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.
htm JRF 11. Lopez test. 164-179 q Emerald Group Publishing Limited 1526-5943 DOI 10. 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. Risk assessment. In this research paper. and Downloaded by UNIVERSITAT DE VALENCIA At 05:03 17 July 2015 (PT) Sait Yılmazer Tekstilbank. 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. 11 No. 2. 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. Bradford. 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.com/1526-5943. Kupiec test. namely. Stock exchanges. Istanbul. Extreme-value theory (EVT) do not follow the . Keywords Stock returns. Emerging markets. Istanbul. Christoffersen test. GARCH with skewed student-t distribution. and the lack of informational efficiency create high volatility and extremes in returns. and h-step ahead forecasting RMSE. root mean squared error (RMSE). The performances of the filtered EVT models are compared to those of generalized autoregressive conditional heteroskedasticity (GARCH). Complex and volatile market conditions in emerging markets require dynamic and flexible econometric models that are able to capture extreme changes in financial variables. Bradford University.2 Filtered extreme-value theory for value-at-risk estimation: evidence from Turkey 164 Alper Ozun School of Management. 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. 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. GARCH with student-t distribution. and FIGARCH by using alternative back-testing algorithms. Findings – The results indicate that filtered EVT performs better in terms of capturing fat-tails in stock returns than parametric VaR models. UK Atilla Cifter S¸ekerbank.1108/15265941011025189 Introduction Estimating losses of financial instruments has become a crucial task for market risk management in the current global economy.The current issue and full text archive of this journal is available at www. unstable regulatory and political environment. 2010 pp.emeraldinsight. Turkey. Turkey Paper type Research paper The Journal of Risk Finance Vol. Diebold and Mariano test.
The back-test results show that the filtered expected shortfall has superior performance in estimating the extremes and presents a new. thus the model captures switching shocks and long-term memory in the stock returns (Ozun and Cifter. 2007 for our empirical research. therefore. Bekaert and Wu (2000) show that the leverage effect in equities determines a strong negative correlation between returns and volatility. For example. 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. extremes in returns might be caused by mechanisms that are structurally different from the usual dynamics of financial markets. 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. EVT which approximates the area under the tail asymptotically might be more powerful than imposing an explicit functional form. As an emerging market with dramatic macroeconomic and regulatory changes in recent years. the normal mixture GARCH (NM-GARCH) model is presented by Alexander and Lazar (2006). 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. 2002 to April 18. EVT has been used in financial risk estimation in recent years. the ISE-100 gives us an opportunity to work with a highly volatile and heavy-tailed data set. EVT is presented as major alternative for generalized autoregressive conditional heteroskedasticity (GARCH) models. In those extreme conditions. and flexible perspective in VaR estimation. EVT focuses on the extremes rather than mean. 2000). creates a crucial source of skewness in returns. we also apply h-step-ahead root mean square error (RMSE) and number of exceptions to measure performance of the filtered expected shortfall.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. First. then it follows Gaussian distribution. What is more. The performance of the model is also empirically compared with those of the parametric models with Kupiec (1995). However. classical VaR models focus on the whole empirical distribution of the returns rather than that of extreme returns. the individual variances are only tied with each other through their dependence on the error term. dynamic. As another original work. extremes might be the result of a major default or a speculative bubble. There are some important reasons to choose EVT against parametric volatility models. 2008). In the NM-GARCH model. managing extreme risk requires estimation of quantiles that usually are not directly captured from the time series data. Lopez (1999). On the other hand. Christoffersen and Jacobs (2004) empirically display that asymmetric GARCH model that captures the leverage effect provides best estimation within all GARCH models. As a recent model. Filtered extremevalue theory 165 . We use the time series of daily returns of the Istanbul Stock Index-100 (Istanbul Stock Exchange (ISE)-100) from January 2. the distribution of returns is heavy-tailed and asymmetric in most financial time series. Christoffersen (1998) and Diebold and Mariano (1995) back-test algorithm. GARCH approach dynamically improves itself in recent years. 99 percent of the time.
A more comprehensive literature review on the EVT . (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. performs best overall. Blum et al. They show that Hungary exhibits higher risk under extreme conditions indicating that its market is much more vulnerable than all other markets under study. McNeil and Frey (2000) and Bali (2003) empirically show that traditional parametric VaR models with normal density fail to estimate losses during financial crises. EVT has been seen as an alternative to GARCH models. Poon et al. Kuester et al. combining a heavy-tailed GARCH filter with an EVT-based approach.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. In literature. By using more than 30 years of daily return data on the NASDAQ Composite Index. They state that a hybrid method. (2005) compare the out-of-sample performance of VaR models and EVT. Inui and Kijima (2005) and Martins and Yao (2006) also empirically show that EVT is superior in risk estimation with financial time series. The empirical evidence shows that these series have significantly fatter tails than normal distribution and therefore suggests the use of EVT. Longin (2000). 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. In the literature. and Turkey. (2005). which combines GARCH model with EVT in applying the residuals from the GARCH process. the empirical evidence on EVT is generally based on the data from hedge funds and emerging financial markets. Morocco. namely Egypt. Especially. Assaf (2006) use the EVT to examine four emerging financial markets belonging to the Middle East and North Africa region. conditional historical simulation. Lhabitant (2002) and Gupta and Liang (2005) also prove that the EVT performs better with hedge-fund indices. in bear markets. (2002). this is no longer the case when skewness is considered. Jordan. (2004) show that extreme-value dependence is usually stronger in bear markets (left tails) than in bull markets (right tails). Empirical evidence from emerging markets is also in favor of EVT. Liang and Park (2007) empirically show that EVT is able to foresight the fat-tails in returns especially regarding high volatility in negative direction.JRF 11. Amin and Kat (2003) empirically show that while hedge funds combine well with stocks and bonds in the mean-variance framework. Acerbi (2002). They show that the parameters of this distribution appear to vary with a tendency to become less fat-tailed over time. By using hedge funds data. fat-tails are usually observed. 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. Kuester et al. and conditional EVT. Kalyvas et al. They underline the conditional quantile problem and apply EVT to the conditional return distribution by using a two-stage method. Extremes in returns are observed in time series data from hedge funds and emerging markets where high volatility and unstable money flows occur. EVT is compared to GARCH-based parametric VaR estimation models. In general. 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. He focuses on the tails of the unconditional distribution of returns in each market and provides their tail index behavior.
the tail estimation can be computed with equation (3): ! ^ ^FðxÞ ¼ 1 2 N u 1 þ j ðx 2 uÞ21=j^ ð3Þ s n In this equation. With Turkish data. (1999) evaluate expected shortfall as an alternative for VaR. (2007) use conditional quantile expected shortfall and generalized Pareto distribution (GPD) for interest rates and find that conditional EVT improves forecasting. the following equation holds: F u ð yÞ ¼ PðX 2 u # yjX . From similar perspective. GPD with constant or conditional quantile is obtained. (1997) and Focardi and Fabozzi (2003. n is the total number of observations in the data set. like EVT. F(u) distribution. measures changes in the market value of portfolio P on time Dt period with probability p (Dowd. by defining u with either constant or conditional quantile. (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. VaR for one day. F u ð yÞ < Gj. Gencay and Selcuk (2004). In this description. for T time period. For pre-defined p . Artzner et al. Altay and Kucukozmen (2006) and Eksi et al. estimated VaR will depend on the distribution of F function. 2004): P½DPDt # VaR ¼ p ð1Þ In the equation. F 2 1 is the reverse of the distribution function. risk is equal to TD-VaR((1 2 p)%). Cifter et al. When we include the time as a variable in the equation. In expected shortfall. are aimed at estimating the returns that are not within the confidence level (p) but extremes and fat-tails. Obviously. VaR( p%). Nu is the violations (extremes) above u. Semi-parametric models. From that perspective. since F(DPD) is a distribution function of changes in the portfolio value. uÞ.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. EVT employs the GPD with thresholds. under the assumptions that x ¼ u þ y. 0 # y # xF 2 u ð2Þ u!1 For negative returns. it is possible to create an equation like (VaR) ¼ F 2 1(p). Filtered extreme-value theory VaR reflects the change in a portfolio with a confidence level on a time period. DPDt. Gencay et al. the expected value of the portfolio return is taken into consideration if there is a violation. (2003). In the perspective of GPD.s ð yÞ. is calculated with the equation (4): ^ ! s n 2j VaRGPD ¼ u þ p 21 ð4Þ Nu j^ In equation (4). Risk for one day should be equal to VaR((1 2 p)%). Filtered extremevalue theory 167 . 2004). for pre-defined j and s.
Kupiec (1995) test. When we nominate the pre-defined VaR with a. For GPD. and 40 days rolling quantile are estimated. 1 parameters as in the equation (6): eðuÞ ¼ EðX 2 ujX . Lopez (1999) test. the probability of failure rate in the VaR estimation is the important point for back-testing. Kupiec test Kupiec (1995) test defines the failure ratio ( f ) as the excess values from VaR(x) to the total observations (T). and h-step ahead forecasting RMSE (70 days). three. likelihood ratio statistics with x 2 distribution for the Kupiec (1995) test can be given in equation (9): LR ¼ 2{logj f x ð1 2 f ÞT2x j 2 logja x ð1 2 aÞT2x j} ð9Þ Christoffersen test According to Christoffersen (1998) test. we can express the mean violation function with j . the first moment of each r exists. In this research paper. 15. To conduct the test. ten. 2002): ESp ¼ VaRp þ EðX 2 VaRp jX . five. 20. VaRp Þ 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. Methodologies of alternative back-tests Alternative back-testing algorithms are employed to compare the performance of the models. Baillie et al. h-step ahead number of exceptions and Diebold and Mariano (1995) encompassing test are used as the back-tests. and FIGARCH for performance comparison of the filtered expected shortfall. GARCH-t. then for all r . We use parametric models like GARCH. 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. but detailed examinations can be found in Chung (1999). four. VaRt ðaÞÞ and test H0 : p a ¼ a against H1 : p a – a. Davidson (2002) and Laurent and Peters (2002). eight different filtered expected shortfall estimation with two. (1996).2 Expected shortfall can be constructed with the following equation (Gilli and Kellezi. 0 From that point of view. uÞ ¼ Downloaded by UNIVERSITAT DE VALENCIA At 05:03 17 July 2015 (PT) ð5Þ s þ ju . number of exceptions. GARCH-skewed student-t. RMSE (70 days).JRF 11. 1/j integer. Christoffersen (1998) test. the expected shortfall is: s^ þ j^ðVaRp 2 uÞ ESp ¼ VaRp þ 1 2 j^ ESp ¼ VaRp s^ 2 j^u þ 1 2 j^ 1 2 j^ ð6Þ ð7Þ ð8Þ If X is GPD. 12j s þ ju . The methodologies for GARCH models are not examined here. one should first define p a ¼ Prð yt . (r). .
tþ1 # VaRtðaÞ ð12Þ By using that methodology. In the first step. Under the null hypothesis. Lopez (1999) defines the violation function LðVaRtðaÞ .The condition of {1ð yt . in order to reach an estimated value for a mean of loss distribution. The test value is calculated as the deviation of the h-step ahead forecasts of a variable. VaRtðaÞ xt. for example 10. In the last step. The smaller its values. 2003). the more accurate the forecasts become.tþ1 Þ that can be given in equation (12): ð1 þ ðxt. Filtered extremevalue theory 169 . back-test is conducted with equation (13): T 1X L^ ¼ LðVaRðaÞ .000. a mean for losses/gains is obtained. The RMSE of E( ytþ h) equals to the square root of the equation (14): vﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ u T þh u1 X ð14Þ RMSE ¼ t ðs^f . T þ 1 is the beginning of the testing sample. employing the model created in the first step and using historical losses/gains VaR and VaR(a) are constructed and Li.000. xt. xt. ytþ h.x 2 sr. The h-step ahead number of exceptions is more sensitive measure of forecasting rather than h-step ahead RMSE as considers tail loss directly.tþ1 2 VaRtðaÞ Þ2 Þ ð0Þ if if xt. the process described above is repeated many times. while T þ 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. Second. VaRt ðaÞÞ and n1 ¼ t¼R 1ð yt . the proper distribution of returns and statistics model is chosen. VaRt ðaÞÞ (Saltoglu. it becomes LðaÞ ¼ ð1 2 aÞn0 a n1 .tþ1 Þ T t ð13Þ H-step ahead RMSE and number of exceptions RMSE is a scaled dependent comparison algorithm for forecasts. Li ¼ 1i ¼ 10. The likelihood ratio test statistics can be given in equation (11): Downloaded by UNIVERSITAT DE VALENCIA At 05:03 17 July 2015 (PT) LR ¼ 22 InðLðaÞÞ d !xð1Þ ^ LðpÞÞ ð11Þ Lopez test Lopez (1999) performs the back-test in three steps. E( ytþ h). VaRðaÞ} has a binomial likelihood and can be given in equation (10): Lð p a Þ ¼ ð1 2 p a Þn0 ð p a Þn1 ð10Þ PT PT where n0 ¼ t¼R 1ð yt . from its observed time path.t Þ2 h t¼Tþ1 In the function.tþ1 .
2 2 2 Define u^ 21.12 Ise positive returns – 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 .2201.15 Ise 0.15 0 0. ISE negative and positive returns 0.02 0.05 – 0.325 daily observations including 610 observations with negative returns from January 2.1671.1 0. By performing augmented Dickey-Fuller (Dickey and Fuller. We constituted the series in log-differenced level.tþ1 21 and d ¼ P t d tþ1 ¼ MSE1 2 MSE2 where MSE represents RMSEs of forecasting models.14 – 0.05 0 1 54 107 160 213 266 319 372 425 478 531 584 637 690 743 796 849 902 955 1.14 1 – 0.12 – 0.273 – 0. 2002 to April 18. Our dataset covers 1. 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 ¼ rﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ P 2 P 22 ðdtþ1 2 dÞ ð15Þ t Data ISE Rate (ISE-100 Index) is received from Bloomberg. 1981) and 0. The main advantage of this statistic is that there is not any assumption on the distribution of forecast errors.04 0.1 Figure 2.tþ1 2 u^ 2. ISE log-differenced series – 0.0081.tþ1 as two forecast errors and estimate dtþ1 ¼ u^ 1.16 0.JRF 11.1141.tþ1 P and u^ 2. 2007.1 – 0.06 – 0.08 0.08 – 0.04 – 0.0611.2 170 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.06 0.1 Figure 1. Figure 1 shows ISE Index in log-differenced series where Figure 2 shows negative and positive returns separately.
1) estimates for ISE index.98092 20. and five days conditional quantile are statistically different at five percent confidence interval and filtered expected shortfall with ten.1)-skewed student-t. As a result. Figures 3 and 4 show filtered expected shortfall and GARCH models graphs. 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. like filtered expected shortfall.00091553 C 0.0000]a 399. 289 0. 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). GARCH(1. at-value 236.Downloaded by UNIVERSITAT DE VALENCIA At 05:03 17 July 2015 (PT) Phillips-Peron test (Philips and Perron. filtered expected shortfall with two. may capture tail loss better compared to GARCH and alternative GARCH models. 15.0707095 6. ten. and 40 days rolling and volatility models as GARCH(1. Also reported in Table IV Diebold-Mariano test (Diebold and Mariano.1)-skewed student-t.5121 * 236. and FIGARCH(1.031 Table I. and 40 days ISE Unit root tests ADF test P-P test Main stats. GARCH(1. Main statistical properties of the log-differenced index is shown in Table I. 20. 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]. four. Unit root test and main statistical properties . Back-testing is done with 95 and 99 percent confidence intervals and Basel guidelines require using a 99 percent confidence interval.0000]a 0.11794 at obs.1) fits better than other GARCH models. 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. Table II reports GARCH(1.1).03 [0.71 [0. 1995) shows that GARCH student-t and GARCH-skewed student-t are not statistically different from GARCH with Gaussian distribution where FIGARCH. GARCH(1.133408 at obs. 15. GARCH(1. five. four.1) are estimated. In this paper. three. Filtered extremevalue theory 171 Empirical evidence Filtered expected shortfall with conditional quantile of two. Kupiec (1995). Table III reports filtered expected shortfall shape (j) and scale (b) parameters of lower and upper tail. 215 876. filtered EVT. All the parameters of GARCH models are statistically significant and according to log-likelihood stat GARCH(1. 1988) we found that ISE Index is stationary at log differenced level as shown in Table I.1)-student-t. three.5267 * 876.1). 20.0212336 20.1)-student-t. and FIGARCH(1.
134 0. 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. filtered expected shortfall with 15 and .5043 (6.44 Notes: *Significance at 5 percent confidence level.187 0.085 * * (4.973) 0.007 0.184 0.300 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 – 0.99 0.395) 0.125 * * (5.072 0.90 0.051) – – – 0.13 0.2 172 Downloaded by UNIVERSITAT DE VALENCIA At 05:03 17 July 2015 (PT) Table II.011 0.139 0.15) – 2 0.022) 0.009 0.009 Upper tail Shape (j) Scale (b) 0. n-Skew d a þ b1 Loglike GARCH GARCH-t GARCH-skew FIGARCH 0.321.215 * * (5.011 0.21) – – – – 0.010 0.170 0.204 * (2.084 * * (4.889 * * (44. The predictive performance of filtered expected shortfall and GARCH models are reported in Tables V and VI with 95 and 99 percent confidence intervals.884 * * (36.0606 (1.0500000 – 0.351.853) 0.089 * * (5.238 0.912) 0.80434 3.011 0.001 * * (2. Estimation results from volatility models v a b1 n-student-t j-Ske.011 0.314) – – – 0.29) 8.115 0.011 0.010 0.325. * *significance at 1 percent level Lower tail Shape (j) Scale (b) Models Table III.96958 3.009 0.348 0.010 0.131 0.001 * * (2.0000000 – 0.2500000 conditional quantile are statistically different from GARCH model at 1 percent confidence interval.490) 0.064 0.1500000 Figure 3.886 * * (37.010 0.505) 8.034 0.325) 0. filtered expected shortfall with two days conditional quantile is the best one based on Lopez test.600 * * (6.JRF 11.008 0.0014 * (2.008 0.226 0.266) – 0.009 0.738) 0.001 * * (3.081 0.1000000 – 0.349. According to 95 percent confidence interval.114 0.97146 3.97840 3.500) 0. Filtered expected shortfall models – 0.
5092 0.185461 * 0. filtered expected shortfall with 15 and 40 days conditional quantile are the best ones based on Kupiec tests.094686 * 0.148612 * 0.06 173 –0.08 –0.1) with Gaussian distribution. and Kupiec tests therefore we applied h-step ahead forecasting of RMSE and number of exceptions.12 Garch Garch-t –0.1) is the best one[2]. Based on h-step ahead forecasting of RMSE GARCH(1.32498 * 0.0 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 –0.14 Garch-skew Figure 4.16 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. 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.077021 * * 0. Table VII shows that based on h-step ahead forecasting of number of exceptions up to 70 days filtered expected Table IV. filtered expected shortfall with two days conditional quantile performs better based on Christoffersen test.02 Filtered extremevalue theory –0. Since 99 percent confidence interval is more significant and Basel guidelines require using 99 percent confidence interval. Christoffersen.04 –0. GARCH models Figarch –0.047025 * * 0. we also consider 99 percent confidence interval for back testing results. Comparing predictive accuracy with the Diebold-Mariano statistic .5194 2 0. * *significance at 1 percent level 40 days conditional quantile performs best based on Christoffersen and Kupiec tests. *significance at 5 percent confidence level.085852 * * Notes: Benchmark model is GARCH(1.1 Downloaded by UNIVERSITAT DE VALENCIA At 05:03 17 July 2015 (PT) Ise –0.191513 * 0. filtered expected shortfall with ten and 20 days conditional quantile performs best based on Lopez test. According to 99 percent confidence interval.09754 * 0.
0501 0.00025 0.696 5.2 Downloaded by UNIVERSITAT DE VALENCIA At 05:03 17 July 2015 (PT) 174 Table V.06051 0.99999 0.20462 0. 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.99980 0.25970 0.23692 0.15074 0.712 7.99601 0.32276 0.257 25.85641 0.00000 0.00396 0.0630 0.00063 0.00000 Lopez Christoffersen Kupiec 0.0383 0.00065 0.0533 0.99846 0.0373 0.0498 0.00120 0.00023 0.0481 0.95528 0.0433 0.500 3.00000 0.06909 0.136 11.99943 0.0419 0.00065 0.01124 0.454 13. 15 13 14 14 22 16 13 12 8 5 8 4 15.85707 0.89262 0.00000 0.0151 RMSE tþ1 Avrg.196 17.98078 0.0640 Note: Average no.15879 0.0411 0.0479 0.74404 0.52972 0.41345 1.78774 0.0449 0.00000 0.01124 0.74404 1.00080 0.00000 0.41345 1.32276 Note: 95 percent confidence interval Models Table VI.0389 0.00595 0.66599 0.JRF 11.85707 0.00226 0. 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.68437 0.99907 0.575 14.11457 0.00022 0.42083 0.08361 0.00000 0.500 13.98265 1.515 13.00146 0.0402 0. 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 0.32134 0. of exceptions tþ1 Avrg.0617 0.01469 0.00409 0.17591 0.00315 0.00249 0.0614 0.00061 0.04985 0.0633 0.42083 0.0405 0.0586 0.00054 0.68437 0.15609 0.68934 1.7272 6.0602 0.00000 0.0537 0. of exceptions and RMSE is estimated with 70 days step-ahead forecasting .0515 0.00766 0.00958 0. 0.00065 0.00000 0.00021 0.
either normality or asymmetric distribution. 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. 40 35 Garch Figarch FES4 FES15 Garch-t FES2 FES5 FES20 Garch-skew FES3 FES10 FES40 30 No. thus indicating that filtering less than ten days conditional quantile may imitate GARCH models. H-step ahead forecasting . Observed extremes and fat tails in returns need to be estimated with relatively more sophisticated models. 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.Filtered extremevalue theory 175 Conclusions As a result of the dynamic and chaotic features of financial markets in emerging economies. of exceptions Downloaded by UNIVERSITAT DE VALENCIA At 05:03 17 July 2015 (PT) shortfall with 40 days conditional quantile is the best one. 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. These assumptions. of exceptions Figure 5. Parametric models have certain strict assumptions on the distribution function of returns. 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). Diebold-Mariano test of equal forecast accuracy (Diebold and Mariano. 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. are not able to make statistically significant estimations. financial forecasting is almost impossible with parametric models.
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 7.045 0.005 Figarch 0 1 Figure 6. if the sum of the variables has a finite variance. In this research paper. H-step ahead forecasting 4 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 0.02 Garch Garch-t 0.035 0.01 Garch-skew 0. we also estimate VaR with . The distribution of extremes in returns is limited to having the same form without relying on the distribution of the parent variable.04 0. By using daily returns of the ISE National 100 Index.05 0.025 0. we investigate filtered EVT with different rolling quantile to estimate VaR.03 0.015 0. on the other hand. According to the theorem.05 RMSE Downloaded by UNIVERSITAT DE VALENCIA At 05:03 17 July 2015 (PT) 0. H-step ahead forecasting H-step ahead forecasting Note: RMSE and filtered expected shortfall models EVT. For comparison of the model performance.07 0.02 0.04 176 RMSE 0.06 0.03 0. employs the central limit theorem for risk estimation. then it follows Gaussian distribution.2 0. we estimate risk with filtered EVT.JRF 11.
G.627426 0. C.. T.1) also the best one. Artzner. (1999). GARCH with student-t distribution. G. Christoffersen (1998) test.1): applications to exchange rate modeling”.T. 13 No. and Mikkelsen. RMSE (70 days). Izmir. D. “Fractionally integrated generalized autoregressive conditional heteroskedasticity”.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. Vol. T. (2006). pp. Journal of Financial and Quantitative Analysis. (2002). 307-36. “Normal mixture GARCH(1. In accordance with this perspective. Journal of Business. new semi-parametric models should be conducted in future research without ignoring econometric methodological concerns. pp. References Acerbi. pp. pp. 74. Vol. Vol. “Hedge fund performance 1990-2000: do the money machines really add value?”. 20 No. “Dependence and mean reversion in stock prices: the case of the MENA region”.2342 0. Journal of Banking & Finance. 1. Amin. C. and Heath. and h-step ahead forecasting RMSE (70 days). 2. A. 3. Notes 1. The success of the estimation of the models are compared by using Kupiec (1995) test. Bekaert. 83-108. 21 No. Comparing predictive accuracy with the Diebold-Mariano statistic . J. C. 3. Proceedings of ICBME. 38. Lopez (1999) test.612275 0. “An extreme value approach to estimating volatility and value at risk”. GARCH. 26 No. Bali..309795 * 0.323022 * 0. Alexander. (2003). pp. Altay. Filtered extremevalue theory 177 Table VIII. Vol. (2000). The results of the back-tests show that the filtered EVT has better risk forecasting performance than parametric VaR models. Vol. pp.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. G. Vol. Based on standard RMSE(t þ 1) GARCH(1. “Asymmetric volatility and risk equity markets”. Research in International Business and Finance.. VaR results of upper tail estimation can be obtained from the authors. Mathematical Finance. Vol. P. Baillie. 1-42. and Wu. (2003). 3-30. Journal of Econometrics. “Coherent measures of risk”. pp. 9.M. S. (2006). F. Journal of Applied Econometrics. GARCH with skewed student-t distribution and FIGARCH. H.264849 * 0. The Review of Financial Studies. Yasar University. Bollerslev. “An assessment of value-at-risk (VaR) and expected tail loss (ETL) under a stress testing framework for Turkish stock market”. *significance at 5 percent confidence level parametric models. 1505-18. 76 No. R. Eber. and Kucukozmen. and Lazar. 7. E. (1996). 203-28. Vol. H. and Kat. namely. “Spectral measures of risk: a coherent representation of subjective risk aversion”. 286-304. 1. Delbaen. We conclude that financial forecasting especially in dynamic markets needs flexible models. (2006). pp. Assaf. 1-24.
77. Occasional Papers. pp. A. “FOMC forecasts of macroeconomic risks”. (2004). “Do hedge funds have enough capital? A value-at-risk approach”. 20. International Economic Review. Christoffersen. P. 49. in Lane. “Asymmetric and fractionally integrated GARCH models with (skewed) student-t and GED distribution in risk management”. Banks Association of Turkey-Bankers’ Journal. Computational Methods in Decision-making. Chung. (2004). Vol. Proceedings of 8th National Finance Symposium. J. (1998). S. A. and Diebold. 3-16. 12. S. F.X. 12-22. “How relevant is volatility forecasting for financial risk management?”. Risk Books. Journal of Risk Finance. (2002). 841-62. and Mikosch. (2004). and Ulugulyagci. Focardi. Gupta. P. Industrial Economics Division. 50. “Comparing predictive accuracy”. Econometrica. Davidson. and Fuller. pp. “Moment and memory properties of linear conditional heteroscedasticity models”. London. E. Izmir. Istanbul.. and Fabozzi. Vol.. pp.A. Gilli. 35-47. (2007). (1997). (1981).ac. working paper. Cifter. pp.JRF 11. Gencay. 5 No.. Embrechts. Proceedings of ICBME. D. 3. F. scaling and stable laws”. Berlin. F. “Alternative risk measures and extreme value theory in finance: implementation on ISE 100 index”. I. Ozun. pp. Kluwer Academic. C. (Ed. and Yilmazer. in Kontoghiorghes.. 33. S. Vol. A. Focardi. R.A.. Vol. 13 No. “Evaluating interval forecasts”. “Extreme value theory and value-at-risk: relative performance in emerging markets”. W. Economics and Finance. 1. The Mathematics of Financial Modeling and Investment Management. “Estimating the fractionally integrated GARCH model”. (2006). A. P. and Yildirak. S. 25-36. “High volatility. Vol. Journal of Business & Economics Statistics. 337-56. and Jacobs.F. 39. Yasar University. 82. (2005). T. working paper. C. (2004). (2003). No. “Likelihood ratio statistics for autoregressive time series with a unit root”. R. Diebold. Wiley. Modeling Extremal Events for Insurance and Finance. K.uk/carbs/econ/davidsonje/ (accessed on April 10. Gencay. K. Springer. pp. Eksi. NY. (1999). Dickey. thick tails and extreme value theory in value-at-risk estimation”. Insurance: Mathematics and Economics. pp. Christoffersen. (2002). Vol.S. K. . Yildirim. A. Dowd. Selcuk. Vol. 253-63. P. and Selcuk.2 Downloaded by UNIVERSITAT DE VALENCIA At 05:03 17 July 2015 (PT) 178 Blum. B. Istanbul Technical University. R. pp. Kluwer Applied Optimization Series. Management Science. Dias. Vol. Z..). K. Review of Economics and Statistics. B. New York. (2003). “Fat tails. pp. New York. M. and Siokos. pp. A. International Journal of Forecasting. 2007). Taipei. 60. NY. P. Vol. Christoffersen. 1057-72. E. “Which GARCH model for option valuation?”. Kluppelberg. “Value-at-risk with expected shortfall and generalized Pareto distribution: some comparative evidence for Turkish interest rate markets”. 1204-21. Alternative Risk Strategies. and Mariano. 167-83. “The art of dependence modeling: the latest advances in correlation analysis”. F. pp. P.J. and Liang. F. and Fabozzi. (2000). (2002). F. (1995). available at: www. National Taiwan University. Cifter. and Kellezi. and Embrechts. Rustem. (Eds). 287-303. (2004).cf. Journal of Financial Economics. “Portfolio optimization with VaR and expected shortfall”.
29 No. (2004). (2000). pp. 4. 75. 1. Sriopoulos. and Park. and Yao. (2006). T. Neftci. Laurent. International Review of Electrical Engineering.S. Vol. J. C. A. 2. S. Kupiec..2: an Ox package for estimating and forecasting various ARCH models”. 3. 335-446. 997-1015. 107-49. Kuester.P. “On the significance of expected shortfall as a coherent risk measure”.emeraldinsight. (2007). McNeil. 580-91. 3. Yaylim Press. “Value-at-risk prediction: a comparison of alternative strategies”. 333-70. 16. 7. 3 No. Vol. International Journal of Financial Services Management. Journal of Economic Survey. Saltoglu. Vol. Vol.H. pp. Vol. 1-22.. European Financial Management. Liang. and Tawn. Journal of Financial Econometrics. pp. A. A. 1097-130. 24. and Peters. K. NY. and Perron. “Estimation of tail-related risk measures for heteroscedastic financial time series: an extreme value approach”. and Brown. (2007). (2002). (2002). Y. and Cifter. F. Vol. “A signal processing model for time series analysis: the effects of international F/X markets on domestic currencies using wavelet networks”. 2. Rockinger.com/reprints Filtered extremevalue theory 179 . P. “Methods for evaluating value-at-risk models”.uk To purchase reprints of this article please e-mail: reprints@emeraldinsight. (2003).C. Poon. (2000). and Paolella. and Georgopoulos.B. B. The Review of Financial Studies. A High Frequency Analysis of Financial Risk and Crisis: An Empirical Study on Turkish Financial Market. 53-89. Journal of Banking & Finance. Vol.com Or visit our web site for further details: www. L. extreme events. pp. Lhabitant. K. 2 Nos 1/2. Vol. Lopez. (1999). pp. pp. 73-84. Hedge Funds: Myths and Limits. pp. Studies in Nonlinear Dynamics and Econometrics. “G@rch 2. Tolikas. Vol. Longin.M. pp. 12-19. models and financial implications”. F.Downloaded by UNIVERSITAT DE VALENCIA At 05:03 17 July 2015 (PT) Inui. Martinz. pp. and Frey. and tail estimation”. Yamai. “Extreme value dependence in financial markets: diagnostics.ac. Mittik. M. A. “An investigation of riskiness in South and Eastern European markets”. Vol. Federal Reserve Bank of San Francisco Economic Review. “Value at risk calculations. S. Journal of Derivatives. (2006). (1995). P. (2005). R. “Testing for a unit root in time series regression”. Vol. pp. B. Corresponding author Alper Ozun can be contacted at: A. “Risk measures for hedge funds: a cross-sectional approach”. F. Istanbul. 2. Journal of Banking & Finance. A. pp.ozun@Brad. Philips. 2. Journal of Empirical Finance. 10 No. (1988). 3-17. F.. “Value-at-risk versus expected shortfall: a practical perspective”. Biometrica. Steftos. R. Spring. (2005). Vol. 12 No. and Kijima. 271-300. 853-64. Journal of Derivatives. 586-610. P.A. H. J. Vol. pp. 17 No.S. pp. 4 No. and Yoshiba. “The distribution of the extreme daily share returns in the Athens Stock Exchange”. (2008). Kalyvas. 13 No. The European Journal of Finance. pp.J..C. (2005). S. (2000). J. pp. “From value at risk to stress testing: the extreme value approach”. S. M. 1. 29. Vol. “Techniques for verifying the accuracy of risk measurement models”. M. 447-85. Vol. Wiley. Ozun. K. pp. “Estimation of value-at-risk and expected shortfall based on nonlinear models of return dynamics and extreme value theory”. 13-21. Journal of Banking & Finance. New York.
[CrossRef] 2. Evaluating value‐at‐risk models before and after the financial crisis of 2008. Stavros Degiannakis. 436-452.This article has been cited by: Downloaded by UNIVERSITAT DE VALENCIA At 05:03 17 July 2015 (PT) 1. Carmen López. 2012. Sonia Benito. Alexandra Livada. The Spanish Review of Financial Economics 12. A comprehensive review of Value at Risk methodologies. Pilar Abad. 2014. Managerial Finance 38:4. [Abstract] [Full Text] [PDF] . 15-32. Christos Floros.