Measuring Financial Risk using

Value at Risk with GARCH and Extreme Value

Theory in the CEE stock markets

Lucian Claudiu ANGHEL, PhD*

Cristian Ioan SOLOMON**

Abstract
Financial industry needs risk measures based on the
statistical distribution of asset returns in order to sustain the
quantitative approach of investment risks. But these risk
measures are static and further models that includes time-
dependent structures are needed, such as conditional volatility
and correlation.
In this paper we will look at Value at risk, the widely used risk
measure. Firstly, we will consider the classical methods such
as historical and Gaussian. Later, we will consider the three
approaches described by Yohan Chalabi and Diethelm Wurtz
on Stock Prices and Time Series on chapter 10 of book
Computational Actuarial Science with R.
Finally, the results of these approaches to Value at Risk will
be compared and analysed for various stock indexes from the
CEE region.
The main objective is to see which of these models predicts
Value-at-Risk most accurately on the past nine years for the
indexes from CEE markets which are less developed. In order
to compare the outcome of these models some developed
market were analysed as well.

*
Lecturer, Department of Management, National University of
Political Studies and Public Administration, Romania
**
Professional Risk Manager

1
 Keywords: asset returns, CEE region, investment risks, risk
measures, stock index, VaR.

1. Introduction
Financial industry needs risk measures based on the
statistical distribution of asset returns in order to sustain the
quantitative approach of investment risks. But these risk measures
are static and further models that includes time-dependent structures
are needed, such as conditional volatility and correlation.
Risk measures are primary focused on modelling of the lower
tail of time series of financial returns distributions. Value-at-risk
(VaR ), the value that one might loss with a given level of probability
, is the most frequent used measure.
In this paper we will look at Value at risk, the widely used risk
measure. Firstly, we will consider the classical methods such as
historical and Gaussian. Later, we will consider the three approaches
described by Yohan Chalabi and Diethelm Wurtz on Stock Prices
and Time Series on chapter 10 of book Computational Actuarial
Science with R, for VaR calculated with GARCH, Extreme Value
Theory (EVT) and GARCH-EVT. At last, we will test the VaR
calculated by fitting a skew t-Student distribution.

Objective
The main objective is to see which of these models predicts Value-at-
Risk most accurately on the past nine years for the indexes from CEE
markets- which are less developed- and for comparison some
developed market were considered as well. If there are models that
are fit for developed markets and not for smaller markets like CEE
region this should be highlighted. In the academic literature the vast
majority of the papers are focusing on developed capital markets. Still
there are particular aspects in the less developed markets that could
affect the quality of prediction using the same model well tested for
developed markets.

2
Methodology
The theories applied in this study will be done by the use R statistical
software, specifically the VaR functions from PerformanceAnalitics
package, the sstdFit function for fitting of t-Student distribution from
fGarch package. Functions for GARCH, EVT and GARCH-EVT were
used as described by Yohan Chalabi and Diethelm Wurtz on Stock
Prices and Time Series on chapter 10 of book Computational
Actuarial Science with R.
Regarding the mathematics and statistics, proofs for functions and
formulae presented will not be shown because this is far beyond the
scope of this article.

Value at Risk
According with the most common definition, value at Risk is an
estimate of the loss from a fixed set of assets (or portfolio) over a
fixed horizon that would be equal or exceed with a specific probability.
There are three parameters that are defining the Value-at-Risk, the
time horizon or holding period, the observation period and the
confidence level. These parameters should be stated when Value-at-
Risk is computed.

Observation Period
Depending on the model used the observation period is an important
parameter. For Gaussian and Historical simulation model the length
of the observation period it has high importance since the
observations have equal weights. For Historical simulation model the
longer the period the larger is the sample of observations. It is
recommended to an observation period of 250 or 500 trading days
meaning one year or two years. For Garch models the observation
period is limited by the model specification being resulted from model
selection process, most frequently being from one to three days.

Confidence Level
The confidence level expresses the probability that observed results
of portfolio or asset returns should exceed the Value-at-Risk number.
Confidence levels most often used are 95% or 99%. In case of 99%
confidence level it is expected the observed loss to exceed the
forecasted Value-at-Risk number once every hundred observations.

3
Review of the models used in this research

1. Historical Method
The historical method simply re-arrange actual historical returns,
ordering them from worst to best. The method it assumes the same
stable distributions of the returns for the future.

2. Gaussian Method
This method it assumes that returns are normally distributed, which is
often incorrect but it has the advantage that is relatively easy to
calculate.
Under the normality assumption VaR can be estimated as:

Where,
is given probability,
C is the coefficient for the confidence level,
is mean of the returns from sample, the observation period
is standard deviation of the returns from sample, the observation
period.

For example, VaR numbers can be easy calculated and compared at

different confidence levels if the mean and are already estimated:

Confidence level 99% 95%

Value-at-Risk -2,326* -1,645*

3. GARCH models for VaR

For this research it was considered the Generalized Autoregressive

Conditional Heteroskedastic (GARCH) model with Student-t
innovation distribution as introduced in 1996 by Bollerslev and
Ghysels. This is an important econometric model because it can
model the conditional volatility of financial returns. (Chalabi, 2014)

4
The variance equation of the GARCH(p,q) model can then be
expressed as follows:

Review of the t-GARCH(1,1) Model

Financial time series were modelled using GARCH(1,1) models

extended by the use of standardized Student-t distributions first time
by Bollerslev. According with his observations, compared with
conditionally normal errors, t-GARCH(1,1) errors captured better the
leptokurtosis seen in the data.

Previous studies have shown that the Student-t distribution performs

well in capturing the observed kurtosis in empirical log-return time
series. The density f*(z|) of the standardized Student-t distribution
can be expressed as follows:

(1)

where > 2 is the shape parameter and B(a,b) = (a) (b)/ (a+b)
the Beta function. If = 0 and If 2 = /( 2) then from formula (1)
results in the usual one-parameter expression for the Student-t
distribution as implemented in the R function dt().(Chalabi, 2014)

5
Forecasting with the heteroskedastic models

Models of heteroskedastic time series can provide forecasts that

includes both conditional mean and conditional variance,
independently of each other. For a GARCH(p,q) process, the n-step-
2
ahead forecast of the conditional variance +| is computed
recursively from

(2)

2 2 2 2 2
where +| = +| for i>0 while +| = + and +| =
2
+ for i <=0

4. VaR based on POT (EVT)

The forth model considered is the Peak-over-Threshold (POT)

method from Extreme Value Theory (EVT).

In the context of Value-at-Risk calculation, the extreme value theory

(EVT) becomes of very high interest because it involves modelling the
tail of the distributions.
Peak-over-threshold (POT) method is one of the approaches used in
the EVT. Peak-over-threshold (POT) results from EVT theorem that
states that when one selects a threshold high enough, the distribution
of the values exceeding the threshold converges in distribution to the
generalized Pareto distribution. (Chalabi, 2014)
The probability density function of the generalized Pareto distribution
can be written:

6
 (3)
where is a scale parameter,
and a shape parameter.
The support of the distribution is, when <= 0, x<= 0, and when < 0
it is 0 >= x >= - / . (Chalabi, 2014)

Further, VaR can then be expressed in terms of the distribution of

the exceeding points as

(4)

where and are the parameters of the GPD, u is the threshold, Nu

is the number of points exceeding the threshold and n is the number
of total values of the dataset.

Practically, VaR based on the POT method can be calculated in two

steps. In step one, the number of points that exceed the threshold u
are counted. The threshold can be kept at 5% or 1%.
In step two, parameters of the GPD are estimated by means of the
maximum likelihood estimation method. Finally, VaR is calculated
based on equation 4.

5. VaR based on the POT (EVT) and GARCH model

The fifth model considered is the combination of GARCH and POT

(EVT) methodologies to estimate tail risk measures, as presented in
2000 by McNeil and Frey. Considering a time series model where the

7
event at time t are composed of a mean term and some stochastic
process where zt ~ D(0,1) and with some time dependent volatility

(5)

where is modelled by a GARCH(1,1) process. The calculation of

VaR forecast can be written:

(6)

where, () is the VaR calculated by the Peak-over-Threshold

(POT) method from Extreme Value Theory (EVT), and Z are the
standardised residuals from fitting of the GARCH process.

6. VaR based on skew t-Student distribution.

The sixth model considered is assuming a stable distribution, the

skew Student-t Distribution in this case. The observation period for
fitting the distribution parameters, including the shape and the
skewness parameters was a window of two years, same as
considered for historical and Gaussian models.
The choice of a leptokurtic distribution was made considering that if it
describes the data better and displays significant excess kurtosis,
there is a good chance that VaR for high significance levels will be a
much better measure of the tail risk.
Further, VaR was calculated by taking the value on the respective
confidence interval from the probability distribution that was fitted.

Backtests
The VaR tests can be grouped on two categories, unconditional
coverage and independence. The unconditional coverage it counts
the frequencies of violations. A violation is when the actual return
exceeds the Value-at-Risk number for that date.
The test for independence makes assumption that observations are
independent of each other, consequently, when a violation happens

8
for two or more consecutive days there should be a question about
the model.

1. Kupiec Test
Kupiec test is an unconditional coverage test and it measures
whether the number of violations is consistent with the chosen
confidence level. The number of exceptions follows the binomial
distribution and it is a hypothesis test, where the nullhypothesis is:

Where,
p is the violation rate from the chosen Value-at-Risk confidence level,
is the observed violation rate
x represents the number of observed violations.
T is the number of observations.

It is conducted as a likelihood-ratio (LR) test and can be formulated

as:

Kupiec Testis asymptotically chi-square distributed with one degree of

freedom, for a confidence level of 95% the critical value is 3.84 and
for a confidence level of 99% is 6.63.

In case the LRUC statistic exceeds the critical value, the null
hypothesis is rejected and therefore the model is inaccurate (Nri,
2005).

2. Christoffersen Test
The Christoffersen test is an independence test described in
Christoffersen & Pelletier (2004). This is a likelihood-ratio test like the
Kupiec test. In order to test the independence it collects data for

9
violations and if they happen subsequently. The observations are
spitted in categories.

A violation occurs when the actual returns exceed the projected

Value-at-Risk number. Results are grouped in the following
categories: a violation followed by non violation (n10), a non violation
followed by a violation (n01), a non violation followed by a non
violation ((n00) and a violation followed by a violation (n00):

Further, values for and are calculated where they represent

the probabilities of a violation occurring conditional on if there was or,
respectively, was not a violation in the previous day. Then, is then
also calculated and represents the violation rate:

Further, the likelihood-ratio test is calculated with the null-hypothesis

that all violations are independent of each other.
The null-hypothesis is the following:

The likelihood statistics for independence is calculated as:

10
The likelihood statistics for independence is asymptotically chi-square
distributed with one degree of freedom, making the critical value at
95% confidence interval 3,84 and 6,83 at 99%. In case that the test
statistics is above this value the nullhypothesis is rejected and the
model is considered to have independence problems. A test statistic
under the critical rate assumes the model to be sound (Nieppola,
2009).

3. Test for Conditional Coverage

Christoffersen created a joint test consisting of the previously
described tests for unconditional coverage and for independence.
The test statistic is as follows:

Kupiec test for unconditional coverage and Christoffersen test for

independence are summed to get the test results for conditional
coverage. The test for conditional coverage is also
asymptotically chi-square distributed and has two degrees of
freedom. The critical value at 95% confidence level is 5,99 and 9,21
for 99%. In order the model pass the test if its test statistic has to be
under the critical value (Christoffersen & Pelletier, 2004).

Practical Research
This research have used the following index data were used for the
period between 2.01.2007 and 29.10.2015:

- BET index from Bucharest Stock Exchange;

- BUX index from Budapest Stock Exchange;
- SOFIX index from Bulgarian Stock Exchange;
- CRO index from Croatia Zagreb Stock Exchange;
- PX index Prague Stock Exchange;
- SPX index - Standard and Poor's 500 Index of companies
listed on the NYSE or NASDAQ;

11
 - UKX index - The Financial Times Stock Exchange 100
Index, also called the FTSE 100 Index, index of the 100
companies listed on the London Stock Exchange with the
highest market capitalization;
- SX5E index - The EURO STOXX 50 is a stock index of
Eurozone stocks designed by STOXX, an index provider
owned by Deutsche Brse Group;
- DAX index - The DAX (Deutscher Aktienindex (German
stock index)) is a blue chip stock market index consisting
of the 30 major German companies trading on the
Frankfurt Stock Exchange;

BET relative performance vs. other CEE indexes

Additionally, the period between 2.01.2009 and 29.10.2015 was

considered in order to avoid the influence of 2008 crisis specially of
historical VaR:

12
Corelation

Entire period correlation

Correlation CEE indexes vs. developed markets

13
3M Rolling Correlation CEE indexes

For the above mentioned indexes the following VaR were calculated:
- Historical VaR
- Gaussian VaR
- Garch VaR
- EVT VaR
- Garch with EVT
- T-Student VaR

14
Test results for Unconditional coverage and Conditional
coverage

Test results for period 2009-2015

BET
Unconditional coverage, H0: Conditional coverage: Correct
uc uc LR cc cc LR
Alph expected actual uc LR critica p cc LR critica p
a VaRtype exceed exceed stat l Val. uc Decision stat l Val. cc Decision
0.95 Hist_VaR 89 61 10.37 3.84 0.00 Reject H0 21.27 5.99 0.00 Reject H0
0.95 Gauss_VaR 89 48 23.71 3.84 0.00 Reject H0 25.47 5.99 0.00 Reject H0
0.95 Garch_VaR 89 75 2.44 3.84 0.12 Fail to Reject H0 8.30 5.99 0.02 Reject H0
0.95 EVT_VaR 89 58 12.89 3.84 0.00 Reject H0 16.76 5.99 0.00 Reject H0
0.95 Garch_EVR_VaR 89 80 0.99 3.84 0.32 Fail to Reject H0 5.50 5.99 0.06 Fail to Reject H0
0.95 Stud_VaR 89 54 16.76 3.84 0.00 Reject H0 19.39 5.99 0.00 Reject H0
0.99 Hist_VaR 17 12 2.16 6.63 0.14 Fail to Reject H0 2.32 9.21 0.31 Fail to Reject H0
0.99 Gauss_VaR 17 25 2.61 6.63 0.11 Fail to Reject H0 3.44 9.21 0.18 Fail to Reject H0
0.99 Garch_VaR 17 19 0.08 6.63 0.78 Fail to Reject H0 1.74 9.21 0.42 Fail to Reject H0
0.99 EVT_VaR 17 12 2.16 6.63 0.14 Fail to Reject H0 5.49 9.21 0.06 Fail to Reject H0
0.99 Garch_EVR_VaR 17 18 0.00 6.63 0.96 Fail to Reject H0 1.85 9.21 0.40 Fail to Reject H0
0.99 Stud_VaR 17 11 3.04 6.63 0.08 Fail to Reject H0 3.17 9.21 0.20 Fail to Reject H0

Test results for period 2011-2015

BET
Unconditional coverage, H0: Correct Conditional coverage: Correct

Alph expected actual uc LR uc uc LR cc LR cc cc LR

a VaRtype exceed exceed stat critical p Val. uc Decision stat critical p Val. cc Decision
0.95 Hist_VaR 62 42 8.27 3.84 0.00 Reject H0 14.46 5.99 0.00 Reject H0
0.95 Gauss_VaR 62 31 20.83 3.84 0.00 Reject H0 20.90 5.99 0.00 Reject H0
0.95 Garch_VaR 62 50 3.01 3.84 0.08 Fail to Reject H0 4.75 5.99 0.09 Fail to Reject H0
0.95 EVT_VaR 62 33 18.02 3.84 0.00 Reject H0 18.04 5.99 0.00 Reject H0
0.95 Garch_EVT_VaR 62 55 1.10 3.84 0.29 Fail to Reject H0 2.08 5.99 0.35 Fail to Reject H0
0.95 Stud_VaR 62 36 14.27 3.84 0.00 Reject H0 14.27 5.99 0.00 Reject H0
0.99 Hist_VaR 12 9 1.15 6.63 0.28 Fail to Reject H0 1.28 9.21 0.53 Fail to Reject H0
0.99 Gauss_VaR 12 17 1.41 6.63 0.24 Fail to Reject H0 2.88 9.21 0.24 Fail to Reject H0
0.99 Garch_VaR 12 13 0.01 6.63 0.91 Fail to Reject H0 2.42 9.21 0.30 Fail to Reject H0
0.99 EVT_VaR 12 7 2.99 6.63 0.08 Fail to Reject H0 7.84 9.21 0.02 Fail to Reject H0
0.99 Garch_EVT_VaR 12 12 0.03 6.63 0.87 Fail to Reject H0 2.73 9.21 0.26 Fail to Reject H0
0.99 Stud_VaR 12 8 1.94 6.63 0.16 Fail to Reject H0 2.04 9.21 0.36 Fail to Reject H0
The full results of the tests are presented in appendix 1

15
Evolution of Value-at-Risk indicators (entire period):

Evolution of Value-at-Risk indicators (2015):

16
Number of exceed cases (daily loss larger than VaR) for BET
Historical VaR

Gaussian VaR

Garch VaR

EVT VaR

17
Garch with EVT VaR

t-Student VaR

Summary of results:

Period 2009-2015 at 95%

Unconditional
Decision
Reject Grand
Fail to Reject H0 H0 Total
EVT_VaR 5 4 9
Garch_EVT_VaR 9 9
Garch_VaR 6 3 9
Gauss_VaR 3 6 9
Hist_VaR 3 6 9
Stud_VaR 4 5 9
Grand Total 30 24 54

18
For Gaussian, Historical and t-student VaRs the null hypothesis
of correct exceedances is rejected in majority of the cases.
For Garch VaR and EVT VaR exceedances are correct in
majority of the cases. And Garch with EVT model has all correct
exceedances in all cases.

Conditional
Decision
Reject Grand
VaR Type Fail to Reject H0 H0 Total
EVT_VaR 9 9
Garch_EVT_VaR 8 1 9
Garch_VaR 5 4 9
Gauss_VaR 9 9
Hist_VaR 9 9
Stud_VaR 9 9
Grand Total 13 41 54

Only two models, can satisfy the condition for correct

exceedances and independence, Garch VaR and Garch with
EVT VaR.

Period 2009-2015 at 99%

Unconditional
Decision
Reject Grand
VaR Type Fail to Reject H0 H0 Total
EVT_VaR 7 2 9
Garch_EVT_VaR 9 9
Garch_VaR 9 9
Gauss_VaR 7 2 9
Hist_VaR 9 9
Stud_VaR 7 2 9

19
Grand Total 48 6 54
All the models performed relatively better for correct
exceedances at 99% comparative with 95% on the
unconditional decision.

Conditional
Decision
Reject Grand
VaR Type Fail to Reject H0 H0 Total
EVT_VaR 5 4 9
Garch_EVT_VaR 8 1 9
Garch_VaR 9 9
Gauss_VaR 5 4 9
Hist_VaR 6 3 9
Stud_VaR 8 1 9
Grand Total 41 13 54

Only the Garch model satisfy the condition for correct

exceedances & independence in all the cases. Garch with EVT
VaR and t-Student VaR models performs relatively well with just
one rejection of the null hypothesis out of nine cases.
Gaussian VaR and EVT VaR have the worst results.

Comparison CEE markets vs. developed markets (at 99%)

CEE
Conditional
Decision
Reject Grand
Row Labels Fail to Reject H0 H0 Total
EVT_VaR 4 1 5
Garch_EVT_VaR 5 5
Garch_VaR 5 5

20
Gauss_VaR 5 5
Hist_VaR 5 5
Stud_VaR 4 1 5
Grand Total 28 2 30
Developed markets
Conditional
Decision
Reject Grand
Row Labels Fail to Reject H0 H0 Total
EVT_VaR 1 3 4
Garch_EVT_VaR 3 1 4
Garch_VaR 4 4
Gauss_VaR 4 4
Hist_VaR 1 3 4
Stud_VaR 4 4
Grand Total 13 11 24

It can be observed that in case of CEE markets the VaR models are
performing well satisfying correct exceedances & independence, with
just two exceptions on EVT and t-Student VaR.
In the developed markets, the situation is completely different, only
Garch models and t-Student VaR models are performing well. The
Gaussian model fails the condition of conditional coverage in all the
cases,

Period 2011-2015 at 99% (eliminating the effects of 2007-2009

volatility)

cc Decision
Reject
Var Type Fail to Reject H0 H0
EVT_VaR 6 3
Garch_EVT_VaR 9
Garch_VaR 9

21
 Gauss_VaR 5 4
Hist_VaR 7 2
Stud_VaR 9
Grand Total 45 9
Garch VaR, and t-Student models perform well in all the cases

CEE markets
Count of cc
Decision Column Labels
Row Labels Fail to Reject H0
EVT_VaR 5
Garch_EVT_VaR 5
Garch_VaR 5
Gauss_VaR 5
Hist_VaR 5
Stud_VaR 5
Grand Total 30
For CEE indexes considered, all models satisfying correct
exceedances & independence

Developed markets
Count of cc
Decision Column Labels
Reject
Row Labels Fail to Reject H0 H0
EVT_VaR 1 3
Garch_EVT_VaR 4
Garch_VaR 4
Gauss_VaR 4
Hist_VaR 2 2
Stud_VaR 4
Grand Total 15 9

22
For developed markets, the Gaussian VaR model it fails in all cases
and Historical VaR has mixed results.

Conclusions

In the following we will look at conditional coverage tests, correct

exceedances and independence. The following table presents the
number of cases where the tests rejects or fail to rejects the null
hypothesis conditional coverage tests, correct exceedances and
independence, at 99% and 95% confidence levels.

Conditional coverage tests summary:

Alpha 0.99
All markets analysed Developed Markets CEE Markets
2007 - 2015 2009 - 2015 2007 - 2015 2009 - 2015 2007 - 2015 2009 - 2015
Fail to Reject Fail to Reject Fail to Reject Fail to Reject Fail to Reject Fail to Reject
Reject H0 Reject H0 Reject H0 Reject H0 Reject H0 Reject H0
H0 H0 H0 H0 H0 H0
Row Labels
EVT_VaR 5 4 6 3 1 3 1 3 4 1 5
Garch_EVT_VaR 8 1 9 3 1 4 5 5
Garch_VaR 9 9 4 4 5 5
Gauss_VaR 5 4 5 4 4 4 5 5
Hist_VaR 6 3 7 2 1 3 2 2 5 5
Stud_VaR 8 1 9 4 4 4 1 5
Grand Total 41 13 45 9 13 11 15 9 28 2 30 0

Alpha 0.95
All markets analysed Developed Markets CEE Markets
2007 - 2015 2009 - 2015 2007 - 2015 2009 - 2015 2007 - 2015 2009 - 2015
Fail to Reject Fail to Reject Fail to Reject Fail to Reject Fail to Reject Fail to Reject
Reject H0 Reject H0 Reject H0 Reject H0 Reject H0 Reject H0
Row Labels H0 H0 H0 H0 H0 H0
EVT_VaR 9 1 8 4 4 5 1 4
Garch_EVT_VaR 8 1 9 4 4 4 1 5
Garch_VaR 5 4 6 3 1 3 1 3 4 1 5
Gauss_VaR 9 9 4 4 5 5
Hist_VaR 9 1 8 4 1 3 5 5
Stud_VaR 9 9 4 4 5 5
Grand Total 13 41 17 37 5 19 6 18 8 22 11 19

Over all, GARCH and GARCH with EVT VaR models performs well
for the indexes analysed not depending on the period chosen for
consideration or confidence level. However, if the longer period,
2007-2015, is consider on the GARCH with EVT VaR model there is
one case rejecting the null hypothesis out of 9 indexes considered.

23
Therefore, only the GARCH VaR model pass the test for conditional
coverage, correct exceedances and independence, for all nine
indexes considered and for the longer period, 2007 2015. The
disadvantage of using these models is the fact that the forecasts
provided by these models are very volatile since are directly
influenced by last observations.

Referring to the confidence intervals considered, 99% and 95%, for

this analysis it should be noted that at 99% confidence level all the
models are performing far better that at 95%. Actually, at 95%
confidence level the Gaussian, Historical, t-Student and EVT VaR
models reject the null hypothesis in almost all the cases. The only
model that performs well at 95% confidence level is the GARCH with
EVT VaR model.

Historical and Gaussian models tend to be affected in period after

crises and are overestimating the VaR. In the periods before crises
these models are underestimating the VaR. The analysis confirms
that in developed markets these models have poor results for both
confidence intervals considered, and with a little improvement for
stable periods. In small markets, as CEE markets are, at 99%
confidence level these models are performing well regardless the
observation period chosen. However, this should be regarded
prudently since this market are characterised by lower liquidity.

There is not a single answer of what method for Value at risk

calculation to be considered. Two or three methods can be used for
calculation in parallel. Periodically testing and back-testing the Value
at risk results are recommended.

24
References:

1. Bollerslev T., 1987 A conditionally heteroskedastic time

series model for speculative prices and rates of return, The
review of economics and statistics
2. Bollerslev T., Ghysels E, (1996), Periodic autoregressive
conditional heteroscedasticity
Journal of Business & Economic Statistics
3. Chalabi Y., Wurtz D., (2014), Computational Actuarial
Science with R: Stock Prices and Time Series - Reprint
Chapter 10
4. Christoffersen, P., & Pelletier, D. (2004). Backtesting Value-
at-Risk: A Duration-Based
Approach. Journal of Financial Econometrics
5. Ghalanos A.,( 2014), rugarch package for R Univariate
GARCH models
6. Nri B, (2005), Backtesting Value-at-Risk Methodologies
7. Ragnarsson F. J., (2011), Comparison of Value-at-Risk
Estimates from GARCH Models
8. Wuertz D.,(2008), fGarch package for R for econometric
functions for modelling GARCH processes

25