Volume 5, Issue 2, February – 2020 International Journal of Innovative Science and Research Technology

ISSN No:-2456-2165

Risk Modelling using EVT, GJR GARCH,

t Copula for Selected NIFTY Sectoral Indices
N. Sai Pranav#1, R. Prabhakara Rao#2
Post Graduate- Research, Economics Department , #2 Dean – Humanities.
#1

SSSIHL, Deemed to be University Vidyagiri, Prasanthi Nilayam, Andhra Pradesh, Pin 515134

Abstract:- This paper makes an attempt to explain the by traditional VaR models. They may often be fairly
procedure as well as estimate the VaR of a selected different, as demonstrated by numerous studies. And most
portfolio of the Nifty Sectoral Indices using approaches of the studies focus on the univariate case of marginal VaR,
such as GJR GARCH-EVT-Copula, Filtered Historical component VaR and incremental VaR making it
Simulation, Generalised Extreme Value Theory and t undesirable for portfolio risk management.
Copula. The GJR GARCH-EVT-t Copula model
extracts the filtered residuals obtained using the GJR An attempt to empirically test and evaluate the VaR
GARCH technique and by using the Gaussian Kernel estimates of the portfolio consisting of NIFTY Sectoral
method for interior of the distribution and Extreme Indices using GJR GARCH, Copula Theory, Extreme
Value Theory for upper and lower tails to estimate the Value Theory and FHS technique is made in this study. The
cumulative distribution of the residuals. A comparison EVT is integrated with a time series model in order to
is made between the estimated VaR simulation by the obtain a conditional EVT which can filter the
Monte-Carlo method, the aforementioned method and heteroscedasticity and the autocorrelations in the financial
by using t Copula to get the joint distribution of each data. The multivariate joint probability density function is
sectorial indices. The normalised maxima of the used to fit the stock market portfolios but it underestimates
sequence is measured by the GEV distribution. An the VaR of the portfolios. The copula method helps in
alternative to the Monte Carlo simulation and the fitting the multivariate dependence model and is simple and
Historical simulation is the FHS technique. The mean flexible.
equation is modelled using the ARMA model while the
volatility is modelled using GARCH with a non- The section 2, 3, 4 and 5 deals with the Literature
parametric specification of the probability distribution Review, Methodology, Results & Discussions and
of asset returns. The VaR estimates of the equally Summary & conclusions respectively.
weighted portfolio of NIFTY Sectoral indices of 95%
and 99% confidence intervals are backtested over a II. LITERATURE REVIEW
2478-day estimation window.
The Heteroskedastic multivariate financial models
Keywords:- Value at Risk, NIFTY Sectorial Indices. have been introduced by Nelson (1981), Kraft and Engel
(1982), Bollerslev et al. (1988), Diebold and Nerlove
I. INTRODUCTION (1989), among others. Different multivariate financial
models impose different restrictions on the dynamic
Traditionally capital markets are considered as behaviour of the variances, co variances and correlations.
barometer of an economy of a country and plays a crucial Since the financial time series are leptokurtic with heavy
role in generating capital required for the economy. One of tails which make VaR being underestimated for i.i.d
the inherent characteristics of these capital markets is that Gaussian distribution. So we tend to adopt the EVT and
they are highly volatile in nature. With the implementation Copula in order to understand and model the tails and
of globalization and liberalization policies by the encapsulate the heavy-tail into the VaR estimation.
developing countries the unrestricted flow of capital among
the markets of the economies has resulted in the financial Lauridsen (2000) in his paper showed several defects
integration with world markets. Especially the developing of the VaR models in modelling the distribution of tails of
markets due to their potential for better returns started profits and losses and extreme value models based on
attracting large capital inflows. As a result, the volatility of GARCH can be improvised by integrating changes in the
these capital markets also became a major concern for the volatility level. Burridge, John, Michael, & Chih (2000)
investors. As the changes in the stock prices are very proposed to estimate the market risk based on Extreme
sensitive to the events happen at economy level there is a Value Theory which attempted to model the rare market
need for the economies to maintain the stable economic events. Mendes & Carvalhal (2003) proposed that the
conditions so that the volatility of stock prices is always Extreme Value Theory to analyse ten Asian stock market
kept under control. for estimating the VaR is a more conservative way to
decide the capital requirements than traditional VaR
Although the VaR has become a very popular models. Selcuk, Gencay & Fatuk (2004) demonstrated that
assessment of risk, it is not a problem-free solution also. the Generalized Pareto Distribution (GDP) and Extreme
First it is not always possible compare the VaR measured Value Theory aptly fits the tails of the return distribution in

IJISRT20FEB443 www.ijisrt.com 1040

Volume 5, Issue 2, February – 2020 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
these markets and are more accurate at higher quantiles. sequence. This process is also known as Generalized
Palaro & Hotta (2006) demonstrated that the conditional Extreme value (GEV) distribution given by:
Copula theory can be an intense tool in estimating the VaR
for a portfolio of Nasdaq and S&P 500 stock indices. Gilli 1
−
𝑥−𝑏 𝑧
& Këllezi (2006) proved that the POT method exp [− (1 + 𝑧 ) ]:𝑧 ≠ 0
demonstrated more prevalent in the long term behaviour 𝑎
𝐻(𝑧; 𝑎, 𝑏) =
and it was favourable to compute in the interval estimates 𝑥−𝑏
as it better endeavours the information about the
{ exp [−𝑒𝑥𝑝 ( 𝑎 )] : 𝑧 = 0
distribution of the model. Bohdalova (2007) presented few
strategies that uses Copula approach for making prudent
When under the condition of x appears to 1+ z(x-b)/a
choices as for the data employed and computational aspects
> 0 and the parameters a and b in the GEV distribution
are concerned can be made to decrease the overall cost and
refers to the location and scale parameters of the limiting
computational time. Marimoutou, Raggad, & Trabelsi
distribution in H, their meaning is close but they are distinct
(2009) results showed that the Conditional Extreme Value
from mean and standard deviation. The final parameter i.e.
Theory and Filtered Historical Simulation procedures are
z is critical and it corresponds to the tail index as it shows
indeed offering a major improvement as suggested for oil
the heaviness of extreme losses in the data sample.
markets. Staudt, FCAS & MAAA (2010) highlighted a few
of the considerations in modelling joint behaviour with
Let’s define Extreme value at risk directly relating to
Copulas such as choosing a Copula which appropriately
the fitted GEV distribution Hn (to n data points) given by: -
catches the tail dependence and representing the skewness
Pr[𝐻𝑛 < 𝐻′ ] = 𝑝 = Pr⁡[𝑋 < 𝐻′ ]𝑛 = [𝛼]𝑛
and kurtosis of the fundamental data and have natural
interpretation. Huang, Chein & Wang (2011), Gondje-
Where α is the VaR confidence level associated with
Dacka & Yang (2014) and Zhang, Zhou, Ming, Yang &
the threshold H’. This can be defined as:
Zhou (2015) has proved that it has the ability to understand
𝑏 [1 − (−𝑛⁡ln⁡(∝))−𝑛𝑧𝑛 ] : (𝐹𝑟𝑒𝑐ℎ𝑒𝑡; 𝑧 > 0)
and process the complex structure among the financial 𝐸𝑉𝑎𝑅 = { 𝑛
market events and even calculated the maximum loss and 𝑏𝑛 − 𝑎𝑛⁡𝑙𝑛[−𝑛𝑙𝑛(∝)] : (𝐺𝑢𝑚𝑏𝑒𝑙; 𝑧 = 0)
maximum gain of the distribution. Yi, Y., Feng, X., &
Huang (2014) and Xiao & Koeniker (2009) proposed a  Copula Function
method to estimate extreme conditional quantiles by Copula are primarily used to minimise tail risk. The
combining quantile GARCH model of an Extreme Value price dependencies of multivariate distribution which can
theory approach. Zhang, H., Guo, J., & Zhou (2015) be split into into k univariate, marginal distribution and a
observed that the prediction effect of VaR is significantly copula theory can be formed. Let 𝑋1 , 𝑋2 , 𝑋3 , … … . , 𝑋𝑑 as the
more beneficial in a relatively stable market and VaR will random variables. Then, their cumulative distribution
underestimate market risk if there are large fluctuations in function is denoted by 𝐻(𝑥1 , 𝑥2 , 𝑥3 , … … , 𝑥𝑑 ) =
market and suggested to utilize stress testing. Singh, Allen 𝑃[𝑋1 < 𝑥1 , 𝑋2 < 𝑥2 , … … , 𝑋𝑑 < 𝑥𝑑 ] and the marginal as per
& Powell (2017) applied GARCH (1,1) based by dynamic the Sklar’s theorem can be seen to be 𝐹𝑖 (𝑥) = 𝑃[𝑋𝑖 ≤ 𝑥] .
EVT approach and appeared with backtesting in stable as Copula can be defined as a multivariate distribution
well as in extreme market conditions for the ASX-AII consisting of random variables in which each of its
ordinaries(Australian) index and the S&P-500 (USA) marginal distributions is uniform. It elucidates the
Index. dependence amongst two or more variables which possess
the characteristic of non-normal distribution.
III. METHODOLOGY
IV. RESULTS AND DISCUSSIONS
 Extreme value Theory (EVT)
Let us assume X represents the random variable of A. GJR-GARCH-EVT-Copula Model
loss and is independently identically distributed given by: It is necessary that the data needs to independent and
identically distributed (i.i.d) before even we use EVT to
𝐹(𝑥) = Pr⁡(𝑋 ≤ 𝑥) model the tails of the distribution (i.e. of an individual
index). The two important which every financial return
EVT heavily uses the Fisher- Trippett theorem and exhibit is autocorrelation and heteroskedascity. Now we
thus giving us practical solutions to n extreme random loss may look at different figures which depict the relation
variables to measure the normalised maxima of the between ACF of returns as well as ACF of squared returns
for a particular nation.

IJISRT20FEB443 www.ijisrt.com 1041

Volume 5, Issue 2, February – 2020 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
 ACF plots of Top 5 NIFTY 50 Sectoral Indices

Fig 1A:- Sample ACF of returns and sample ACF of squared returns of NIFTY Bank

Fig 1B:- Sample ACF of returns and sample ACF of squared returns of NIFTY FMCG

Fig 1C:- Sample ACF of returns and sample ACF of squared returns of NIFTY Private Bank

Fig 1D:- Sample ACF of returns and sample ACF of squared returns of NIFTY IT

IJISRT20FEB443 www.ijisrt.com 1042

Volume 5, Issue 2, February – 2020 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165

Fig 1E:- Sample ACF of returns and sample ACF of squared returns of NIFTY Financial Services
Fig 1:- ACF plots of Top 5 NIFTY 50 Sectoral Indices

We require the GARCH model essentially to

condition the data for the tail estimation process. The
reason being that the squared returns illustrates a high
degree of persistence w.r.t to variance. GARCH will also
be crucially helpful in filtering out the serial dependence
which is exhibited by the data. One which is quite
noticeable is the fact that the returns are not independent
from one day to the next. But AR (1)-GJR GARCH (1, 1)
model helps in producing i.i.d observation which sorts out
the requirement necessary for EVT. Now we try to fit AR After we fit AR (1)-GJR GARCH (1, 1) models to
(1)-GJR GARCH (1, 1) models to each index: - each index, we can then compare model residuals as well as
the equivalent conditional standard deviation which is
separated out from the raw returns.

IJISRT20FEB443 www.ijisrt.com 1043

Volume 5, Issue 2, February – 2020 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165

Fig 2:- Filtered residuals and filtered conditional standard deviations for NIFTY Sectorial Indices

When we closely perceive the lower graphs we observe a persistent variation in volatility present in the filtered residuals.
Later on we can standardize the residuals. These standardized residuals follow zero – mean and unit- variance (i.i.d series).
Therefore, it shows the EVT estimation of the sample CDF tail.

 The ACF of the standardized residuals and squared standardized residuals.

Fig 3A:- Sample ACF of the standardized returns and squared standardized residuals of NIFTY Bank

Fig 3B:- Sample ACF of the standardized returns and squared standardized residuals of NIFTY FMCG

IJISRT20FEB443 www.ijisrt.com 1044

Volume 5, Issue 2, February – 2020 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165

Fig 3C:- Sample ACF of the standardized returns and squared standardized residuals of NIFTY Private Bank

Fig 3D:- Sample ACF of the standardized returns and squared standardized residuals of NIFTY IT

Fig 3E:- Sample ACF of the standardized returns and squared standardized residuals of NIFTY Financial Services
Fig 3:- Sample ACF of the standardized returns & squared standardized residuals of NIFTY Sectorial Indices

The next task involves fitting a probability poorly when it is applied it to upper and lower tails. In the
distribution for each index so that their daily movements practice of risk management, we notice that it is of utmost
can be traced (this can be done after we filter out the data). importance that we accurately portray the tails of the
While doing we are not concerned whether the data that is distribution, even when the observed data in the tails is
being analyzed is from normal distribution or any other scarce. This gap is bridged with the help of GPD
form of simple parametric distribution. (generalized Pareto distribution).

Interior of the distribution is where we find the After approximating three distinct regions of
majority of the data, so we can use kernel density estimate composite semi-parametric empirical CDF, we graphically
for it. One of the biggest drawback of it is that it executes join them and we will be able to see the results.

IJISRT20FEB443 www.ijisrt.com 1045

Volume 5, Issue 2, February – 2020 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165

Fig 4:- Empirical CDF of a Top 5 NIFTY Sectoral Indices

IJISRT20FEB443 www.ijisrt.com 1046

Volume 5, Issue 2, February – 2020 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
We then suture together three distinct regions of the composite semi parametric empirical CDF which were estimated before
and the results are displayed above. By observation we get to know that both lower and upper tail regions are appropriate for
extrapolation. While the kernel smooth interior denoted in black can be used for interpolation.

We already know that the older graph illustrated CDF so it is indispensable to check whether the GPD would fit in detail.
The parameterized Cumulative density function of GPD is given as:

Let us plot the empirical CDF of upper tail in excess of the residuals. This has to be supplemented with the CDF being fitted
with the GPD.

Fig 5:- Upper Tail of Standardized Residuals of NIFTY Sectoral Indices

IJISRT20FEB443 www.ijisrt.com 1047

Volume 5, Issue 2, February – 2020 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
From the last figure we can perceive that the each index to uniform scale. This form is crucial for fitting
empirically generated CDF curve for a specific nation a Copula.
matches well with the fitted GPD results. Thus far we have
only used 10% of the standardized residuals and we notice B. Filtered Historical Simulation
that the fitted distribution quiet closely resembles the FHS combines a relatively sophisticated model-based
exceedances data. Hence we can conclude that the GPD treatment of volatility (GARCH) with a nonparametric
model is a good choice. Consequently, for each of the five specification of the probability distribution of asset returns.
NIFTY Sectoral Indices we have five separate univariate FHS retains the non-parametric nature of historical
models which describes the distribution of daily gains and simulation by bootstrapping (sampling with replacement)
losses. But the problem arises in tying these models from standardised residuals.
together and this is done by Copula model. As per the
definition of the Copula we know that it is a multivariate This method requires the observations to be i.i.d. But
probability distribution whose individual variables are as we have already seen that the vast majority of the
uniformly distributed. Thus we take these resultant financial return series display various degrees of
univariate distributions to transform the individual data of autocorrelation and heteroskedascity.

Fig 6:- Sample ACF and Sample ACF of Squared of the Portfolio Returns

The sample ACF of the portfolio returns exhibit a

mild serial correlation. However, when the Sample ACF is
squared it illuminates the degree of persistence in variance.
Thus makes it necessary for the GARCH model to
condition the data used in the bootstrapping method.

Thus for generating a series of i.i.d observations, we

can fit AR (1) +EGARCH (1, 1) model given below:

𝑟𝑡 = 𝑐 + 𝜃𝑟𝑡−1 + 𝜖𝑡 , 𝜖𝑡⁡𝑁(0, 𝜎𝑡 ) Table 1:- Result of ARMA (1,0,0) Model

And a symmetric EGARCH for conditional variance

looks like

Thus this shows that where AR model could only

compensate for auto correlation, EGARCH model is able to
compensate for heteroskedascity.

Table 2:- Result of EGARCH (1,1) Conditional Variance

Model:

We see that the estimation depicts that there are six

estimated parameters accompanied by their corresponding
standard errors. (I.e. AR conditional mean model has two
parameters while EGARCH conditional variance model has
four parameters.

IJISRT20FEB443 www.ijisrt.com 1048

Volume 5, Issue 2, February – 2020 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
Thus the fitted model can be written as: Now next major step involves in modelling the
𝑟𝑡 = −0.00065 ∗ 10−7 + ⁡0.0645⁡𝑟𝑡−1 + 𝜖𝑡, ⁡𝜖𝑡 = 𝑁(0, 𝜎𝑡 ) residuals and the resultant standard deviation which are
log⁡[𝜎 2 𝑡 ] = ⁡ −0.0936049 + 0.99log⁡[𝜎 2 𝑡 ] + 0.13⁡(|𝑧𝑡−1 | filtered out from the raw returns. Below graph depicts the
− 𝐸[|𝑧𝑡−1 |) − 0.08𝑧𝑡−1 variation in heteroskedascity present in the filtered residual.
The i.i.d property is of significant for it allows
The t-statistic of AR (1) in ARCH (1, 0,0) model is bootstrapping that uses sampling procedures to safely avoid
greater than two which means that the parameter should be the downsides of choosing the sample from a population.
statistically significant, while for GARCH and ARCH in The reason being that the successive observation is
EGARCH (1,1) model). critically dependent upon each other.

Now let us take a look at the ACF of the standardized

residual as well as the squared standardized residual.

Fig 7:- Sample ACF of Standardized Residuals and Sample ACF of Squared Standardized Residuals

As we try to match both the ACF of the standardized residuals as well as the raw returns, it is revealed that the standardized
residuals are exhibiting properties of being approximately i.i.d. therefore it is more amenable for subsequent bootstrapping.
We then sample for 10000 times on the filtered standard residual based on the bootstrapping method. This may be taken to input
of i.i.d noise process of the holding period.

The below figure shows the cumulative distribution function and probability density function of simulation of one-month
return.

Fig 8:- Simulated One-month of Top 5 NIFTY 50 Sectoral Indices Portfolio Returns CDF

IJISRT20FEB443 www.ijisrt.com 1049

Volume 5, Issue 2, February – 2020 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
C. Copula Simulation
Computing value at risk is shown in the following example. These are used for portfolio using multivariable Copula
simulation with fat tailed marginal distribution. To calculate optimal risk-return portfolios, these simulations are used.

 Returns & Marginal Distributions:

The distributions of returns of each index are characterized individually to make Copula modelling. Although each return
series distribution can be featured parametrically, it is needed to fit a semi-parametric model by utilizing a piecewise distribution
with generalized pareto tails. To improve the behaviour in each tail, extreme value theory is used.

Fig 9:- Pairwise Correlation of Historical Returns

For top 5 NIFTY 50 Sectoral Index return series the code segment makes an object of type pareto tails. To create a
composite semi-parametric CDF for every index, pareto tail objects encloses the estimates of parametric pareto lower and upper
tail and the nonparametric kernel – smoothed interior.

The outcome which is a piecewise distribution object permits interpolation index in interior of CDF whereas extrapolation
(function evaluation) in each tail. To estimate quantities out of historical record, extrapolation can be used though it has not valued
for operations of risk management. The fit coming from pareto tail distribution is compared with normal distribution here.

IJISRT20FEB443 www.ijisrt.com 1050

Volume 5, Issue 2, February – 2020 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165

Fig 10:- Semi-Parametric Piecewise Probability Plot for NIFTY Sectoral Indices

IJISRT20FEB443 www.ijisrt.com 1051

Volume 5, Issue 2, February – 2020 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
 Copula Calibration
The Statistics toolbox function can be used to calibrate and simulate a t-Copula to data. Daily index returns are used to
estimate the parameters of Gaussian and t-Copula are used to estimate the function Copula fit. When the scalar degrees of freedom
become infinitely large then the t-Copula becomes a Gaussian Copula. These two Copula shares linear correlation as basic
parameter and become same family.

Fig 11:- Transformed returns prior to fitting a Copula

The Calibration of a linear correlation matrix of a

Gaussian Copula is straightforward whereas it is not the
same case for t-Copula. So in order to calibrate a t-Copula,
Statistics tool box software give two techniques. The
following code segment first transform the daily centred
returns into uniform variates by using the piecewise, semi-
parametric Cumulative Distribution Functions derived from
above and then Gaussian and t-Copulas are fitted into the
transformed data.

The expected correlation matrix is related but not

identical to the linear correlation matrix

The estimated correlation matrix is quite similar to

linear correlation matrix though they are not identical.

 Copula Simulation
As the parameter of the Copula have been estimated.
The combined dependent uniform variates have to be
simulated by utilizing the function Copularnd. The uniform
variates from Copularnd has to be transformed into daily
central returns by extrapolating pareto tails and
interpolating smoothed interior. The historical data set is
t-Copula parameters have to be by the parameters tallied with simulated centred returns and the returns
obtained from t-Copula calibration which are of lower obtained are consistent. The returns obtained are assumed
degree of freedoms have to be noted and a significant to independent of time but at any instant may possess
exodus from the Gaussian situation has to be indicated. dependence and rank correlation induced by Copula.

IJISRT20FEB443 www.ijisrt.com 1052

Volume 5, Issue 2, February – 2020 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165

Fig 13:- Probability Density Function of Generalised

Extreme Value Distribution

Fig 12:- Pairwise correlation of simulated returns The best estimation of the 1 month EVAR is given as: -
𝑎
𝐸𝑉𝑎𝑅 = 𝑏25 − −𝑧25 [1 – (-n ln(0.95))nz25] = -0.0729
25
D. Generalised extreme value theory and extreme VaR 𝑎
GEV distribution alone can be used to measure the 𝐸𝑉𝑎𝑅 = 𝑏25 − −𝑧25 [1 – (-n ln (0.99)) nz25] = -0.1241
25
normalised maxima of the sequence. This distribution is
also called Fisher Tippet Distribution because it measures The EVaR value is indicative of the fact that among
the chance of deviation of an event from the probability the 5 NIFTY Sectoral Indices in our portfolio we are
distribution’s central tendency i.e. median. The family of definitely expecting an extreme loss of 12.41% & 7.29% on
GEV have converged to three types of Extreme value the following month (taken from the last 2478 trading
distributions i.e. Gamma, Gumbel and the Frechet days). The cumulative distribution function for NIFTY 50
distributions. Sectoral Indices are given as:

So let’s look at the sample of N=5 largest losses on

per NIFTY sectoral indices over last 2478 days, we can
effortlessly fit it with GEV distribution and get the best
estimates for the parameter z, an and b parameters. But if
you see is a very small parameter. Instead of that why can’t
extract 5 worst losses that took place in the last 2478 days.
thus it increases our n substantially to n=25.

As mentioned earlier the MATLAB’s cell array is

holding 2478 return series (each 2478 day long). We
Increase the sample size to n=30 points by taking the top 5
maximal daily losses for each stock. Now we fit the GEV
distribution
1 Fig 14:- Cumulative Distribution Function of Generalised
𝑥 − 𝑏 −𝑧 Extreme Value Distribution
𝐻(𝑧; 𝑎, 𝑏) = exp [− (1 + 𝑧 ) ]
𝑎
 Computing VaR using different models
While engaging the ready to use function gevfit from In this section, we go through the methods of
Matlab statistics toolbox we get, calculation and the approach adopted to establish our
findings. Initially, we embark upon the task of transforming
the individual standardized residuals pertaining to the AR
(1)-GJR GARCH (1, 1) models into uniform variates. We
attain this by utilizing the semi-parametric empirical CDF
after which we fit the t Copula to the transformed data. It is
important to note that the estimated optimal degrees of
freedom for the t Copula is 8.4108. In this research, we
also adopt the vital techniques of Filtered Historical
Simulation, t Copula and Generalized Extreme Value
The best estimates of the model’s parameters are Z25 , method for comparison. Using the same, we simulate 1,
a25, b25. The negative value of z actually comes from the 00,000 independent random trails of dependent
Fretchet distribution since we fitted data with negative standardized index residuals spanning a month-long
signs. horizon of 22 trading days. Lastly, we form a 1/5 equally
weighted index portfolio composed of the individual
indices assuming that we are given the simulated returns of

IJISRT20FEB443 www.ijisrt.com 1053

Volume 5, Issue 2, February – 2020 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
each index. Next, we work on calculating the VaR at 99% estimated figures of 95% and 99% VaRs for t(8.4108) and
and 95% confidence levels, again spanning the month-long other models.
risk horizon. In the table constructed below, we list out the

Value at Risk for different models

Models CEVT+t(8.4108) copula FHS t Copula GEV appraoch
Max Loss 27.33% 30.86% 29.86% 35.81%
max Gain 17.69% 15.18% 16.20% 21.69%
95% VaR -6.73% -5.55% -4.32% -7.29%
99% VaR -11.73% -9.94% -4.22% -12.41%
Table 3:- Value-at-Risk Calculations for the different models

We can see that the table 3, that the CEVT-Copula [4]. Huang, S. C., Chein, Y. H., & Wang, R. C. (2011).
based approach given the estimated optimal degree of Applying Garch-Evt-Copula Models for Portifolio
freedom as 8.4108 performs best to be only followed by t Value-at-Risk on G7 Currency Markets. International
Copula. Finally, it is to be noted that The Generalized Research Journal of Finance and Economics(74).
Extreme Value approach and Filtered Historical Simulation [5]. Lauridsen, S. (2000). Estimation of Value at Risk by
overestimate the portfolio VaR. Extreme Value Methods. Extremes, 3(2), 107-144.
[6]. Mendes, B. V., & Carvalhal, A. (2003). Value-at-Risk
V. CONCLUSIONS and Extreme Returns in Asian Stock Markets.
International Journal of Business,, 8(1).
The paper is an attempt to find an appropriate VaR [7]. Marimoutou, V., Raggad, B., & Trabelsi, A. (2009).
model among the set of models namely GJR GARCH, Extreme Value Theory and Value at Risk :
EVT-Copula, t copula, Filtered Historical Simulation and Application to Oil Market. Energy Economics, 31(4),
General Extreme Value Distribution to estimate the VaR of 519-530.
returns of the NIFTY Sectoral Indices. [8]. Palaro, H. P., & Hotta, L. K. (2006). Using
Conditional Copula to Estimate Value at Risk.
First, the market risk of the NIFTY Sectoral Indices Journal of Data Science, 93-115.
portfolio is modelled by the Monte Carlo simulation using [9]. Palaro, H. P., & Hotta, L. K. (2006). Using
the t copula and EVT. Second, the distribution of the Conditional Copula to Estimate Value at Risk.
residuals is modelled using the POT based EVT. Lastly, the Journal of Data Science, 93-115.
data and the simulated residuals are checked for their strong [10]. Selcuk, Gencay, R., & Fatuk. (2004). Extreme value
or weak correlation by fitting a seven-dimensional t copula. theory and Value-at-Risk: Relative performance in
emerging markets. International Journal of
Hence there is a perennial conflict as the method Forecasting, 20(2), 287-303.
chosen by a financial institution decides the risk capital it [11]. Singh, A., Allen, D., & Powell, R. (2017). Value at
holds. It is a non-trivial issue because choosing a method Risk Estimation Using Extreme Value Theory.
for portfolio VaR problems inaccurately measures market [12]. Staudt, A., FCAS, & MAAA. (2010). Tail Risk,
risk can have adverse impact on the functioning of the Systemic Risk and Copulas. Semantic Scholar.
financial institutions. So the results of this study can be [13]. Xiao, Z., & Koenker, R. (2009). Conditional Quantile
used to perform a good risk management on Global Estimation for Garch models. Journal of the American
investors. Statistical Association.
[14]. Yi, Y., Feng, X., & Huang, Z. (2014). Estimation of
REFERENCES Extreme Value-at-Risk: An EVT Approach for
Quantile GARCH Model. Elsevier, 124(3), 378-381.
[1]. Bohdalova, M. (2007). A comparison of Value at Risk [15]. Zhang, H., Guo, J., & Zhou, L. (2015). Study on
methods for measurement. Financial Market Risk Measurement Based on GJR
[2]. Burridge, P., John, C., Michael, T., & Chih, H. L. GARCH and FHS. Science Journal of Applied
(2000). Value-at-risk: Applying the extreme value Mathematics and Statistics, 70-74.
approach to Asian markets in the recent financial [16]. Zhang, H., Zhou, L., Ming, S., Yang, Y., & Zhou, M.
turmoil. Pacific-Basin Finance Journal, 8(2), 249- (2015). Empirical Research on VaR Model based on
275. GJR GARCH, Evt and Copula. Science Journal Of
[3]. Gondje-Dacka, I.-M., & Yang, Z. (2014). Modelling Applied Mathematics and Statistics, 3(3), 136-143.
risk of foreign exchange portifolio based on Garch-
Evt-Copula and filtered historical simulation
approaches . TheEmpirical Econometrics and
Quantitative Economics Letters, 3(2), 33-46.

IJISRT20FEB443 www.ijisrt.com 1054