (0, R)

when t ∼ N (0, Σt ) are univariate GARCH standardized residuals. The uni-
variate volatility process can be any kind of GARCH process (E-GARCH,
GJR-GARCH or some other) and doesn't have to be the same for all as-
sets in the portfolio. The correlation is estimated as a constant historical
correlation.

R = (ρij ) (2.28)
The advantages with the CCC method is that it is much simpler then the
full multivariate GARCH model (VECH), a univariate GARCH process can
be used for the estimation and the formulation ensures positive deniteness
H
of t . The disadvantages is that it assumes the correlation to be constant,
?
which is unrealistic ( ).

Dynamic conditional covariance, DCC

Here the idea is the same in all steps as in CCC except to let the correlation
be time varying. A structure for estimating the dynamic correlation was
introduced by Engle (2002) as:

Qt = (1 − α − β)Q̄ + α(t−1t−1 + βQt−1) 0

(2.29)
where Q̄
is the k × k unconditional covariance matrix of t and α and β are
parameters > 0, that have to satisfy; α + β < 0. The correlation matrix t R
can now be obtained as:
18 FINANCIAL TIME SERIES MODELING

Rt = Q?−1
t QtQ?−1
t (2.30)
where Q?−1
t is obtained as:
 
√1 0 0
q11
Qt?−1 =  0
..
. 0


 (2.31)
0 0 √1
qkk

where qii are the i-th diagonal element of the matrix Qt , where i ∈ [1, k].

Some of the main advantages of the DCC model is that it doesn't rely on
a constant correlation matrix, can be calculated in steps, only few extra
parameters are needed and univariate estimate can be used for obtaining a
large part of the parameters.
Chapter 3

Parametric methods

3.1 Univariate parametric methods

In the parametric approach a distribution is tted to the data and the VaR
is estimated from the tted distribution. The parametric approach is more
appealing mathematically than the non-parametric, since it has a distribu-
tion (and density) function, which can give a relatively straight forward way
of calculating VaR. For example if the normal distribution ts the data well,
the VaR at α condence level can be calculated as

VaRα% = µ + σ · zα (3.1)
where zα comes from the standard normal distribution table (' 2.326 for
99% condence level), see gure 3.1.

Figure 3.1: VaR assuming normal distribution, LAIS long period data
20 PARAMETRIC METHODS

If the governing distribution is assumed be student's t-distribution, then the

VaR can be calculated as:
r
υ−2
VaRα% = µ + σ · tα (3.2)
υ
with υ degrees of freedom and tα comes from standard t-distribution ta-
ble. If for example the interest would be to nd VaR99% assuming that
t-distribution with 4 degrees of freedom, υ = 4, ts the data well while
µ = 0 and σ = 0.02, then it would be:

r
4−2
VaR99% = · 0.02 · 3.747 = 0.053
4
while normal distribution would have given VaR99% = 0.047. Here it is im-
portant to understand that volatility tends to be time varying, see section 2.3,
therefore the distribution of the residuals is expressed as for example:

rt | Θt ∼ N (0, σt2 )

where Θt is information set know at time t, for example past returns

{r0 , . . . , rt−1 } and/or past volatilities {σ0 , . . . , σt−1 }. Therefore it is said
that the returns rt are conditionally normal distributed, meaning that the
returns at time t are normal distributed conditional on the information set.

The choice of distribution can dier a lot. The most common is to assume
that normality (that is a normal distribution) is sucient to t the data
well, although this has been debated1 . Other common choice of distributions
are for example log-normal distribution and extreme value distribution. As
stated before, nancial data tend to be clustered, have fat tails and are pos-
sibly skewed and thus we would like to t a distribution to the data that can
show these characteristics.

3.2 Multivariate parametric methods

Multivariate parametric methods are analogous to the univariate case where

assumptions are made about the portfolio rather then a single asset, although
making the assumption that each asset in the portfolio is normal is the same
as assuming that the portfolio is normal distributed (holds only for normal
distribution). With k assets in a portfolio, assuming normal distribution,
the VaR can be obtained by:
1
see for example Hull (2007)
3.2 MULTIVARIATE PARAMETRIC METHODS 21

√
VaRα% = wTr + wTΣwzα (3.3)
see section 2.2 for further details. As before zα is obtained from the standard
normal distribution. Likewise the for student's t-distributed data portfolio
VaRα% can be obtained as:

υ − 2√ T
wr w w
r
VaRα% = T + Σ zα (3.4)
υ
Correlation can have much eect on VaR estimate which can be shown with
a simple example. Suppose we have two equal weighted (w1 = w2 = 0.5)
assets, A1 and A2 , in a portfolio which both are; A1 , A2 ∼ N(0, 1). The VaR
of each asset is obtained by equation 3.1. The portfolio volatility could be
calculated as:
r
1+ρ
q
2 2 2 2
σP = w1 σ1 + w2 σ2 + 2ρw1 w2 σ1 σ2 =
2
where ρ is the correlation between the two assets. Since µ1 = µ2 = 0 and
σ1 = σ2 = 1 the VaR estimate can then be written as:
r
1+ρ
VaRα% = rP + σP zα = zα (3.5)
2
The VaR estimate will be less then the individual VaR estimate for all values
of ρ except for ρ = 1 (when short-selling2 is not allowed). So generally it
could be stated that:
N
X
VaRportf olio < VaRi
i=1

where VaRi is the Value-at-Risk for asset i in the portfolio. This is one of
the fundamentals with portfolios, called to diversify (see section 2.2). Let's
take a simple example. Say that we have two equal weighted assets, A1 and
A2 , with are found to have following characteristics; A1 , A2 ∼ N(0, 0.022 )
and the correlation is found to be 0.6. We are interested in nding VaR99% .
Start by nding the portfolio volatility as:

p
σP = 0.52 · 0.022 + 0.52 · 0.022 + 2 · 0.6 · 0.5 · 0.5 · 0.02 · 0.02 = 0.0179

and therefore the VaR99% can be obtained as:

VaRα% = 0 + 0.0179 · 2.326 = 0.0416

2
In nance short-selling is the act of getting of selling a asset you do not own, in hope
repurchasing it back for a lower price.
22 PARAMETRIC METHODS

while student's t-distribution would have given VaR99% = 0.0474. Both give
lower VaR estimate then in the individual case, found for the univariate sec-
tion.

It is easy to see the advantages with the parametric approach, but as strong
as they can be they can be equally weak if assumptions about the tted
distributions are bad. Therefore obtaining the parametric assumption right
is the most important part of the parametric approaches (Dowd, 2005). The
main disadvantages with the parametric approaches is dealing with non-
linear instruments, such as options. In that case a linear approximation is
needed which cannot capture the behavior of the instrument and therefore
lead to large error.
Chapter 4

Non-parametric methods

4.1 Basic historical simulation

The attempt with the non-parametric models is to let the data (prot/loss
or return series) speak for themselves as much as possible, rather than some
tted distribution. The main assumption with non-parametric models is
that the recent past can be used to model the near future, meaning that
some past returns, say two years, are used to model tomorrow's VaR. This
way the data (returns) can accommodate any behavior, such as fat tails and
skewness, without having to make any distributional assumptions, if the past
returns showed that kind of behavior.

The most popular and known non-parametric model is the basic historical
simulation (HS). For the basic HS the general idea is to sort the historical
returns and estimate the VaR from the sorted historical returns at preferred
condence level α. Suppose for example we have 1000 observation of his-
torical returns and would want to estimate VaR for tomorrow at a 99%
condence level. We would start by sorting the data, then we would know
that 10 returns would lie in `left' of the VaR estimate (1% · 1000) and there-
fore a rational estimate of tomorrow's VaR would be the 11th one (or some
interpolation between the 10th and the 11th one). Meaning that 99% of the
time the loss is not more than the VaR99% .

If for example the 15 worst returns the last 4 years have been the following
(LAIS data):

{ -0.0693, -0.0685, -0.0645, -0.0604, -0.0569, -0.0548, -0.0537, -0.0509,

-0.0497, -0.0459, -0.0443, -0.0441, -0.0435, -0.0434, -0.0420, ... }

and the interest is to examine the 1% quantile for the period (i.e. the
VaR99% ) a reasonable estimate would be the 11th one (since 4 years are
24 NON-PARAMETRIC METHODS

equals to 1000 days1 and 10 observation are allowed to lie `left' of the esti-
mate)

VaR99% = 0.0443
therefore it could be said that we are 99% sure of not getting worse re-
turn than 4.43% for tomorrow. Graphically this can be done by plotting a
histogram and examine the tail as is shown in gure 4.1

(a) Histogram LAIS data (b) The tail

Figure 4.1: Shows how VaR is obtained graphically with the basic HS

The advantages with the basic HS is that it is really simple. The main dis-
advantages is that in the basic HS approach all observations are treated the
same, that is all observations have the same weight (called equally-weighted).
If all observation are treated the same a shock on the market today could
be `averaged' out if the sample size is large enough and not noticed at all
except at high condence levels. In other words risk grows without VaR
showing it. Another example could be a major nancial crisis in the past.
This shock could produce high VaR estimate while it is in the sample space,
called `ghost eects', and then produce a jump in the estimate when it falls
out of the sample space, see section 2.4.1.

One of the most attractive facts with the non-parametric approaches is that
they can be applied as well for a multivariate case as well as univariate case
and there is no need for estimating a variance-covariance matrix Σ, which is
often the `dicult' part of a multivariate estimation.

There are several implementations that can be added to the basic HS, such as
bootstrapping (re-sampling the data over and over) and combination of non-
parametric density function (for being able to treat the data as continuous,
not discrete). One of the most popular implementation to the basic HS
is to weight the data certain way so that not all observations are treated
1
There are roughly 250 trading days each year
4.2 WEIGHTED HISTORICAL SIMULATION 25

equally. These method's are called `Weighted Historical Simulation' and can
be thought as `semi-parametric method' since they combine features of both
non-parametric and parametric methods (Dowd, 2005).

4.2 Weighted historical simulation

There are various ways to adjust the data to overcome problems such as
`ghost eects'. Here I will introduce few of them.

4.2.1 Age weighted historical simulation (Age-WHS)

Observations are given weight according to their age as them name implies,
so that recent (in time) observation will have more weight than older ones.
Boudoukh, Richardson and Whitelaw (1998) introduced a formula for calcu-
lating observations weight as function of the decay factor λ

1−λ
w(i) = λi−1 (4.1)
1 − λn
where w(i) is the weight to i days old observation (i.e w(1) is the weight for
the newest observation) and λ is the rate of decay, 0 < λ < 1. High λ (close
to 1) gives slow rate of decay and low λ gives high rate of decay, Boudoukh
et al. (1998) recommend using λ = 0.98. As said returns are given weight
according to their age, then the returns are sorted. Their weight's are then
summed up, until preferred condence level is achieved and corresponding
return will give the VaR estimate.

Boudoukh et al. (1998) age-weighting formula is a nice generalization of

basic equal weighted HS (the same as λ → 1) and gives a more responsive
VaR estimates with a well chosen decay factor, λ. The method is also helpful
in reducing ghost eects, since old observation will have had weight close to
zero and a large jump is thus less likely to be observed in the sample space.

4.2.2 Volatility weighted historical simulation (VWHS)

Proposed by Hull and White (1998) to update the return with volatility
changes. As pointed out by them if for example the volatility on market
today is 1.5% per day on average and two months ago it was 1% on average
then the `old' volatility will give an underestimate for changes in near future
and vice versa. They therefore introduced a formula for updating return
with volatility as:

rt
rt? = σT · (4.2)
σt
26 NON-PARAMETRIC METHODS

where σt is the estimated daily volatility at time t, rt is the historical return

and σT is the most recent estimate of volatility made at the end of date T
Hull and White (1998). More generally, daily returns are standardized with
their volatility and then scaled with current volatility. This approach is a
straight forward extension of the basic HS where volatility uctuations has
been taken into account in estimating VaR. Advanced techniques as EWMA
or GARCH process could be used to estimate the volatility process to explain
for example volatility clustering. This method has proved a higher estimate
of VaR then basic HS (Dowd, 2005).

4.2.3 Filtered historical simulation (FHS)

Approach developed by Barone-Adesi, Bourgoin and Giannopoulos (1998)2
has becoming more and more popular among risk analyzers. First a volatil-
ity process is tted to the return (EWMA, GARCH or any other), then the
returns (or the residuals, t ) are standardized with the volatility estimate
as zt = rt /σt . Here heteroscedasticity should be removed (can be checked
by for example looking at autocorrelation plot, see section 5.2.1). The stan-
dardized returns zt are then bootstrapped. Bootstrapping involves drawing
observations randomly from the sample, until the original sample size is
reached. This is done N times (typically 500, 1000 or 5000 times). It should
be mentioned that in the bootstrapping procedure the same observation can
be drawn more often than once. Finally the new samples are scaled with
current volatility σT , and then each sample can give an estimate of tomor-
row's return and the VaR can be obtained at preferred condence level. The
procedure is shown graphically on gure 4.2.

Figure 4.2: Bootstrap procedure

2
and other papers by the same authors
Chapter 5

Analysis

5.1 Methodology

All methods described in chapters 3 and 4 are now compared with the goal
of nding the method that gives fewest exceptions and lowest VaR estimate
while fullling regulations stipulated by the Basel Committee (2006). By
stating as low as possible, the aim is to minimize the regulatory capital, as
that is depended on the VaR gure if the internal approach as set forth by
the Basel Committee is used. If VaR limit is set very high or overestimated
for some reasons, more capital will have to be kept in reserve.

All calculation were done in Matlab. For the multivariate GARCH cal-
culations the UCSD Matlab toolbox by Kevin Sheppard was used1 . As I
mentioned in section 1.2, VaR has the unit of money, although in the analy-
sis I will calculate VaR as a percentage of return for the sake of comparison
between dierent instruments. Before going further I will give a short intro-
duction to the regulations set by the Basel Committee (see Basel Committee
(2006)).

5.1.1 Backtesting

Backtesting is a test performed to check the accuracy of internal VaR model,

historically, meaning that over some period (at least 1 year) estimated daily
VaR and actual P/L series2 are compared and exceptions, when
−VaRα > Rt , are counted. The internal model is given as:

60
k X
max{VaRt , VaRt−i+1 } (5.1)
60
i=1

1
see http://www.kevinsheppard.com/wiki/UCSD_GARCH
2
In my case return series, since VaR is in percentage
28 ANALYSIS

where VaRt is the 10-day VaR estimate for day t, and k is known as the hys-
teria factor which is determined by the bank's backtesting result (somewhere
between 3 and 4, see section 5.1.2). The reason for using the 10-day VaR, or
10-day holding period, is that it may take that long time to liquidate a posi-
tion3 . For interpolating the 1-day VaR to 10-day VaR the Basel Committee
allows that the infamous `square-root of time rule' should be used. The rule
is given as:
√
N-day VaR = N × 1-day VaR (5.2)
The origin of this rule comes from that if you have 2 independent and iden-
tical normal distributed (normal iid) variables xt and xt+1 with variance σ 2 ,
the sum of their variance will be:

V ar(xt + xt+1 ) = V ar(xt ) + V ar(xt+1 ) = 2σ 2

√
Which implies that their volatility is scaled by 2. However this holds only
?
if all observations are assumed to be normal iid, else it is a approximation ( ).

Stylized facts such as heteroscedasticity violates the normal iid assumptions

and therefore many have debated Basel Committee's recommendation using
the square-root of time rule, see for example Daníelsson and Zigrand (2005).
The most straight forward way of calculating a 10-day VaR is to use 10 times
more data and divide it into 10-day intervals instead of 1-day. This means
that for the short period 20 years of data would be needed and for the long
period 40 years of data would be needed. The problem is that there are
not that many stocks, indices or other nancial time series with such a long
history, and those who exist are likely to have changed drastically over last
decades (it could thus be debated to use the same model for such a long
period). Because of these debates and approximations I will concentrate on
calculating 1-day VaR and skip any scaling or interpolations to other time
intervals.

5.1.2 Basel zones

The Basel Committee requires that models are at least 99% accurate and
uses a general hypothesis test, in order to balance two types of errors; (I) the
possibility that an accurate risk model would be classied as inaccurate on
the basis of its backtesting result, and (II) the possibility that an inaccurate
model would not be classied that way based on its backtesting result. The
Basel Committee categorizes backtesting result into 3 zones to minimize the
type I and type II errors. Green zone, indicating that the model is probably
good, minimal probabilities of type I error, yellow zone indicating uncertainty
3
Meaning that it could take 10 days to sell the position.
5.1 METHODOLOGY 29

and possibilities of both types of error and red zone indicating a probably
bad model with a minimal chance of type II error. If the model ends in
yellow zone it is up to the nancial institution to prove it's goodness (Basel
Committee, 2006). Table (1.2) shows the categorization for 1 year of data.

Number of Increase in Cumulative

Zone
exceptions hysteria factor probability
0 0 8,11 %
1 0 28,58%
Green Zone 2 0 54,32%
3 0 75,81%
4 0 89,22%
5 0,40 95,88%
6 0,50 98,63%
Yellow Zone 7 0,65 99,60%
8 0,75 99,89%
9 0,85 99,97%
Red zone 10 or more 1 99,99%
Table 5.1: Backtesting zones in Basel accord based on 250 observations

Increase in hysteria factor is what adds to the default, k = 3, in equation 5.1.

The cumulative probability is the binomial probability of getting the num-
ber of exceptions or fewer, see equation 5.3. For example the probability of
getting 5 exceptions or less is equal to 95,88%, when dealing with 1 year of
data and 99% condence.

To interpolate the table to other time intervals the boundaries between green
and yellow zone is when the cumulative binomial distribution is equal to/or
exceeds 95% and the boundaries between yellow and red zone 99,99%. Lim-
its for the short period are therefore, up to 8 exceptions is green zone, up to
14 is yellow zone and 15 or more will be red zone. For the long period the
limits will be, up to 14 is green zone, up to 22 is yellow zone and 23 or more
will be red zone.

5.1.3 The basic frequency test

The basic frequency test was proposed by Kupiec (1995) where the idea was
to check with simple hypothesis testing whether to accept or reject a model.
When checking number of exceptions the cumulative binomial distribution
can be used:

K
X n!
P (K ≤ x) = pi (1 − p)n−i (5.3)
i!(n − i)!
i=0
30 ANALYSIS

where n is the number of observations (days), K is the number of exceptions

and p is the probability (p = 1 − α). In hypothesis testing the idea is to
propose a `null-hypothesis' H0 which is assumed to be true (in this case ex-
pected number of exceptions) and a `alternative-hypothesis' H1 which is the
actual outcome from the model. If for example the total number of observa-
tion n = 1000, the actual number of exceptions K = 20, the condence level
α = 0.99 (⇒ p = 0.01) therefore the expected number of exceptions would
be 10 (1000 × 0,01) which is less then 20, the hypothesis test proposed could
be:
(
H0 : p = 0.01
H1 : p > 0.01
More generally, the idea is to check whether the model used for obtaining K
exceptions is ok, when the expected number of exceptions is n·p. This is done
by putting the values of n, K and p are put into equation 5.3 and the results
checked. P (K ≥ 20) = 1 − P (K ≤ 19) = 0.0033. Normally 5% condence is
used for validating the statistical test (Hull, 2007), therefore the `alternative-
hypothesis' would be rejected if P (K = x) is less than the condence level
of the test. In this case P (K ≥ 20) = 0.0033 < 0.05, therefore the null-
hypothesis is rejected, which leads to that the model used for calculating
this number of exceptions is rejected. If the number of exceptions had been
less than the expected number of exceptions, say K = 7, the hypothesis had
looked like:
(
H0 : p = 0.01
H1 : p < 0.01
and the result had been P (K ≤ 7) = 0.2189 > 0.05, and thus the `alternative-
hypothesis' is not rejected and therefore the model used is not rejected (found
to be ok).

By using a combination of the two theories (Basel zones and frequency test-
ing), we can sort the data into zones and also reject models who have too few
exceptions according to 5% condence interval, for preventing that the VaR
limit is set to high. This will be used as a main criteria between methods.
If models seem to be giving similar results, further analyzing criteria can be
achieved by looking at means, standard deviations, minimum values and the
models complexity.
5.2 UNIVARIATE CASE 31

5.2 Univariate case

First lets look at univariate analysis. Univariate means that there is only one
asset (stock, currency pair, index) underlying. First I will present the results
for the parametric approach, then non-parametric and nally compare them
together.

5.2.1 Parametric methods

Before modeling the data using parametric approach it is good to try to

get as much information from the data as possible, to make as rational
decisions as possible. One of those things is to examine the autocorrelation
of the data which can give information about repeating patterns in the data.
Autocorrelation is given as:

E[(Xi − µ)(Xi+k − µ)]

R(k) = (5.4)
σ2
where k is the lag between observations. By calculation autocorrelation for
the rst two moments, i.e. rt and rt2 (or t and 2t ), repeating patterns in
the mean and variance can be examined. As has been said before, one of
the most stylized facts about nancial time series is that they tend to have
autocorrelation in the second moment (the variance, called heteroscedastic-
ity). This fact cannot be dismissed, and therefore tting a volatility process
that can explain these characteristics should be rational. As was explained
in section 2.4.3 a GARCH process is good in simulating heteroscedasticity
and could thus be a wise choice.

Here I shown the autocorrelation for the rst two moments for LAIS data
long period, see section 1.3 for information on time series.

(a) Autocorrelation in t (b) Autocorrelation in 2t

Figure 5.1: Examining autocorrelation in rst two moments of the residuals, LAIS
long period. The red bars show the autocorrelation for respective lag and the blue
lines are the 95% condence interval.
32 ANALYSIS

From gure 5.1 there is no indicator of autocorrelation in the rst moment,

the mean (gure (a)) while there is a strong indicator of autocorrelation in
the second moment, the variance (gure (b)), almost all lags have autocorre-
lation higher then the 95% condence interval. A GARCH process was tted
to the data and then the residuals standardized (zt = t /σt ). By examining
autocorrelation in the standardized residuals (gure 5.2), it is possible to see
whether the tted volatility process did a good job in removing the autocor-
relation or not.

(a) Autocorrelation in zt (b) Autocorrelation in zt2

Figure 5.2: Here the autocorrelation is shown for the standardized residuals, LAIS
long period.

Figure 5.2 shows that there is no longer any clear autocorrelation in the stan-
dardized residuals and the next step would be to make some assumptions
about the distribution. Similar results were achieved for other assets (see
appendix A.1.1).

By looking at qq-plots, which are plots where empirical quantiles are plotted
against theoretical quantiles, it is possible to check wheter a certain distribu-
tion ts the data. If a chosen distribution ts the data well the qq-plot should
form a straight line. As before I show the results for LAIS long period.
5.2 UNIVARIATE CASE 33

(a) Normal distribution (b) Student's t-distribution

Figure 5.3: Estimating distribution, LAIS long period

From the qq-plots it's clear that Student's t-distribution (with 4 degrees of
freedom) ts the LAIS data better then the normal distribution and should
therefore be the choice (between those two), although I try both. Figure 5.4
shows the dierence between a normal distribution and Student-t distribu-
tion with 4 degrees of freedom.

(a) Normal vs. t distribution (b) Tail behavior

Figure 5.4: The dierence between normal distribution and Student's t-distribution
with 4 degrees of freedom

Figure 5.4 shows the characteristic dierence between normal and student's
t-distribution (with 4 degrees of freedom).The gure shows that student's
t-distribution has fatter tails then the normal distribution, which leads to
higher VaR estimate, since the estimate is equal to the α% area under the
curve.

Now simulation can take place. Methods described in chapter 3 are calcu-
lated and results for LAIS are plotted in the following page. First gures for
the long period, gures 5.5 and 5.6 then gures 5.7 and 5.8 shows how the
methods react dierently to shock (zoomed in on a shock). Figures for other
stocks are shown in appendix A.1.1. For the EWMA model decay factor
λ = 0.94 is used for stock and λ = 0.97 is used for the currency pair.
34 ANALYSIS

Figure 5.5: LAIS long period, assuming normal distribution

Figure 5.6: LAIS long period, assuming student's t-distribution

5.2 UNIVARIATE CASE 35

Figure 5.7: LAIS shock behavior, assuming normal distribution

Figure 5.8: LAIS shock behavior, assuming student's t-distribution

36 ANALYSIS

Exceptions were counted and are presented in tables 5.2 to 5.5. In the tables
the green color stands for green zone, no color for yellow zone and red color
for red zone. Outlined numbers are those where a model has been rejected
due to hypothesis testing. The number in brackets is observed exceptions
divided by expected number of exceptions.
Method LAIS MARL ERIC NDA ISXI15 OMX
EWMA 8 (1,6) 9 (1,8) 11 (2,2) 9 (1,8) 11 (2,2) 14 (2,8)
GARCH(1,1) 8 (1,6) 8 (1,6) 13 (2,6) 9 (1,8) 13 (2,6) 16 (3,2)
E-GARCH 6 (1,2) 8 (1,6) 17 (3,4) 8 (1,6) 12 (2,4) 14 (2,8)
GJR-GARCH 6 (1,2) 8 (1,6) 12 (2,4) 9 (1,8) 12 (1,4) 14 (2,8)

Table 5.2: Exceptions for the short period, assuming normal distribution

Method LAIS MARL ERIC NDA ISXI15 OMX

EWMA 6 (1,2) 5 (1,0) 9 (1,8) 8 (1,6) 6 (1,2) 10 (2,0)
GARCH(1,1) 4 (0,8) 3 (0,6) 10 (2,0) 7 (1,4) 7 (1,4) 9 (1,8)
E-GARCH 2 (0,4) 2 (0,4) 7 (1,4) 7 (1,4) 9 (1,8) 9 (1,8)
GJR-GARCH 2 (0,4) 3 (0,6) 9 (1,8) 6 (1,2) 7 (1,4) 10 (2,0)

Table 5.3: Exceptions for the short period, assuming student's t-distribution

Method LAIS MARL ERIC NDA ISXI15 OMX

EWMA 16 (1,6) 13 (1,3) 23 (2,3) 18 (1,8) 20 (2,1) 28 (2,8)
GARCH(1,1) 15 (1,5) 11 (1,1) 20 (2,0) 16 (1,6) 26 (2,6) 25 (2,5)
E-GARCH 16 (1,6) 13 (1,3) 24 (2,4) 16 (1,6) 25 (2,5) 23 (2,3)
GJR-GARCH 16 (1,6) 13 (1,3) 18 (1,8) 15 (1,5) 25 (2,5) 23 (2,3)

Table 5.4: Exceptions for the long period, assuming normal distribution

Method LAIS MARL ERIC NDA ISXI15 OMX

EWMA 10 (1,0) 8 (0,8) 13 (1,3) 13 (1,3) 13 (1,3) 18 (1,8)
GARCH(1,1) 7 (0,7) 3 (0,3) 15 (1,5) 10 (1,0) 13 (1,3) 16 (1,6)
E-GARCH 8 (0,8) 2 (0,2) 12 (1,2) 11 (1,1) 17 (1,7) 16 (1,6)
GJR-GARCH 8 (0,8) 3 (0,3) 13 (1,3) 10 (1,0) 14 (1,4) 16 (1,6)

Table 5.5: Exceptions for the long period, assuming student's t-distribution

All models show improvements when t-distribution is assumed to t the data

rather then a normal distribution, except for GARCH methods on MARL
data. The reason for that can be instability and that the model fails on
nding a local maxima in the log-likelihood function. A further examination
of means, standard deviation and minimum values of the VaR estimate are
given in tables 5.6 - 5.9.
5.2 UNIVARIATE CASE 37

LAIS MARL
Method Exceptions VaR σ VaR max(VaR) Exceptions VaR σ VaR max(VaR)
EWMA 8 0.0324 0.0077 0.0529 9 0.0244 0.0060 0.0558
GARCH(1,1) 8 0.0338 0.0068 0.0526 8 0.0268 0.0065 0.1203
E-GARCH 6 0.0359 0.0087 0.0599 8 0.0268 0.0055 0.0795
GJR-GARCH 6 0.0346 0.0073 0.0532 8 0.0262 0.0063 0.1094

ERIC NDA
Method Exceptions VaR σ VaR max(VaR) Exceptions VaR σ VaR max(VaR)
EWMA 11 0.0513 0.0248 0.1587 9 0.0367 0.0108 0.0667
GARCH(1,1) 13 - - 0.1242 9 0.0345 0.0094 0.0734
E-GARCH 17 0.0428 0.0098 0.0728 8 0.0346 0.0086 0.0732
GJR-GARCH 12 - - 0.1315 9 0.0344 0.0098 0.0809

ISXI15 OMX
Method Exceptions VaR σ VaR max(VaR) Exceptions VaR σ VaR max(VaR)
EWMA 11 0.0272 0.0107 0.0557 14 0.0307 0.0083 0.0548
GARCH(1,1) 13 0.0274 0.0123 0.0755 16 0.0285 0.0085 0.0636
E-GARCH 12 0.0270 0.0110 0.0636 14 0.0277 0.0094 0.0592
GJR-GARCH 12 0.0273 0.0120 0.0678 14 0.0280 0.0100 0.0656

Table 5.6: Detailed information assuming normal distribution, short period

LAIS MARL
Method Exceptions VaR σ VaR max(VaR) Exceptions VaR σ VaR max(VaR)
EWMA 6 0.0369 0.0088 0.0602 5 0.0278 0.0068 0.0635
GARCH(1,) 4 0.0399 0.0091 0.0657 3 0.0375 0.0132 0.1771
E-GARCH 2 0.0418 0.0105 0.0685 2 0.0662 0.0384 0.2276
GJR-GARCH 2 0.0408 0.0095 0.0681 3 0.0363 0.0118 0.1445

ERIC NDA
Method Exceptions VaR σ VaR max(VaR) Exceptions VaR σ VaR max(VaR)
EWMA 9 0.0585 0.0282 0.1807 8 0.0418 0.0123 0.0760
GARCH(1,1) 10 0.0556 0.0208 0.1641 7 0.0395 0.0116 0.0922
E-GARCH 7 0.0551 0.0203 0.1466 7 0.0399 0.0112 0.0923
GJR-GARCH 9 0.0566 0.0263 0.2058 6 0.0395 0.0122 0.1001

ISXI15 OMX
Method Exceptions VaR σ VaR max(VaR) Exceptions VaR σ VaR max(VaR)
EWMA 6 0.0310 0.0122 0.0634 10 0.0349 0.0094 0.0625
GARCH(1,1) 7 0.0316 0.0146 0.0891 9 0.0330 0.0103 0.0762
E-GARCH 9 0.0316 0.0135 0.0772 9 0.0320 0.0112 0.0689
GJR-GARCH 7 0.0316 0.0144 0.0804 10 0.0324 0.0119 0.0774

Table 5.7: Detailed information assuming student's t-distribution, short period

38 ANALYSIS

LAIS MARL
Method Exceptions VaR σ VaR max(VaR) Exceptions VaR σ VaR max(VaR)
EWMA 16 0.0363 0.0121 0.0796 13 0.0277 0.0093 0.0703
GARCH(1,1) 15 0.0370 0.0117 0.0928 11 0.0310 0.0078 0.1203
E-GARCH 16 0.0375 0.0111 0.0775 13 0.0305 0.0074 0.0822
GJR-GARCH 16 - - 0.0799 13 0.0307 0.0080 0.1094

ERIC NDA
Method Exceptions VaR σ VaR max(VaR) Exceptions VaR σ VaR max(VaR)
EWMA 23 0.0464 0.0207 0.1587 18 0.0317 0.0109 0.0667
GARCH(1,1) 20 - - 0.1242 16 0.0319 0.0087 0.0734
E-GARCH 24 0.0443 0.0136 0.0885 16 0.0315 0.0083 0.0732
GJR-GARCH 18 - - 0.1315 15 0.0318 0.0089 0.0809

ISXI15 OMX
Method Exceptions VaR σ VaR max(VaR) Exceptions VaR σ VaR max(VaR)
EWMA 20 0.0257 0.0102 0.0557 28 0.0254 0.0097 0.0556
GARCH(1,1) 26 0.0254 0.0105 0.0755 25 0.0252 0.0084 0.0658
E-GARCH 25 0.0254 0.0093 0.0636 23 0.0247 0.0084 0.0592
GJR-GARCH 25 0.0253 0.0102 0.0678 23 0.0250 0.0092 0.0686

Table 5.8: Detailed information assuming normal distribution, long period

LAIS MARL
Method Exceptions VaR σ VaR max(VaR) Exceptions VaR σ VaR max(VaR)
EWMA 10 0.0413 0.0138 0.0906 8 0.0316 0.0106 0.0801
GARCH(1,1) 7 0.0449 0.0149 0.1033 3 0.0476 0.0179 0.2257
E-GARCH 8 0.0460 0.0145 0.1012 2 0.1192 0.1659 1.9051
GJR-GARCH 8 0.0457 0.0147 0.1082 3 - - 0.2439

ERIC NDA
Method Exceptions VaR σ VaR max(VaR) Exceptions VaR σ VaR max(VaR)
EWMA 13 0.0529 0.0236 0.1807 13 0.0362 0.0124 0.0760
GARCH(1,1) 15 0.0526 0.0188 0.1641 10 0.0366 0.0101 0.0922
E-GARCH 12 - - 0.1466 11 0.0366 0.0100 0.0923
GJR-GARCH 13 - - 0.2058 10 0.0366 0.0107 0.1001

ISXI15 OMX
Method Exceptions VaR σ VaR max(VaR) Exceptions VaR σ VaR max(VaR)
EWMA 13 0.0293 0.0116 0.0634 18 0.0289 0.0110 0.0633
GARCH(1,1) 13 0.0300 0.0132 0.0891 16 0.0288 0.0098 0.0762
E-GARCH 17 0.0300 0.0125 0.0772 16 0.0283 0.0099 0.0689
GJR-GARCH 14 - - 0.0823 16 0.0288 0.0108 0.0774

Table 5.9: Detailed information assuming student's t distribution, long period

5.2 UNIVARIATE CASE 39

The models give fairly similar results. EWMA, E-GARCH and GJR-GARCH
give 11 green zones out of 24 and GARCH(1,1) 10 out of 24. When means,
standard deviations and minimum values are examined, it comes clear that
EWMA usually has the lowest mean and minimum values while GARCH
based methods have the lowest standard deviations. This supports descrip-
tions given in section 2.3, i.e. GARCH models are quicker to simulate market
condition and spike higher, while EWMA is slower to follow market uctu-
ations (see gures 5.7 and 5.8). The GARCH models are more sensitive
than the EWMA model, especially GJR-GARCH, and fails on getting re-
sults when high jumps occur in time series (MARL and ERIC). Finally I
compare the parametric results with the extra long currency pair time series
(10 years). Exceptions, means, standard deviation and minimum values are
given in table 5.10.

Normal distribution Student's t-distribution

Method Exceptions VaR σ VaR min(VaR) Exceptions VaR σ VaR min(VaR)
EWMA 44 0.0179 0.0076 0.0591 22 0.0204 0.0086 0.0673
GARCH(1,1) 31 0.0179 0.0073 0.1036 14 0.0206 0.0081 0.0855
E-GARCH 42 0.0172 0.0081 0.1537 18 0.0199 0.0079 0.0853
GJR-GARCH 32 - - 0.1389 18 - - 0.1046

Table 5.10: Detailed information, currency pair

For the currency pair GARCH and GJR-GARCH has green zones for both
distribution, while EWMA and E-GARCH has only green zone when t-
distribution is assumed to be the governing distribution of the residuals.
Means, standard deviation and maximum values are pretty similar between
the methods. Here it is worth mention that although a frequency test would
suggest to reject models who have 17 or fewer exceptions they will not be
rejected because of central limit theorem4 .

Altogether none of the method is showing any superior characteristics. E-

GARCH and GJR-GARCH do not improve the regular GARCH signicantly
and seem to be more sensitive than the regular one. E-GARCH has though
the tendency to lower the VaR, which is appealing. Since GARCH is known
to be better in explaining fat tails and heteroscedasticity it is recommended
as a rst choice, but since EWMA isn't giving any fewer green zones it is
recommended as second choice. It could thus be wise to model both for the
sake of comparison and to minimize model and implementation risk.

4
Central limit theorem indicates that if you for example toss a fair dice 10 times the
probability of getting say 7 heads isn't that unlikely, but as you toss the dice more often
a fair dice would give equal probabilities of getting head and tails. In other words the
probability of getting 7 heads in 10 tosses isn't the same as getting 700 heads out of 1000
tosses. The limit gets narrower.
40 ANALYSIS

5.2.2 Non-parametric methods

For the non-parametric case any distributional assumption aren't necessary
and therefore a pre analysis on the data isn't needed. Calculation for all
the non-parametric method's described in the chapter 4 were done for the
data. For the FHS simulation a GARCH(1,1) process was used to standard-
ize returns and 1000 bootstraps were used. The mean of those 1000 were
taken as the nal solution. Solutions for all calculation are in tables. Here
I present only LAIS gures (see gures 5.9 and 5.10) , the rest is in appendix.

Figure 5.9: LAIS long period

5.2 UNIVARIATE CASE 41

Figure 5.10: LAIS short period

Exceptions where counted and are displayed in tables 5.11 and 5.12. Over-
lined numbers indicates that the model is rejected due to hypothesis testing
(too few exceptions). Color indicates which zone the model ends in (green
means green zone, no color means yellow zone and red means red zone).

Method LAIS MARL ERIC NDA ISXI15 OMX

Basic HS 1 (0.2) 4 (0.8) 8 (1.6) 6 (1.2) 7 (1.4) 7 (1.4)

Age WHS 13 (2.6) 13 (2.6) 12 (2.4) 12 (2.4) 15 (3.0) 14 (2.8)
EWMA VWHS 11 (2.2) 13 (2.6) 11 (2.2) 7 (1.4) 11 (2.2) 8 (1.6)
GARCH(1,1) VWHS 7 (1.4) 6 (1.2) 10 (2.0) 8 (1.6) 8 (1.6) 8 (1.6)
E-GARCH VWHS 5 (1.0) 8 (1.6) 11 (2.2) 6 (1.2) 9 (1.8) 10 (2.0)
FHS 7 (1.4) 9 (1.8) 9 (1.8) 8 (1.6) 8 (1.6) 7 (1.4)

Table 5.11: Exceptions for the short period

Method LAIS MARL ERIC NDA ISXI15 OMX

Basic HS 15 (1,5) 8 (0,8) 13 (1,3) 13 (1,3) 19 (1,9) 14 (1,4)

Age WHS 27 (2,7) 30 (3,0) 26 (2,6) 27 (2,7) 31 (3,1) 31 (3,1)
EWMA VWHS 22 (2,2) 21 (2,1) 23 (2,3) 18 (1,8) 27 (2,7) 22 (2,2)
GARCH(1,1) VWHS 17 (1,7) 11 (1,1) 17 (1,7) 13 (1,3) 24 (2,4) 15 (1,5)
E-GARCH VWHS 16 (1,6) 13 (1,3) 17 (1,7) 11 (1,1) 23 (2,3) 17 (1,7)
FHS 17 (1,7) 11 (1,1) 15 (1,5) 13 (1,3) 22 (2,2) 15 (1,5)

Table 5.12: Exceptions for the long period

Means, standard deviations and maximum values are given in table 5.13 and
5.14.
42 ANALYSIS

LAIS MARL
Method Exceptions VaR σ VaR max(VaR) Exceptions VaR σ VaR max(VaR)
Basic HS 1 0.0475 0.0048 0.0563 4 0.0278 0.0018 0.0316
Age WHS 13 0.0296 0.0096 0.0450 13 0.0218 0.0078 0.0352
EWMA VWHS 11 0.0305 0.0076 0.0490 13 0.0223 0.0061 0.0545
GARCH(1,1) VWHS 7 0.0365 0.0074 0.0578 6 0.0273 0.0071 0.1285
E-GARCH VWHS 5 0.0383 0.0090 0.0632 8 0.0264 0.0056 0.0862
FHS 7 0.0358 0.0075 0.0581 7 0.0270 0.0069 0.1259

ERIC NDA
Method Exceptions VaR σ VaR max(VaR) Exceptions VaR σ VaR max(VaR)
Basic HS 8 0.0525 0.0078 0.0641 6 0.0351 0.0042 0.0453
Age WHS 12 0.0654 0.0669 0.3124 12 0.0337 0.0119 0.0572
EWMA VWHS 11 0.0539 0.0311 0.1940 7 0.0387 0.0130 0.0750
GARCH(1,1) VWHS 10 0.0518 0.0157 0.1356 8 0.0388 0.0118 0.0888
E-GARCH VWHS 11 0.0504 0.0135 0.0943 6 0.0387 0.0097 0.0833
FHS 9 0.0522 0.0169 0.1376 8 0.0381 0.0113 0.0861

ISXI15 OMX
Method Exceptions VaR σ VaR max(VaR) Exceptions VaR σ VaR max(VaR)
Basic HS 7 0.0338 0.0026 0.0395 7 0.0344 0.0039 0.0394
Age WHS 15 0.0263 0.0112 0.0470 14 0.0316 0.0098 0.0500
EWMA VWHS 11 0.0274 0.0113 0.0583 8 0.0353 0.0100 0.0647
GARCH(1,1) VWHS 8 0.0302 0.0139 0.0849 8 0.0358 0.0105 0.0773
E-GARCH VWHS 9 0.0304 0.0125 0.0733 10 0.0323 0.0102 0.0658
FHS 8 0.0301 0.0138 0.0838 7 0.0361 0.0115 0.0819

Table 5.13: Detailed information, short period

LAIS MARL
Method Exceptions VaR σ VaR max(VaR) Exceptions VaR σ VaR max(VaR)
Basic HS 15 0.0414 0.0096 0.0563 8 0.0336 0.0075 0.0514
Age WHS 27 0.0331 0.0154 0.0717 30 0.0239 0.0098 0.0455
EWMA VWHS 22 0.0332 0.0114 0.0724 21 0.0247 0.0084 0.0573
GARCH(1,1) VWHS 17 0.0361 0.0118 0.0821 11 0.0308 0.0080 0.1285
E-GARCH VWHS 16 0.0378 0.0119 0.0751 13 0.0305 0.0082 0.0862
FHS 18 0.0358 0.0118 0.0817 11 0.0309 0.0084 0.1257

ERIC NDA
Method Exceptions VaR σ VaR max(VaR) Exceptions VaR σ VaR max(VaR)
Basic HS 13 0.0634 0.0191 0.1429 13 0.0370 0.0091 0.0714
Age WHS 26 0.0551 0.0511 0.3124 27 0.0288 0.0119 0.0572
EWMA VWHS 23 0.0464 0.0251 0.1940 18 0.0326 0.0124 0.0750
GARCH(1,1) VWHS 17 0.0511 0.0165 0.1356 13 0.0358 0.0104 0.0888
E-GARCH VWHS 17 0.0498 0.0163 0.0989 11 0.0354 0.0092 0.0833
FHS 15 0.0516 0.0170 0.1372 13 0.0355 0.0099 0.0860

ISXI15 OMX
Method Exceptions VaR σ VaR max(VaR) Exceptions VaR σ VaR max(VaR)
Basic HS 19 0.0290 0.0061 0.0395 14 0.0310 0.0058 0.0433
Age WHS 31 0.0246 0.0118 0.0487 31 0.0261 0.0108 0.0500
EWMA VWHS 27 0.0243 0.0111 0.0583 22 0.0274 0.0124 0.0647
GARCH(1,1) VWHS 24 0.0266 0.0124 0.0849 15 0.0302 0.0111 0.0816
E-GARCH VWHS 23 0.0265 0.0114 0.0733 17 0.0282 0.0098 0.0692
FHS 22 0.0266 0.0123 0.0831 15 0.0304 0.0116 0.0818

Table 5.14: Detailed information, long period

The basic HS has the most green zone results 8 out of 12, then GARCH
VWHS with 7 out of 12 and then FHS with 6 out of 12. The only model
rejected due to hypothesis testing is the basic HS, rejected twice, which also
has the highest means in all cases for the long period and for half of the
cases in the short period (indicating that basic HS is overestimating VaR,
can be checked by looking at plots 5.9 and 5.10). As for the parametric case
5.2 UNIVARIATE CASE 43

GARCH VWHS has the highest maximum values in most of the times, al-
though the dierence between the VWHS isn't that much most of the time.
The Age WHS gives the poorest result, a red zone for all the cases in the
long period.

Results for the extra long currency pair modeling are given in table 5.15.
USDISK
Method Exceptions VaR σ VaR max(VaR)

Basic HS 44 0.0177 0.0036 0.0277

Age WHS 57 0.0172 0.0077 0.0523
EWMA VWHS 38 0.0178 0.0057 0.0408
GARCH(1,1) VWHS 29 0.0188 0.0067 0.0992
E-GARCH VWHS 37 0.0180 0.0077 0.1514
FHS 32 0.0186 0.0067 0.0976

Table 5.15: Detailed information, currency pair

Result for the currency pair supports the result from the stock and index
analysis. GARCH VWHS and FHS have green zones, while AGE WHS gives
red zone and the rest yellow zone.

Since the basic HS is rejected twice in the stock and index analysis (and has
the highest means), GARCH VWHS is regarded as the best of the univari-
ate non-parametric models, since it gives more green zones then the other
models and is slightly simpler then the FHS.
44 ANALYSIS

5.3 Multivariate case

As for the univariate case I rst present the results for the parametric ap-
proach and then the non-parametric approach. For the parametric approach
methods described in section 3.2 are calculated with covariance estimated as
described in section 2.6.1.

5.3.1 Parametric methods

As in the univariate case plotting qq-plots can be helpful to see what distri-
bution ts the data well. As before I check how the residuals t to normal
distribution and students t-distribution.

(a) Normal distribution (b) Student's t-distribution

Figure 5.11: Estimating distribution for Portfolio 1

As for the univariate case, student's t-distribution (with 4 degrees of freedom)

ts the residuals better then the normal distribution, although as before I
will try both. Portfolios described in section 1.3 are analyzed and VaR esti-
mate is obtained both with multivariate EWMA and multivariate GARCH
models. For the multivariate GARCH models, CCC and DCC, the univari-
D
ate volatility estimates in matrix t (see equation 2.25) are all obtained by
GARCH(1,1) process, since it showed the best result for the univariate case.

Figure 5.12 shows a comparison between multivariate EWMA, CCC and

DCC model assuming normal distribution of the residuals and gure 5.13
shows a comparison between multivariate EWMA, CCC and DCC model
assuming student's t-distribution of the residuals. Here I present only g-
ures for Portfolio 1 (gures for Portfolio 2 is in appendix).
5.3 MULTIVARIATE CASE 45

Figure 5.12: Multivariate case, assuming normal distribution

Figure 5.13: Multivariate case, assuming Student's t-distribution

Then I examine if taking covariances into account really matters. This can
be done by estimating the VaR of each asset in the portfolio separately, then
summing them together (depended on their weight) and comparing to regu-
lar multivariate case.
46 ANALYSIS

Figure 5.14: Multivariate case, assuming normal distribution

Figure 5.15: Multivariate case, assuming Student's t-distribution

Finally exceptions are counted and presented in tables 5.14 and 5.15. EWMA†
and GARCH(1,1)† stands for EWMA and GARCH(1,1) without taking co-
variance into account.
5.3 MULTIVARIATE CASE 47

Method Portfolio 1 Portfolio 2

†
EWMA 1 (0.1) 0 (0.0)
†
GARCH(1,1) 2 (0.2) 4 (0.4)
EWMA 22 (2.2) 14 (1.4)
CCC 18 (1.8) 15 (1.5)
DCC 17 (1.7) 16 (1.6)

Table 5.16: Exceptions for the multivariate case, assuming normality

Method Portfolio 1 Portfolio 2

†
EWMA 0 (0.0) 0 (0.0)
†
GARCH(1,1) 1 (0.1) 0 (0.0)
EWMA 13 (1.3) 12 (1.2)
CCC 13 (1.3) 12 (1.2)
DCC 19 (1.9) 13 (1.3)

Table 5.17: Exceptions for the multivariate case, assuming t-distribution

As can be seen from the gures 5.14 and 5.15 and tables 5.16 and 5.17 not
taking covariances into account raises the VaR estimate a lot, as was ex-
pected (see section 2.5), and with hypothesis testing all of the cases when
correlation are not taken into account are rejected due to too few excep-
tions. Further analysis of means, standard deviations and maximum values
are presented in tables 5.18 and 5.19.

Portfolio 1 Portfolio 2
Method
†
Exceptions VaR σ VaR max(VaR) Exceptions VaR σ VaR max(VaR)

EWMA 1 0.0353 0.0087 0.0615 0 0.0341 0.0085 0.0632

†
GARCH(1,1) 2 0.0373 0.0060 0.0668 4 0.0355 0.0066 0.0673
EWMA 22 0.0223 0.0073 0.0492 14 0.0235 0.0070 0.0461
CCC 18 0.0240 0.0047 0.0538 15 0.0249 0.0067 0.0535
DCC 17 0.0223 0.0059 0.0464 16 0.0247 0.0067 0.0530

Table 5.18: Detailed information assuming normal distribution

Portfolio 1 Portfolio 2
Method
†
Exceptions VaR σ VaR max(VaR) Exceptions VaR σ VaR max(VaR)

EWMA 0 0.0403 0.0098 0.0702 0 0.0389 0.0096 0.0720

†
GARCH(1,1) 1 0.0424 0.0068 0.0761 0 0.0404 0.0075 0.0766
EWMA 13 0.0255 0.0082 0.0560 12 0.0267 0.0080 0.0526
CCC 13 0.0240 0.0047 0.0538 12 0.0263 0.0064 0.0658
DCC 19 0.0219 0.0044 0.0532 13 0.0258 0.0064 0.0657

Table 5.19: Detailed information assuming Student's t-distribution

Although the multivariate GARCH models are theoretically more `fancy'

then multivariate EWMA, they do not provide more green zones then multi-
variate EWMA. The multivariate EWMA model has 3 green zones out of 4,
while CCC has 2 out of 4 and nally DCC 1 out of 4. Detailed information
48 ANALYSIS

shows that multivariate EWMA has the lowest maximum VaR value in 3 out
of 4, but the highest standard deviation in all cases. Mean values are pretty
similar for all the multivariate methods.

Since multivariate EWMA has the most green zones and multivariate GARCH
doesn't show any superior skills multivariate EWMA is valued as the best
method for the multivariate parametric case. An really important factor is
also that the multivariate EWMA is much more simpler to calculate than
the very complex multivariate GARCH cases.

5.3.2 Non-parametric methods

For the Non-parametric case the methods described in chapter 4 were com-
pared. The portfolio return and the portfolio volatility was calculated with
equation 2.4 and 2.5 for the VWHS case, and then proceeded as in univariate
case. For the FHS a GARCH(1,1) volatility process was used to standard-
ize the portfolio returns, 1000 bootstraps were made and the mean of those
gave the nal VaR estimate. Here I only present the gure for portfolio 1
(portfolio 2 is in appendix).

Figure 5.16: Multivariate case, assuming Student's t-distribution

5.3 MULTIVARIATE CASE 49

Exceptions are presented in table 5.20

Method Portfolio 1 Portfolio 2
Basic HS 14 (1.4) 17 (1.7)
Age-WHS 25 (2.5) 24 (2.4)
EWMA VWHS 23 (2.3) 15 (1.5)
GARCH(1,1) VWHS 16 (1.6) 13 (1.3)
E-GARCH VWHS 13 (1.3) 12 (1.2)
FHS 15 (1.5) 14 (1.4)

Table 5.20: Exceptions for the Non-parametric multivariate case

The only method that has green zones for both portfolios is the E-GARCH.
Age-WHS gives red zones for both portfolios, which was also the case in
the univariate solutions, see 5.12. For further examination means, standard
deviations and maximum values are given in table 5.21.
Portfolio 1 Portfolio 2
Method Exceptions VaR σ VaR min(VaR) Exceptions VaR σ VaR min(VaR)
Basic HS 14 0.0264 0.0052 0.0427 17 0.0262 0.0034 0.0302
Age WHS 25 0.0226 0.0119 0.0693 24 0.0222 0.0098 0.0446
EWMA VWHS 23 0.0219 0.0085 0.0534 15 0.0215 0.0081 0.0479
GARCH(1,1) VWHS 16 0.0251 0.0086 0.0573 13 0.0247 0.0080 0.0561
E-GARCH VWHS 13 0.0253 0.0083 0.0507 12 0.0240 0.0077 0.0478
FHS 15 0.0250 0.0086 0.0563 14 0.0242 0.0074 0.0541

Table 5.21: Detailed information

By examining the two tables ( 5.20 and 5.21) it is clear that GARCH based
methods give fewer exceptions then the other, and of the GARCH based
methods, E-GARCH seems to be the best, since it has green zones for both
portfolios and lower minimum value than the other two. The basic HS has
pretty good results for the portfolios. It doesn't have that many exceptions
and has the lowest standard deviation and minimum value for both portfo-
lios. But as was shown in univariate case they can overestimate the VaR
value drastically. Therefore E-GARCH is deemed the best method for the
multivariate non-parametric case.
Chapter 6

Conclusion

As was said in section 1.1 the aim of the thesis was to evaluate dierent
methods for calculating VaR in order to obtain the optimal method for banks
trading book. Many calculations have been carried out with various results.
Of course it is dicult to point out one method that is `denitely' superior
than the others and results are dependent on assumptions and prerequisites.
Factors such as whether the data can be regarded as univariate or multi-
variate, how long the time period is, which distribution is assumed, what
complexity is allowed and how long does it take to perform calculation are
all things that can have great impact on the valuation.

If the conclusion could be separated into to two main elds, the univariate
and the multivariate case, which both have parametric and non-parametric
approaches the best models would be as in table 6.1

Univariate Multivariate
Parametric GARCH(1,1) (EWMA) EWMA
Non-Parametric GARCH(1,1) VWHS E-GARCH VWHS
Table 6.1: Separated conclusion

Of the parametric approaches the GARCH(1,1) and EWMA are found to

give the best result in the univariate case, but despite GARCH's `fancy' be-
havior, it's complexity aects the multivariate result and EWMA model is
regarded as the best. In all cases student's t-distribution lowered the num-
ber of exceptions, although it made minimum values larger in most cases.
Student's t-distribution is more sensitive to nding a local maxima and for
example failed for the MARL data. I would therefore recommend that nor-
mal distribution is always used as well as a alternative distribution to avoid
such incidents. In the non-parametric case volatility weighted historical sim-
ulations (VWHS) seem to be the best. They are relatively easy to implement
and can be used for both univariate and multivariate case without compli-
 51

cations. They manage to minimize ghost eects and react to volatility uc-
tuations. Filtered historical simulation also gave satisfactory results and are
quite interesting with their bootstrapping procedure, which could developed
further by for example enlarging the number of bootstraps, and/or obtaining
the nal estimate by some other way then taking the average. Age weighted
historical simulation didn't give good results in any case and the basic his-
torical simulation showed tendency to overestimate VaR drastically.

GARCH based models are more responsive and react to market conditions
better than other models. Their main failure is that they become intolerably
complex as the dimensionality grows, while EWMA is always quite simple to
implement. Bank's trading book is usually very large and holds all sorts of
instruments. It can therefore regarded as strictly multivariate and complex.
When evaluating VaR for bank's trading book, I would recommend using a
mixture of parametric and non-parametric models both to get a comparison
as well as to prevent failures (such as model and implementation failures).
My proposition for multivariate book would be parametric EWMA model
and a non-parametric GARCH based VWHS. But as said, bank's trading
book is usually very large and complex and therefore it is important to know
the underlying risk factors when choosing a model.

It is also worth mentioning that the last 2 years have been really unique in
nancial markets. Stock and indices have never been as high, and they have
been fallen really quickly past 2 years therefore the results of the this thesis
are colored by that. Not meaning that the results are any less important,
but just that it is important to keep in mind this extreme uctuations on
nancial markets over the past years when results are examined. Figure 6.1
shows the development of the OMX index from January 1993.

Figure 6.1: OMX index since January 1993

52 CONCLUSION

Finally I will list the main advantages and disadvantages with the paramet-
ric and the nonparametric approaches.

The main advantages with the parametric approach are;

• It's mathematically more advanced to use the parametric approach.

You will get a density and distribution function so there will be a
straight forward way of calculating VaR.

• They will give a better results than the non-parametric approach, if

assumptions about distributions are correct (or to be more specic
good) so the most important step is to make rational decisions about
the distribution.

while the main disadvantages are

• That you have to make an assumption about the distribution, and since
nancial data tend to be clustered, skewed and/or fat tailed picking a
distribution can be hard.

• Computationally more complex than non-parametric approach

• For non-linear assets they are an approximation and very complicated.

For the non-parametric methods, the main advantages could be summed up

as;

• They don't have any distribution and can therefore accommodate fea-
tures as fat tails, skew and any non-normal features, which are harder
to describe with parametric methods.

• They are relatively easy and simple

• Can be used for any type of risk, linear and non-linear (derivatives)

• They don't have any high dimensional problems

• Several implementations (or renements) have been developed for the

Basic HS, which can reduce `ghost eects'.

while their main disadvantages are;

• They will only simulated from the past history, meaning that results
will never be worse then what is in the sample (if the period was quiet
the VaR estimate will be low and vice versa).

• If the past (sample space) has some extreme losses, the VaR estimate
will be reected by that loss (unless some renement are used).
6.1 CRITICISM ON VAR 53

• Ghost eects are likely to corrupt the estimate.

• Can be computationally heavy, especially for more complicated assets,

as the assets has to be revalued.

6.1 Criticism on VaR

Although VaR is a very popular measure of risk in the nancial sector there
are many who have criticized that it isn't a very sophisticated risk mea-
sure. Artzner, Delbaen, Eber and Heath (1999) debated that VaR isn't
sub-additive, meaning that the VaR of a portfolio may be larger than the
sum of individual instruments in the portfolio. This is a fundamental rule,
since it means that diversication isn't guaranteed to reduce risk. Let's take
an example.

Imagine two identical bonds A and B both with the default probability
of 4%. Now if default occurs the loss will be the value of the bond, say
equal to 100. Therefore the VaR95% of each bond is 0 (higher than the
4% chance of default) and the VaR95% (A) = VaR95% (B) = 0. Now imag-
ine a portfolio of A and B . The probability of a loss equal to 200 is P(A
will default) × P(B will default) = 0.042 = 0.0016, and likewise the prob-
ability of no loss is 0.962 = 0.9216. Therefore the loss equal to 100 (one
of the bond will default) is P(A or B will default) = 1 - P(no default)
- P(both will default) = 1 − 0.9216 − 0.0016 = 0.0768. Therefore the
VaR95% (A+B) = 100 > VaR95% (A)+ VaR95% (B). Diversication has failed.

Generally speaking VaR doesn't say anything about the potential loss when
the loss occurs. VaR only gives the amount we are α percent sure of losing
not more than. Say that the VaR99% = 1 million euros. Now if the unlikely
occurs and we suer from a potential loss greater than VaR99% we have no
idea if it will be 1.1 million euros or 100 million euros. (Daníelsson, 2002) The
loss is only depending on the tail of the distribution (in left of the VaR value).

Other criticism proposed is for example by Taleb (1997) and Hoppe (1998)
who argued the statistical assumptions of VaR could lead to major errors,
Beder (1995) argued that dierent VaR models can give dierent VaR esti-
mates which makes the estimate impercise, furthermore Marshall and Siegel
(1997) argued that a similar techniques could give dierent estimates due
to implementations of the models. All this uncertainty concerning VaR can
lead to that experts and traders do not fully trust the VaR proposed and
take on larger risk than suggested by the VaR number, making the VaR
biased downwards (Ju and Pearson, 1999). Daníelsson (2002) argued that
regulatory constraints might discourage good risk management.
54 CONCLUSION

6.2 Further analysis

There are various ways to extend the research of this thesis. In the paramet-
ric approach dierent distribution could be tested, for example log-normal,
extreme value distribution or perhaps some skewed distribution, and in the
non-parametric approach tting a distribution to the histogram could be
quite interesting and would combine the features of both parametric and
non-parametric approach (called Kernel's). Comparing other types of stocks,
portfolios and currency pairs is straight forward and a whole new landscape
would be obtained by trying modeling non-linear instruments, such as op-
tions and futures, where Monte Carlo method's could come handy. Alter-
native volatility estimates such as stochastic and implied volatility are in-
teresting as well as Copula theory for estimating covariance. There would
also be interesting to research model/methods for interpolating VaR to other
time intervals, both longer, for example 10-day VaR which Basel Committee
demands, as well as shorter intra-day VaR (10 or 20 minute VaR) which are
common among traders.
Appendix A

Pictures

A.1 Univariate case

A.1.1 Parametric methods

First I show how GARCH(1,1) removes the autocorrelation in the 2nd mo-
ment by plotting autocorrelation plot of the squared residuals, 2t , and the
squared standardized residuals, zt2 .

Figure A.1: Autocorrelation for 2t and zt2 , MARL

Figure A.2: Autocorrelation for 2t and zt2 , ERIC

56 PICTURES

Figure A.3: Autocorrelation for 2t and zt2 , NDA

Figure A.4: Autocorrelation for 2t and zt2 , ISXI15

Figure A.5: Autocorrelation for 2t and zt2 , OMX

A.1 UNIVARIATE CASE 57

As can be seen in gures GARCH(1,1) succeeds well in removing heteroscedas-

ticity. Following are rest of the gures for the Parametric approach. Only
the long period is showed (the short period is the latter half of the long
period).

Figure A.6: LAIS and MARL

Figure A.7: ERIC and NDA

Figure A.8: ISXI15 and OMX

58 PICTURES

Then gures for the long period assuming student's t-distribution

Figure A.9: LAIS and MARL

Figure A.10: ERIC and NDA

Figure A.11: ISXI15 and OMX

A.1 UNIVARIATE CASE 59

A.1.2 Non-parametric methods

Following are the rest of the gures for the Parametric approach. Only the
long period is showed (the short period is the latter half of the long period).

Figure A.12: LAIS and MARL

Figure A.13: ERIC and NDA

60 PICTURES

Figure A.14: ISXI15 and OMX

A.2 Multivariate case

A.2.1 Parametric methods

Figures for Portfolio 2 are presented here below.

Figure A.15: Normal distribution and student's t-distribution

A.2 MULTIVARIATE CASE 61

Then pictures showing that not taking covariances into account overestimates
the value of VaR for portfolio 2.

Figure A.16: Normal distribution and student's t-distribution

A.2.2 Non-parametric methods

Finally the non-parametric result for portfolio 2

Figure A.17: Portfolio 2, non-parametric

Bibliography

Artzner, P., Delbaen, J., Eber, M. and Heath, D. (1999). Coherent measure
of risk. Mathematical Finance. 9. 203228.

Barone-Adesi, G., Bourgoin, F. and Giannopoulos, K. (1998). Don't look

back. Risk. 11. 100104.

Basel Committee on Banking Supervision. (2006). International convergence

of capital measurement and capital standards. Basel: Bank for Interna-
tional Settlements.

Beder, T. (1995). VaR: Seductive but dangerous. Financial Analysts Journal.

51. 1224.

Bollerslev, T. (1986). Generalized autoregressive conditional heteroskedas-

ticity. Journal of Econometrics. 31. 307327.

Bollerslev, T., Engle, R. and J. Wooldridge. (1988). A capital asset pricing

model with time varying covariances, Journal of Political Economy. 96.
116131.

Bollerslev, T. (1990) Modelling the coherence in short-run nominal exchange

rates: A multivariate generalized ARCH model, The Review of Economics
and Statistics, 72, 498505.
Boudoukh, J., Richardson, M. and Whitelaw, R. (1998). The best of both
worlds: A hybrid approach to calculate value at risk. Risk. 11. 6467.

Daníelsson, J. (2002). The emperor has no clothes: Limits to risk modelling.

Journal of Banking and Finance. 26. 12731296.
Daníelsson, J. and Zigrand, J. P. (2005). On time-scaling of risk and the
square-root-of-time-rule. EFA 2004 Maastricht Meetings. No. 5339.

Dowd, K. (1998). Beyond value at risk: The new science of risk management.
New Yotk: John Wiley and sons Ltd.

Dowd, K. (2005). Measuring market risk (2nd edition). New York: John
Wiley and sons Ltd.
Engle, R. (1982). Autoregressive conditional heteroskedasticity with esti-
mates of the variance of U.K. ination. Econometrica. 50. 9871008.
Engle, R. (2002). Dynamic conditional correlation - A simple class of mul-
tivariate generalized autoregressive conditional heteroskedasticity models.
Journal of Business and Economic Statistics. 20. 339350.
Glosten, L. R., Jagannathan, R. and Runkle, D. E. (1993). On the relation
between the expected value and volatility of the nominal excess return on
stocks. The Journal of Finance. 48. 17791801.
Hansen, P. R. and Lunde, A. (2005). A forecast comparison of volatility
models: Does anything beat a GARCH(1,1)? Journal of Applied Econo-
metrics. 20. 873889.
Hoppe, R. (1998). VaR and the unreal world. Risk. 11. 4550.
Hull, J. C. and White, A. (1998). Incorporating volatility updating into the
historical simulation method for value at risk. Journal of Risk. 1. 519.
Hull, J. C. (2007). Risk management and nancial institutions. New Jersey:
Pearson Prentice Hall.
Ju, X. and Pearsson, N. D. (1999). Using value-at-risk to control taking:
How wrong can you be?. Journal of Risk. 1. 536
Kaplan, S. and Garrick, B.J. (1981). On the quantitative denition of risk.
Risk Analysis, 1, 1127.
Kolluru, R. V. (1995). Risk assessment and management: Handbook for en-
vironmental. New York: McGraw-Hill Education
Kupiec, P. H. (1995). Techniques for verifying the accurancy of risk manage-
ment models. Journal of Derivatives. 3. 7384.
Linsmeier, T. J. and Pearson, N. D. (1996). Risk measurement: An introduc-
tion to value at risk. Mimeo, University of Illinois at Urbana-Champaign.
Marshall, C. and Siegel, M. (1997). Value at risk: Implementing a risk mea-
surement standard. Journal of Derivatives. 4. 91110.
Morgan, J. P. (1996). RiskMetrics technical document. New York, J.P. Mor-
gan Bank.
Nelson D. B. (1991). Conditional heteroskedasticity in asset returns: A new
approach. Econometrica. 59. 347370.
Sheppard, K. (2003). Multi-step estimation of multivariate GARCH models.
Proceedings of the international ICSC symposium: Advanced computing
in nancial markets.
Sigurdarson, H. A. (2007). Margvid GARCH líkön og verdlagning aeidna.
Unpublished M.Sc. thesis. Háskóli Íslands.

Taleb, N. (1997). Against VaR. Derivatives Strategy. 2. 2126.