The Journal of Risk Finance

Filtered extreme-value theory for value-at-risk estimation: evidence from Turkey

Alper Ozun Atilla Cifter Sait Y#lmazer

Article information:

Downloaded by UNIVERSITAT DE VALENCIA At 05:03 17 July 2015 (PT)

To cite this document:

Alper Ozun Atilla Cifter Sait Y#lmazer, (2010),"Filtered extreme-value theory for value-at-risk estimation:
evidence from Turkey", The Journal of Risk Finance, Vol. 11 Iss 2 pp. 164 - 179
Permanent link to this document:
http://dx.doi.org/10.1108/15265941011025189
Downloaded on: 17 July 2015, At: 05:03 (PT)
References: this document contains references to 46 other documents.
To copy this document: permissions@emeraldinsight.com
The fulltext of this document has been downloaded 726 times since 2010*

Users who downloaded this article also downloaded:

Stavros Degiannakis, Christos Floros, Alexandra Livada, (2012),"Evaluating value-at-risk models before
and after the financial crisis of 2008: International evidence", Managerial Finance, Vol. 38 Iss 4 pp. 436-452
http://dx.doi.org/10.1108/03074351211207563
Fotios C. Harmantzis, Linyan Miao, Yifan Chien, (2006),"Empirical study of value-at-risk and expected
shortfall models with heavy tails", The Journal of Risk Finance, Vol. 7 Iss 2 pp. 117-135 http://
dx.doi.org/10.1108/15265940610648571
Martin Odening, Jan Hinrichs, (2003),"Using extreme value theory to estimate value-at-risk", Agricultural
Finance Review, Vol. 63 Iss 1 pp. 55-73 http://dx.doi.org/10.1108/00215000380001141

Access to this document was granted through an Emerald subscription provided by emerald-srm:213039 []

For Authors
If you would like to write for this, or any other Emerald publication, then please use our Emerald for
Authors service information about how to choose which publication to write for and submission guidelines
are available for all. Please visit www.emeraldinsight.com/authors for more information.

About Emerald www.emeraldinsight.com

Emerald is a global publisher linking research and practice to the benefit of society. The company
manages a portfolio of more than 290 journals and over 2,350 books and book series volumes, as well as
providing an extensive range of online products and additional customer resources and services.
Emerald is both COUNTER 4 and TRANSFER compliant. The organization is a partner of the Committee
on Publication Ethics (COPE) and also works with Portico and the LOCKSS initiative for digital archive
preservation.
*Related content and download information correct at time of download.

The current issue and full text archive of this journal is available at
www.emeraldinsight.com/1526-5943.htm

JRF
11,2

Filtered extreme-value theory

for value-at-risk estimation:
evidence from Turkey

164

Alper Ozun
School of Management, Bradford University, Bradford, UK

Atilla Cifter
Sekerbank, Istanbul, Turkey, and
Downloaded by UNIVERSITAT DE VALENCIA At 05:03 17 July 2015 (PT)

Sait Ylmazer
Tekstilbank, Istanbul, Turkey
Abstract
Purpose The purpose of this paper is to use filtered extreme-value theory (EVT) model to forecast
one of the main emerging market stock returns and compare the predictive performance of this model
with other conditional volatility models.
Design/methodology/approach This paper employs eight filtered EVT models created with
conditional quantile to estimate value-at-risk (VaR) for the Istanbul Stock Exchange. The
performances of the filtered EVT models are compared to those of generalized autoregressive
conditional heteroskedasticity (GARCH), GARCH with student-t distribution, GARCH with skewed
student-t distribution, and FIGARCH by using alternative back-testing algorithms, namely, Kupiec
test, Christoffersen test, Lopez test, Diebold and Mariano test, root mean squared error (RMSE), and
h-step ahead forecasting RMSE.
Findings The results indicate that filtered EVT performs better in terms of capturing fat-tails in
stock returns than parametric VaR models. An increase in the conditional quantile decreases h-step
ahead number of exceptions and this shows that filtered EVT with higher conditional quantile such as
40 days should be used for forward looking forecasting.
Originality/value The research results show that emerging market stock return should be
forecasted with filtered EVT and conditional quantile days lag length should also be estimated based
on forecasting performance.
Keywords Stock returns, Emerging markets, Risk assessment, Stock exchanges, Turkey
Paper type Research paper

The Journal of Risk Finance

Vol. 11 No. 2, 2010
pp. 164-179
q Emerald Group Publishing Limited
1526-5943
DOI 10.1108/15265941011025189

Introduction
Estimating losses of financial instruments has become a crucial task for market risk
management in the current global economy. The importance of that task is more
critical in emerging financial markets where fluctuations in the volume of hot money
from international portfolio investments and hedge funds, unstable regulatory and
political environment, and the lack of informational efficiency create high volatility and
extremes in returns.
Complex and volatile market conditions in emerging markets require dynamic and
flexible econometric models that are able to capture extreme changes in financial
variables. In this research paper, we use filtered (conditional quantile) expected
shortfall as filtered extreme-value approach for value-at-risk (VaR) estimation to
capture the extremes in the returns. Extreme-value theory (EVT) do not follow the

Downloaded by UNIVERSITAT DE VALENCIA At 05:03 17 July 2015 (PT)

central limit theorem in mathematics arguing that if the sum of the variables has a
finite variance, then it follows Gaussian distribution. EVT focuses on the extremes
rather than mean.
There are some important reasons to choose EVT against parametric volatility
models. First, the distribution of returns is heavy-tailed and asymmetric in most
financial time series. EVT which approximates the area under the tail asymptotically
might be more powerful than imposing an explicit functional form. What is
more, extremes in returns might be caused by mechanisms that are structurally
different from the usual dynamics of financial markets. For example, extremes might
be the result of a major default or a speculative bubble. In those extreme conditions, the
distributional characteristics of financial time series might shift and require the
separation of tail estimation from the estimation of the rest of the distribution
(Neftci, 2000).
EVT has been used in financial risk estimation in recent years. The originality of
our paper is that we use conditional quantile expected shortfall with different lags and
compare them to find the optimal model capturing the extremes. As another original
work, we also apply h-step-ahead root mean square error (RMSE) and number of
exceptions to measure performance of the filtered expected shortfall. The performance
of the model is also empirically compared with those of the parametric models with
Kupiec (1995), Lopez (1999), Christoffersen (1998) and Diebold and Mariano (1995)
back-test algorithm.
We use the time series of daily returns of the Istanbul Stock Index-100 (Istanbul
Stock Exchange (ISE)-100) from January 2, 2002 to April 18, 2007 for our empirical
research. As an emerging market with dramatic macroeconomic and regulatory
changes in recent years, the ISE-100 gives us an opportunity to work with a highly
volatile and heavy-tailed data set. The back-test results show that the filtered expected
shortfall has superior performance in estimating the extremes and presents a new,
dynamic, and flexible perspective in VaR estimation.
Theoretical framework
Estimating VaR has become a crucial task of risk management functions of banks and
financial institutions since the Basel Committee stated that banks should be able to
cover losses on their trading portfolios over a ten-day horizon, 99 percent of the time.
However, classical VaR models focus on the whole empirical distribution of the returns
rather than that of extreme returns. On the other hand, managing extreme risk requires
estimation of quantiles that usually are not directly captured from the time series data.
EVT is presented as major alternative for generalized autoregressive conditional
heteroskedasticity (GARCH) models. GARCH approach dynamically improves itself in
recent years. Christoffersen and Jacobs (2004) empirically display that asymmetric
GARCH model that captures the leverage effect provides best estimation within all
GARCH models. Bekaert and Wu (2000) show that the leverage effect in equities
determines a strong negative correlation between returns and volatility, therefore,
creates a crucial source of skewness in returns. As a recent model, the normal mixture
GARCH (NM-GARCH) model is presented by Alexander and Lazar (2006). In the
NM-GARCH model, the individual variances are only tied with each other through
their dependence on the error term, thus the model captures switching shocks and
long-term memory in the stock returns (Ozun and Cifter, 2008).

Filtered extremevalue theory

165

JRF
11,2

Downloaded by UNIVERSITAT DE VALENCIA At 05:03 17 July 2015 (PT)

166

The distinguishing characteristic of EVT is to quantify the stochastic behavior of a

process at unusual levels. Especially, in bear markets, fat-tails are usually observed.
Poon et al. (2004) show that extreme-value dependence is usually stronger in bear
markets (left tails) than in bull markets (right tails). Longin (2000), McNeil and Frey
(2000) and Bali (2003) empirically show that traditional parametric VaR models with
normal density fail to estimate losses during financial crises.
In general, EVT has been seen as an alternative to GARCH models. EVT with
conditional quantile is constructed by McNeil and Frey (2000) under the assumption
that the tail of the conditional distribution of GARCH is approximated by a
heavy-tailed distribution. They underline the conditional quantile problem and apply
EVT to the conditional return distribution by using a two-stage method, which
combines GARCH model with EVT in applying the residuals from the GARCH
process.
In the literature, EVT is compared to GARCH-based parametric VaR estimation
models. Yamai and Yoshiba (2005) find out the empirical fact that VaR models do not
give the proper risk estimation in volatile market conditions while the EVT has a better
prediction performance. Acerbi (2002), Kuester et al. (2005), Inui and Kijima (2005) and
Martins and Yao (2006) also empirically show that EVT is superior in risk estimation
with financial time series. By using more than 30 years of daily return data on the
NASDAQ Composite Index, Kuester et al. (2005) compare the out-of-sample
performance of VaR models and EVT. They state that a hybrid method, combining
a heavy-tailed GARCH filter with an EVT-based approach, performs best overall.
Extremes in returns are observed in time series data from hedge funds and
emerging markets where high volatility and unstable money flows occur. In literature,
the empirical evidence on EVT is generally based on the data from hedge funds and
emerging financial markets. Amin and Kat (2003) empirically show that while hedge
funds combine well with stocks and bonds in the mean-variance framework, this is no
longer the case when skewness is considered. By using hedge funds data, Liang and
Park (2007) empirically show that EVT is able to foresight the fat-tails in returns
especially regarding high volatility in negative direction. Blum et al. (2002), Lhabitant
(2002) and Gupta and Liang (2005) also prove that the EVT performs better with
hedge-fund indices.
Empirical evidence from emerging markets is also in favor of EVT. Kalyvas et al.
(2007) present evidence from three former emerging and currently transition economies
along with two EU member countries of South and Eastern Europe using historical
simulation, conditional historical simulation, EVT, and conditional EVT. They show
that Hungary exhibits higher risk under extreme conditions indicating that its market
is much more vulnerable than all other markets under study.
Assaf (2006) use the EVT to examine four emerging financial markets belonging to
the Middle East and North Africa region, namely Egypt, Jordan, Morocco, and Turkey.
He focuses on the tails of the unconditional distribution of returns in each market and
provides their tail index behavior. The empirical evidence shows that these series have
significantly fatter tails than normal distribution and therefore suggests the use of
EVT. Tolikas and Brown (2006) use EVT to examine the asymptotic distribution of the
lower tail for daily returns in the Athens Stock Exchange over the period 1986-2001.
They show that the parameters of this distribution appear to vary with a tendency to
become less fat-tailed over time. A more comprehensive literature review on the EVT

Downloaded by UNIVERSITAT DE VALENCIA At 05:03 17 July 2015 (PT)

with methodological concerns can be followed in the works of Embrechts et al. (1997)
and Focardi and Fabozzi (2003, 2004).
With Turkish data, Cifter et al. (2007) use conditional quantile expected shortfall
and generalized Pareto distribution (GPD) for interest rates and find that conditional
EVT improves forecasting. Gencay et al. (2003), Gencay and Selcuk (2004), Altay and
Kucukozmen (2006) and Eksi et al. (2006) use EVT with unconditional quantile EVT to
estimate fat-tails in the stock returns in Turkey and they find that EVT performs better
than classical VaR models.
Filtered extreme-value theory
VaR reflects the change in a portfolio with a confidence level on a time period. In this
description, DPDt, measures changes in the market value of portfolio P on time Dt
period with probability p (Dowd, 2004):
PDPDt # VaR
p

In the equation, since F(DPD) is a distribution function of changes in the portfolio

value, it is possible to create an equation like (VaR) F 2 1(p). Obviously, F 2 1 is
the reverse of the distribution function. From that perspective, estimated VaR will
depend on the distribution of F function. Risk for one day should be equal to
VaR((1 2 p)%). When we include the time as a variable in the equation, for T time
period, risk is equal to TD-VaR((1 2 p)%).
Semi-parametric models, like EVT, are aimed at estimating the returns that are not
within the confidence level (p) but extremes and fat-tails.
EVT employs the GPD with thresholds. In the perspective of GPD, for pre-defined j
and s, the following equation holds:
F u y PX 2 u # yjX . u;
F u y < Gj;s y;

0 # y # xF 2 u

u!1

For negative returns, under the assumptions that x u y, the tail estimation can be
computed with equation (3):
!
^
^Fx 1 2 N u 1 j x 2 u21=j^
3
s
n
In this equation, n is the total number of observations in the data set, Nu is the
violations (extremes) above u. For pre-defined p . F(u) distribution, VaR for one day,
VaR( p%), is calculated with the equation (4):


^ !
s
n 2j
VaRGPD u
p 21
4
Nu
j^
In equation (4), by defining u with either constant or conditional quantile, GPD with
constant or conditional quantile is obtained. From similar perspective, Artzner et al.
(1999) evaluate expected shortfall as an alternative for VaR. In expected shortfall, the
expected value of the portfolio return is taken into consideration if there is a violation.

Filtered extremevalue theory

167

JRF
11,2

Expected shortfall can be constructed with the following equation (Gilli and
Kellezi, 2002):
ESp VaRp EX 2 VaRp jX . VaRp

168

The second nomination on the right side of the equation is the mean of the violation
distribution of FVaRp( y) on the VaRp threshold. For GPD, we can express the mean
violation function with j , 1 parameters as in the equation (6):
eu EX 2 ujX . u

Downloaded by UNIVERSITAT DE VALENCIA At 05:03 17 July 2015 (PT)

s ju
;
12j

s ju . 0

From that point of view, the expected shortfall is:

s^ j^VaRp 2 u
ESp VaRp
1 2 j^
ESp

VaRp s^ 2 j^u

1 2 j^ 1 2 j^

If X is GPD; then for all r , 1/j integer, (r), the first moment of each r exists.
In this research paper, eight different filtered expected shortfall estimation with two,
three, four, five, ten, 15, 20, and 40 days rolling quantile are estimated. The rolling
quantile days are randomly selected and maximum rolling is estimated as 40 days
since conditional EVT approximate to unconditional EVT more than 40 days rolling.
We use parametric models like GARCH, GARCH-t, GARCH-skewed student-t, and
FIGARCH for performance comparison of the filtered expected shortfall. The
methodologies for GARCH models are not examined here, but detailed examinations
can be found in Chung (1999), Baillie et al. (1996), Davidson (2002) and Laurent and
Peters (2002).
Methodologies of alternative back-tests
Alternative back-testing algorithms are employed to compare the performance of the
models. Kupiec (1995) test, Christoffersen (1998) test, Lopez (1999) test, RMSE (70
days), and h-step ahead forecasting RMSE (70 days), number of exceptions, h-step
ahead number of exceptions and Diebold and Mariano (1995) encompassing test are
used as the back-tests.
Kupiec test
Kupiec (1995) test defines the failure ratio ( f ) as the excess values from VaR(x) to the
total observations (T). When we nominate the pre-defined VaR with a, likelihood ratio
statistics with x 2 distribution for the Kupiec (1995) test can be given in equation (9):
LR 2{logj f x 1 2 f T2x j 2 logja x 1 2 aT2x j}

Christoffersen test
According to Christoffersen (1998) test, the probability of failure rate in the VaR
estimation is the important point for back-testing. To conduct the test, one should first
define p a Pr yt , VaRt a and test H0 : p a a against H1 : p a a.

The condition of {1 yt , VaRa} has a binomial likelihood and can be given in

equation (10):
L p a 1 2 p a n0 p a n1
10
PT
PT
where n0 tR 1 yt . VaRt a and n1 tR 1 yt , VaRt a (Saltoglu, 2003).
Under the null hypothesis, it becomes La 1 2 an0 a n1 . The likelihood ratio test
statistics can be given in equation (11):

Downloaded by UNIVERSITAT DE VALENCIA At 05:03 17 July 2015 (PT)

LR

22 InLa d
!x1
^
Lp

11

Lopez test
Lopez (1999) performs the back-test in three steps. In the first step, the proper
distribution of returns and statistics model is chosen. Second, employing the model
created in the first step and using historical losses/gains VaR and VaR(a) are
constructed and Li, a mean for losses/gains is obtained. In the last step, the process
described above is repeated many times, for example 10,000, in order to reach an
estimated value for a mean of loss distribution, Li 1i 10,000.
Lopez (1999) defines the violation function LVaRta ; xt;t1 that can be given in
equation (12):
1 xt;t1 2 VaRta 2
0

if

if

xt;t1 . VaRta

xt;t1 # VaRta

12

By using that methodology, back-test is conducted with equation (13):

T
1X
L^
LVaRa ; xt;t1
T t

13

H-step ahead RMSE and number of exceptions

RMSE is a scaled dependent comparison algorithm for forecasts. The smaller its
values, the more accurate the forecasts become. The test value is calculated as the
deviation of the h-step ahead forecasts of a variable, E( yt h), from its observed time
path, yt h. The RMSE of E( yt h) equals to the square root of the equation (14):
v
u T h
u1 X
14
RMSE t
s^f ;x 2 sr;t 2
h tT1
In the function, T 1 is the beginning of the testing sample, while T h is the end of
the testing sample.
The h-step ahead number of exceptions is calculated with the same methodology
but where RMSE is replaced with number of exceptions. The h-step ahead number of
exceptions is more sensitive measure of forecasting rather than h-step ahead RMSE as
considers tail loss directly.

Filtered extremevalue theory

169

JRF
11,2

170

Diebold and Mariano test of forecast accuracy

Diebold and Mariano (1995) developed the forecast comparison between a benchmark
and selected models based on forecast errors. The main advantage of this statistic is
that there is not any assumption on the distribution of forecast errors.
2
2
2
Define u^ 21;t1
P and u^ 2;t1 as two forecast errors and estimate dt1 u^ 1;t1 2 u^ 2;t1
21


and d P
t d t1 MSE1 2 MSE2 where MSE represents RMSEs of forecasting
models.
Diebold-Mariano test for equal MSE is defined as in equation (15):

Downloaded by UNIVERSITAT DE VALENCIA At 05:03 17 July 2015 (PT)

d

DM r
P

2
P 22 dt1 2 d

15

Data
ISE Rate (ISE-100 Index) is received from Bloomberg. Our dataset covers 1,325 daily
observations including 610 observations with negative returns from January 2, 2002 to
April 18, 2007. We constituted the series in log-differenced level. Figure 1 shows ISE
Index in log-differenced series where Figure 2 shows negative and positive returns
separately. By performing augmented Dickey-Fuller (Dickey and Fuller, 1981) and
0.15
Ise
0.1
0.05
0

54 107 160 213 266 319 372 425 478 531 584 637 690 743 796 849 902 955 1,0081,0611,1141,1671,2201,273

0.05
0.1

Figure 1.
ISE log-differenced series

0.15
0

0.14

0.02

0.12

Ise positive returns

0.04
0.1
0.06
0.08
0.08
0.06
0.1

Figure 2.
ISE negative and positive
returns

0.04

0.12
0.14
0.16

0.02
Ise negative returns
0

1 24 47 70 93 116 139 162 185 208 231 254 277 300 323 346 369 392 415 438 461 484 507 530 553 576 599 622 645 668 691 714

Downloaded by UNIVERSITAT DE VALENCIA At 05:03 17 July 2015 (PT)

Phillips-Peron test (Philips and Perron, 1988) we found that ISE Index is stationary at
log differenced level as shown in Table I.
Main statistical properties of the log-differenced index is shown in Table I.
Although x 2 normality test and Jargue-Bera statistics indicate that the index is
normally distributed kurtosis and skewness values show that distribution is skewed
and heavy-tailed. As a result, filtered EVT, like filtered expected shortfall, may capture
tail loss better compared to GARCH and alternative GARCH models.

Filtered extremevalue theory

171

Empirical evidence
Filtered expected shortfall with conditional quantile of two, three, four, five, ten, 15, 20,
and 40 days rolling and volatility models as GARCH(1,1), GARCH(1,1)-student-t,
GARCH(1,1)-skewed student-t, and FIGARCH(1,1) are estimated. Kupiec (1995), Lopez
(1999) and Christoffersen (1998) backtesting procedures and h-step ahead forecasting
of RMSE and number of exceptions are applied to compare predictive performance of
the models. Back-testing is done with 95 and 99 percent confidence intervals and Basel
guidelines require using a 99 percent confidence interval.
Table II reports GARCH(1,1), GARCH(1,1)-student-t, GARCH(1,1)-skewed student-t,
and FIGARCH(1,1) estimates for ISE index. All the parameters of GARCH models are
statistically significant and according to log-likelihood stat GARCH(1,1) fits better than
other GARCH models. Table III reports filtered expected shortfall shape (j) and scale
(b) parameters of lower and upper tail. In this paper, we only estimate lower tail of VaR
estimation based on GARCH and filtered expected shortfall models so that only lower
tail parameters are used to estimate filtered expected shortfall models[1].
Figures 3 and 4 show filtered expected shortfall and GARCH models graphs. Also
reported in Table IV Diebold-Mariano test (Diebold and Mariano, 1995) shows that
GARCH student-t and GARCH-skewed student-t are not statistically different from
GARCH with Gaussian distribution where FIGARCH, filtered expected shortfall with
two, three, four, and five days conditional quantile are statistically different at five
percent confidence interval and filtered expected shortfall with ten, 15, 20, and 40 days

ISE
Unit root tests
ADF test
P-P test
Main stats.
Asymptotic test: (x 2)
Normality test: (x 2)
Mean (m)
SD (s)
Skewness (S)
Kurtosis (K)
Minimum
Maximum
Jarque-Bera statistic
Notes: *Statistical significance at 5 percent level (at least); at-value

236.5121 *
236.5267 *
876.03 [0.0000]a
399.71 [0.0000]a
0.00091553 C
0.0212336
20.0707095
6.98092
20.133408 at obs. 289
0.11794 at obs. 215
876.031

Table I.
Unit root test and main
statistical properties

JRF
11,2

172

Downloaded by UNIVERSITAT DE VALENCIA At 05:03 17 July 2015 (PT)

Table II.
Estimation results from
volatility models

v
a
b1
n-student-t
j-Ske.
n-Skew
d
a b1
Loglike

GARCH

GARCH-t

GARCH-skew

FIGARCH

0.001 * * (2.738)
0.089 * * (5.325)
0.889 * * (44.21)

0.97840
3,321.90

0.001 * * (3.395)
0.084 * * (4.500)
0.886 * * (37.29)
8.125 * * (5.314)

0.97146
3,349.99

0.0014 * (2.912)
0.085 * * (4.490)
0.884 * * (36.15)

2 0.0606 (1.505)
8.215 * * (5.266)

0.96958
3,351.13

0.001 * * (2.973)
0.204 * (2.853)
0.600 * * (6.051)

0.5043 (6.022)
0.80434
3,325.44

Notes: *Significance at 5 percent confidence level; * *significance at 1 percent level

Lower tail
Shape (j)
Scale (b)

Models

Table III.
Extreme-value
parameters

Filtered
Filtered
Filtered
Filtered
Filtered
Filtered
Filtered
Filtered

ES
ES
ES
ES
ES
ES
ES
ES

2 days
3 days
4 days
5 days
10 days
15 days
20 days
40 days

0.064
0.081
0.114
0.034
0.170
0.131
0.139
0.184

0.011
0.011
0.009
0.011
0.010
0.011
0.010
0.009

Upper tail
Shape (j)

Scale (b)

0.072
0.115
0.134
0.187
0.300
0.238
0.226
0.348

0.011
0.010
0.010
0.008
0.007
0.009
0.009
0.008

0.0000000

0.0500000

0.1000000

0.1500000

Figure 3.
Filtered expected shortfall
models

0.2000000

Ise
FES 3 days
FES 5 days
FES 15 days
FES 40 days

FES 2 days
FES 4 days
FES 10 days
FES 20 days

0.2500000

conditional quantile are statistically different from GARCH model at 1 percent

confidence interval.
The predictive performance of filtered expected shortfall and GARCH models are
reported in Tables V and VI with 95 and 99 percent confidence intervals. According to
95 percent confidence interval, filtered expected shortfall with two days conditional
quantile is the best one based on Lopez test, filtered expected shortfall with 15 and

0
1

23 45 67 89 111 133 155 177 199 221 243 265 287 309 331 353 375 397 419 441 463 485 507 529 551 573 595

0.02

Filtered extremevalue theory

0.04
0.06

173

0.08
0.1
Downloaded by UNIVERSITAT DE VALENCIA At 05:03 17 July 2015 (PT)

Ise

0.12

Garch
Garch-t

0.14

Garch-skew

Figure 4.
GARCH models

Figarch

0.16

Models
GARCH
GARCH-t
GARCH-skew
FIGARCH
Filtered ES 2 days
Filtered ES 3 days
Filtered ES 4 days
Filtered ES 5 days
Filtered ES 10 days
Filtered ES 15 days
Filtered ES 20 days
Filtered ES 40 days

Ratio DM

0.5092
0.5194
2 0.09754 *
0.094686 *
0.148612 *
0.191513 *
0.185461 *
0.32498 *
0.047025 * *
0.077021 * *
0.085852 * *

Notes: Benchmark model is GARCH(1,1) with Gaussian distribution; *significance at 5 percent

confidence level; * *significance at 1 percent level

40 days conditional quantile performs best based on Christoffersen and Kupiec tests.
According to 99 percent confidence interval, filtered expected shortfall with ten and
20 days conditional quantile performs best based on Lopez test, filtered expected
shortfall with two days conditional quantile performs better based on Christoffersen
test, filtered expected shortfall with 15 and 40 days conditional quantile are the best
ones based on Kupiec tests. Since 99 percent confidence interval is more significant and
Basel guidelines require using 99 percent confidence interval, we also consider
99 percent confidence interval for back testing results. According to all back testing
procedures filtered expected shortfall models predictive performance is better than
GARCH models. There is not one filtered expected shortfall model that beats other
models based on Lopez, Christoffersen; and Kupiec tests therefore we applied h-step
ahead forecasting of RMSE and number of exceptions. Based on h-step ahead
forecasting of RMSE GARCH(1,1) is the best one[2]. Table VII shows that based on
h-step ahead forecasting of number of exceptions up to 70 days filtered expected

Table IV.
Comparing predictive
accuracy with the
Diebold-Mariano statistic

JRF
11,2

Downloaded by UNIVERSITAT DE VALENCIA At 05:03 17 July 2015 (PT)

174

Table V.
Back testing

Models
GARCH
GARCH-t
GARCH-skew
FIGARCH
Filtered ES 2 days
Filtered ES 3 days
Filtered ES 4 days
Filtered ES 5 days
Filtered ES 10 days
Filtered ES 15 days
Filtered ES 20 days
Filtered ES 40 days

Lopez

Christoffersen

Kupiec

0.68934
1.00409
0.78774
0.89262
0.15879
0.52972
0.85641
0.98265
1.41345
1.74404
1.41345
1.74404

0.00315
0.00025
0.00146
0.00063
0.17591
0.00958
0.00065
0.00021
0.00000
0.00000
0.00000
0.00000

0.00249
0.00023
0.00120
0.00054
0.11457
0.00766
0.00061
0.00022
0.00000
0.00000
0.00000
0.00000

Lopez

Christoffersen

Kupiec

0.32134
0.15609
0.25970
0.20462
0.66599
0.23692
0.08361
0.04985
0.00065
0.01124
0.00065
0.01124

0.00080
0.01469
0.00226
0.00595
0.00000
0.00396
0.06909
0.15074
0.85707
0.42083
0.85707
0.42083

0.99980
0.99601
0.99943
0.99846
0.99999
0.99907
0.98078
0.95528
0.68437
0.32276
0.68437
0.32276

Note: 95 percent confidence interval

Models

Table VI.
Back testing

GARCH
GARCH-t
GARCH-skew
FIGARCH
Filtered ES 2 days
Filtered ES 3 days
Filtered ES 4 days
Filtered ES 5 days
Filtered ES 10 days
Filtered ES 15 days
Filtered ES 20 days
Filtered ES 40 days
Note: 99 percent confidence interval

Models

Table VII.
H-step ahead forecasting

GARCH
GARCH-t
GARCH-skew
FIGARCH
Filtered ES 2 days
Filtered ES 3 days
Filtered ES 4 days
Filtered ES 5 days
Filtered ES 10 days
Filtered ES 15 days
Filtered ES 20 days
Filtered ES 40 days

No. of exceptions
t1
Avrg.
15
13
14
14
22
16
13
12
8
5
8
4

15.136
11.454
13.500
13.257
25.196
17.575
14.515
13.712
7.696
5.7272
6.500
3.0151

RMSE
t1

Avrg.

0.0373
0.0405
0.0383
0.0389
0.0449
0.0481
0.0501
0.0498
0.0586
0.0602
0.06051
0.0614

0.0402
0.0433
0.0411
0.0419
0.0479
0.0515
0.0537
0.0533
0.0617
0.0630
0.0633
0.0640

Note: Average no. of exceptions and RMSE is estimated with 70 days step-ahead forecasting

Filtered extremevalue theory

175

Conclusions
As a result of the dynamic and chaotic features of financial markets in emerging
economies, financial forecasting is almost impossible with parametric models.
Observed extremes and fat tails in returns need to be estimated with relatively more
sophisticated models. Parametric models have certain strict assumptions on the
distribution function of returns. These assumptions, either normality or asymmetric
distribution, are not able to make statistically significant estimations.
40
35

Garch
Figarch
FES4
FES15

Garch-t
FES2
FES5
FES20

Garch-skew
FES3
FES10
FES40

30
No. of exceptions

Downloaded by UNIVERSITAT DE VALENCIA At 05:03 17 July 2015 (PT)

shortfall with 40 days conditional quantile is the best one. We observed that an
increase in the conditional quantile decreases h-step ahead forecasting of number of
exceptions and this shows that filtered expected shortfall with 40 days conditional
quantile should be used for forward looking forecasting such as more than one month
forecasting. Christoffersen and Diebold (2000) show that volatility models such as
GARCH and other GARCH models can be used for forecasting up to 15-20 days ahead
for the USA financial instruments and Cifter (2004) shows that volatility models can be
used for forecasting up to 10-14 days ahead for Turkish interest rates.
Figure 5 shows h-step ahead forecasting of number of exceptions and Figures 6 and 7
show h-step ahead forecasting of RMSE up to 70 steps. The h-step ahead forecasting of
number of exceptions shows that filtered expected shortfall from 15 to 40 days
conditional quantile beats all GARCH models and filtered expected shortfall less than
15 days conditional quantile. Diebold-Mariano test of equal forecast accuracy (Diebold
and Mariano, 1995) is also applied to reveal statistical difference between filtered
expected shortfall models and found that filtered expected shortfall models with less
than ten days conditional quantile are not statistically different than two days
conditional quantile estimation (Table VIII), thus indicating that filtering less than ten
days conditional quantile may imitate GARCH models.

25
20
15
10
5
0

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
H-step ahead forecasting
Note: No. of exceptions

Figure 5.
H-step ahead forecasting

JRF
11,2

0.05
0.045
0.04

176

RMSE

0.035
0.03
0.025
0.02

Garch
Garch-t

0.01

Garch-skew

0.005

Figarch

0
1

Figure 6.
H-step ahead forecasting

7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 52 55 58 61 64 67 70
H-step ahead forecasting

Note: RMSE and GARCH models

0.07
0.06
0.05

RMSE

Downloaded by UNIVERSITAT DE VALENCIA At 05:03 17 July 2015 (PT)

0.015

0.04
0.03
0.02
0.01

FES2

FES3

FES4

FES5

FES10

FES15

FES20

FES40

0
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51 53 55 57 59 61 63 65 67 69

Figure 7.
H-step ahead forecasting

H-step ahead forecasting

Note: RMSE and filtered expected shortfall models

EVT, on the other hand, employs the central limit theorem for risk estimation.
According to the theorem, if the sum of the variables has a finite variance, then it
follows Gaussian distribution. The distribution of extremes in returns is limited to
having the same form without relying on the distribution of the parent variable.
In this research paper, we investigate filtered EVT with different rolling quantile to
estimate VaR. By using daily returns of the ISE National 100 Index, we estimate risk
with filtered EVT. For comparison of the model performance, we also estimate VaR with

Models
Filtered
Filtered
Filtered
Filtered
Filtered
Filtered
Filtered
Filtered

Ratio DM
ES
ES
ES
ES
ES
ES
ES
ES

2 days rolling
3 days rolling
4 days rolling
5 days rolling
10 days rolling
15 days rolling
20 days rolling
40 days rolling

1.2342
0.612275
0.627426
0.323022 *
0.264849 *
0.309795 *
0.297323 *

Downloaded by UNIVERSITAT DE VALENCIA At 05:03 17 July 2015 (PT)

Notes: Benchmark model is filtered expected shortfall with two days conditional quantile;
*significance at 5 percent confidence level

parametric models, namely, GARCH, GARCH with student-t distribution, GARCH

with skewed student-t distribution and FIGARCH. The success of the estimation of
the models are compared by using Kupiec (1995) test, Christoffersen (1998) test, Lopez
(1999) test, RMSE (70 days), and h-step ahead forecasting RMSE (70 days). The results
of the back-tests show that the filtered EVT has better risk forecasting performance
than parametric VaR models.
We conclude that financial forecasting especially in dynamic markets needs flexible
models. In accordance with this perspective, new semi-parametric models should be
conducted in future research without ignoring econometric methodological concerns.
Notes
1. VaR results of upper tail estimation can be obtained from the authors.
2. Based on standard RMSE(t 1) GARCH(1,1) also the best one.
References
Acerbi, C. (2002), Spectral measures of risk: a coherent representation of subjective risk
aversion, Journal of Banking & Finance, Vol. 26 No. 7, pp. 1505-18.
Alexander, C. and Lazar, E. (2006), Normal mixture GARCH(1,1): applications to exchange rate
modeling, Journal of Applied Econometrics, Vol. 21 No. 3, pp. 307-36.
Altay, S. and Kucukozmen, C. (2006), An assessment of value-at-risk (VaR) and expected tail
loss (ETL) under a stress testing framework for Turkish stock market, Proceedings of
ICBME, Yasar University, Izmir.
Amin, G. and Kat, H. (2003), Hedge fund performance 1990-2000: do the money machines really
add value?, Journal of Financial and Quantitative Analysis, Vol. 38, pp. 1-24.
Artzner, P., Delbaen, F., Eber, J.M. and Heath, D. (1999), Coherent measures of risk,
Mathematical Finance, Vol. 9, pp. 203-28.
Assaf, A. (2006), Dependence and mean reversion in stock prices: the case of the MENA region,
Research in International Business and Finance, Vol. 20 No. 3, pp. 286-304.
Baillie, R.T., Bollerslev, T. and Mikkelsen, H. (1996), Fractionally integrated generalized
autoregressive conditional heteroskedasticity, Journal of Econometrics, Vol. 74, pp. 3-30.
Bali, T. (2003), An extreme value approach to estimating volatility and value at risk, Journal of
Business, Vol. 76 No. 1, pp. 83-108.
Bekaert, G. and Wu, G. (2000), Asymmetric volatility and risk equity markets, The Review of
Financial Studies, Vol. 13 No. 1, pp. 1-42.

Filtered extremevalue theory

177
Table VIII.
Comparing predictive
accuracy with the
Diebold-Mariano statistic

JRF
11,2

Downloaded by UNIVERSITAT DE VALENCIA At 05:03 17 July 2015 (PT)

178

Blum, P., Dias, A. and Embrechts, P. (2002), The art of dependence modeling: the latest
advances in correlation analysis, in Lane, M. (Ed.), Alternative Risk Strategies, Risk
Books, London.
Christoffersen, P. (1998), Evaluating interval forecasts, International Economic Review, No. 39,
pp. 841-62.
Christoffersen, P. and Diebold, F.X. (2000), How relevant is volatility forecasting for financial
risk management?, Review of Economics and Statistics, Vol. 82, pp. 12-22.
Christoffersen, P. and Jacobs, K. (2004), Which GARCH model for option valuation?,
Management Science, Vol. 50, pp. 1204-21.
Chung, C.F. (1999), Estimating the fractionally integrated GARCH model, working paper,
National Taiwan University, Taipei.
Cifter, A. (2004), Asymmetric and fractionally integrated GARCH models with (skewed)
student-t and GED distribution in risk management, Proceedings of 8th National Finance
Symposium, Istanbul Technical University, Istanbul.
Cifter, A., Ozun, A. and Yilmazer, S. (2007), Value-at-risk with expected shortfall and
generalized Pareto distribution: some comparative evidence for Turkish interest rate
markets, Banks Association of Turkey-Bankers Journal, Vol. 60, pp. 3-16.
Davidson, J. (2002), Moment and memory properties of linear conditional heteroscedasticity
models, working paper, available at: www.cf.ac.uk/carbs/econ/davidsonje/ (accessed on
April 10, 2007).
Dickey, D.A. and Fuller, W.A. (1981), Likelihood ratio statistics for autoregressive time series
with a unit root, Econometrica, Vol. 49, pp. 1057-72.
Diebold, F. and Mariano, R.S. (1995), Comparing predictive accuracy, Journal of Business &
Economics Statistics, Vol. 13 No. 3, pp. 253-63.
Dowd, K. (2004), FOMC forecasts of macroeconomic risks, Occasional Papers, 12, Industrial
Economics Division.
Eksi, Z., Yildirim, I. and Yildirak, K. (2006), Alternative risk measures and extreme value theory
in finance: implementation on ISE 100 index, Proceedings of ICBME, Yasar University, Izmir.
Embrechts, P., Kluppelberg, C. and Mikosch, T. (1997), Modeling Extremal Events for Insurance
and Finance, Springer, Berlin.
Focardi, S. and Fabozzi, F. (2003), Fat tails, scaling and stable laws, Journal of Risk Finance,
Vol. 5 No. 1, pp. 25-36.
Focardi, S. and Fabozzi, F. (2004), The Mathematics of Financial Modeling and Investment
Management, Wiley, New York, NY.
Gencay, R. and Selcuk, F. (2004), Extreme value theory and value-at-risk: relative performance
in emerging markets, International Journal of Forecasting, Vol. 20, pp. 287-303.
Gencay, R., Selcuk, F. and Ulugulyagci, A. (2003), High volatility, thick tails and extreme value
theory in value-at-risk estimation, Insurance: Mathematics and Economics, Vol. 33,
pp. 337-56.
Gilli, K. and Kellezi, E. (2002), Portfolio optimization with VaR and expected shortfall,
in Kontoghiorghes, E.J., Rustem, B. and Siokos, S. (Eds), Computational Methods in
Decision-making, Economics and Finance, Kluwer Applied Optimization Series, Kluwer
Academic, New York, NY, pp. 167-83.
Gupta, A. and Liang, B. (2005), Do hedge funds have enough capital? A value-at-risk approach,
Journal of Financial Economics, Vol. 77, pp. 35-47.

Downloaded by UNIVERSITAT DE VALENCIA At 05:03 17 July 2015 (PT)

Inui, K. and Kijima, M. (2005), On the significance of expected shortfall as a coherent risk
measure, Journal of Banking & Finance, Vol. 29, pp. 853-64.
Kalyvas, L., Steftos, A., Sriopoulos, C. and Georgopoulos, A. (2007), An investigation of
riskiness in South and Eastern European markets, International Journal of Financial
Services Management, Vol. 2 Nos 1/2, pp. 13-21.
Kuester, K., Mittik, S. and Paolella, M.S. (2005), Value-at-risk prediction: a comparison of
alternative strategies, Journal of Financial Econometrics, Vol. 4 No. 1, pp. 53-89.
Kupiec, P.H. (1995), Techniques for verifying the accuracy of risk measurement models,
Journal of Derivatives, Vol. 3, pp. 73-84.
Laurent, S. and Peters, J.P. (2002), G@rch 2.2: an Ox package for estimating and forecasting
various ARCH models, Journal of Economic Survey, Vol. 16, pp. 447-85.
Lhabitant, F.S. (2002), Hedge Funds: Myths and Limits, Wiley, New York, NY.
Liang, B. and Park, H. (2007), Risk measures for hedge funds: a cross-sectional approach,
European Financial Management, Vol. 13 No. 2, pp. 333-70.
Longin, F.M. (2000), From value at risk to stress testing: the extreme value approach, Journal of
Banking & Finance, Vol. 24, pp. 1097-130.
Lopez, J.A. (1999), Methods for evaluating value-at-risk models, Federal Reserve Bank of
San Francisco Economic Review, Vol. 2, pp. 3-17.
McNeil, A.J. and Frey, R. (2000), Estimation of tail-related risk measures for heteroscedastic
financial time series: an extreme value approach, Journal of Empirical Finance, Vol. 7,
pp. 271-300.
Martinz, F.C. and Yao, F. (2006), Estimation of value-at-risk and expected shortfall based on
nonlinear models of return dynamics and extreme value theory, Studies in Nonlinear
Dynamics and Econometrics, Vol. 10 No. 2, pp. 107-49.
Neftci, S. (2000), Value at risk calculations, extreme events, and tail estimation, Journal of
Derivatives, Spring, pp. 12-19.
Ozun, A. and Cifter, A. (2008), A signal processing model for time series analysis: the effects of
international F/X markets on domestic currencies using wavelet networks, International
Review of Electrical Engineering, Vol. 3 No. 3, pp. 580-91.
Philips, P.C.B. and Perron, P. (1988), Testing for a unit root in time series regression,
Biometrica, Vol. 75, pp. 335-446.
Poon, S., Rockinger, M. and Tawn, J. (2004), Extreme value dependence in financial markets:
diagnostics, models and financial implications, The Review of Financial Studies, Vol. 17
No. 2, pp. 586-610.
Saltoglu, B. (2003), A High Frequency Analysis of Financial Risk and Crisis: An Empirical Study
on Turkish Financial Market, Yaylim Press, Istanbul.
Tolikas, K. and Brown, R. (2006), The distribution of the extreme daily share returns in the
Athens Stock Exchange, The European Journal of Finance, Vol. 12 No. 1, pp. 1-22.
Yamai, Y. and Yoshiba, T. (2005), Value-at-risk versus expected shortfall: a practical
perspective, Journal of Banking & Finance, Vol. 29 No. 4, pp. 997-1015.
Corresponding author
Alper Ozun can be contacted at: A.ozun@Brad.ac.uk
To purchase reprints of this article please e-mail: reprints@emeraldinsight.com
Or visit our web site for further details: www.emeraldinsight.com/reprints

Filtered extremevalue theory

179

This article has been cited by:

Downloaded by UNIVERSITAT DE VALENCIA At 05:03 17 July 2015 (PT)

1. Pilar Abad, Sonia Benito, Carmen Lpez. 2014. A comprehensive review of Value at Risk methodologies.
The Spanish Review of Financial Economics 12, 15-32. [CrossRef]
2. Stavros Degiannakis, Christos Floros, Alexandra Livada. 2012. Evaluating valueatrisk models before
and after the financial crisis of 2008. Managerial Finance 38:4, 436-452. [Abstract] [Full Text] [PDF]