Printer Friendly

El nuevo enfoque hibrido de value at risk basado en la teoria de valores extremos.

The new hybrid value at risk approach based on the extreme value theory

1. INTRODUCTION

Although there is a widespread agreement about the use of VaR as a general measure of market risk and the economic loss that banks and other financial institutions may suffer due to exposure to the market risk, there is no consensus on the preferred approach to VaR's calculation. The difficulties in obtaining reliable VaR estimates derive from the fact that all the existing approaches involve certain compromises and simplifications (Boudoukh, Richardson and Whitelaw, 1998). They are based on certain assumptions which greatly simplify the real market conditions. The assumptions on which these approaches are based represent a compromise between the efficiency of implementation, on one hand, and the statistical precision of market risk estimates on the other. Hence, determining the best approach to VaR estimation is an empirical question and a question of implementation. In other words, the choice of the optimal approach represents a decision between the efficiency of implementation and the statistic reliability of the approach. It may depend on the number and types of assets in portfolios and their sensitivity to changes in risk factors. The validity of the VaR models application primarily depends on a degree of compatibility between the characteristics of the real environment and the assumptions on which the models are based. Apart from that, the VaR estimates do not always fulfill all the characteristics of the coherent risk measures. The VaR estimates do not universally exhibit sub-additivity. The risk of a portfolio can be greater than the sum of the stand-alone risks of its components.

Despite the significant differences which derive from the differences in theoretical postulates on which approaches are based, a common feature of the most popular VaR approaches is their inability to be simultaneously effective in capturing leptokurtosis and strong time-varying volatility. In general, it can be said that nonparametric approaches are effective in capturing kurtosis and fat tails, but they will not be successful in capturing heteroscedasticity. On the negative side, nonparametric approaches depend too much on the historical data set, react slowly to changes in the market and are subject to predictable jumps in their forecasts of volatility (Zikovic, 2010). On the other hand, parametric approaches can be expected to be successful in capturing the dynamics of the time series of returns, but also quite unreliable when empirical distribution is deviates from the theoretical.

Since many empirical studies of the emerging markets show that the series of returns are characterized by excess kurtosis when compared to normal distribution (extreme financial returns are more likely than the normal distribution implies), they show a significant degree of autocorrelation and heteroscedasticity. It means that VaR approaches which are based on constant volatility, such as the Historical Simulation approach (HS), or VaR approaches that take simple techniques of modeling conditional volatility, such as equally weighted and exponentially weighted (e.g. RiskMetrics) models, will not be able to capture adequately the dynamics of returns in these markets (to be noticed that some authors, such as Schittenkopf, Lehar and Scheicher (2002), Harmantzis, Miao and Chien (2006) claim that complex techniques of modeling conditional volatility bring no significant improvement to VaR estimates in conditions of high volatility). This means that, in the context of emerging markets, some popular and most widely used VaR approaches are based on false assumptions. This is very indicative for risk management in banks and other financial institutions, because when elementary assumptions of most VaR approaches are not satisfied, VaR estimates will be unreliable and, at best, they can only provide unconditional coverage. The optimal approach for emerging markets is the one that can equally well capture both leptokurtosis (excess kurtosis) and time-varying volatility (heteroscedasticity). In order to capture successfully these specifics of emerging markets, it is necessary to design a sophisticated model of conditional volatility, which will also be easy to implement.

Considering these requirements, a new VaR approach for estimating market risk of the portfolio of banks, which operate in emerging markets, has been developed in this paper. The approach is designated as the new hybrid VaR approach based on the EVT. It is based on the solution proposed by McNeil and Fray (2000) in the application of EVT to estimate the market risk. Also, this approach incorporates the solutions proposed by Hull and White (1998) in improvement of HS approach and the solutions proposed by Zikovic (2010) in the development of semi-parametric model to VaR estimation. The approach is designed to combine the best features of HS approach, the application of EVT to VaR estimation and the advantages of GARCH (p,q) model to capture the heteroscedasticity. In other words, the approach is designed to capture successfully the two most conspicuous characteristics of financial asset returns with emerging markets, namely strong time varying volatility and excess kurtosis.

The paper is organized as follows: Section 1 contains the introduction. Section 2 gives an overview of the most significant, recent empirical researches in the area of VaR models. Section 3 presents a theoretical background of the new hybrid VaR approach based on the EVT. Section 4 provides a brief description of the analyzed data, the methodology used and the descriptive statistics of selected emerging markets. The backtesting results are presented in section 5. Since both backtesting tests used in the paper and the Kupiec's and Christoffersen's conditional coverage test are based on certain asymptotic assumptions and don't show the desired characteristics when performed on finite size samples, verification of their results was necessary. The Dufour (2006) Monte Carlo testing technique was used for that purpose. The final section summarizes the conclusions.

2. LITERATURE REVIEW

The recent studies about the applicability of VaR approaches in the emerging markets in terms of meeting the backtesting rules of Basel Committee, such as the studies conducted by Diamandis et al. (2011), Sener, Baronyana and Menguturk (2012), Rossignolo, Fethib and Shaban (2012, 2013), Cui et al. (2013), Louzis, Xanthopoulos-Sisinis and Refenes (2014) and Del Brio, Mora-Valencia and Javier (2014), indicate the importance of developing the most appropriate VaR approach for measuring the market risk at the emerging markets. At the same time, they suggest that regulatory authorities should determine the use of approaches which can capture the heavy tails (particularly EVT approach) and discourage or prohibit the use of traditional VaR approaches, especially the Linear, HS and Filtered Historical Simulation (FHS) approaches. In other words, the results of these studies confirm the conclusions of earlier researches conducted by Lucas (2000), Berkowitz and O'Brien (2002), Wong, Cheng and Wong (2002), Gencay and Selcuk (2004), pointing out that neither the popular nonparametric, hybrid nor parametric approaches can provide reliable VaR estimates when the volatility is not constant over time and that they also cannot manage the extreme events and losses that fall at the end of the distribution tail. The first one to try to design the VaR approach adequate for emerging markets was Zikovic (2005). Zikovic's approach, known in literature as the Volatility and Time Weighted Historical Simulation approach (VTWHS), represents a simple combination of the two popular VaR approaches, Hull-White's and Hybrid approach, proposed by Boudoukh, Richardson and Whitelaw. Essentially, the approach represents an attempt to exploit the advantages of both approaches. The importance of this approach is reflected in the fact that it is a pioneer work in developing an adequate (optimal) approach to emerging markets. To improve the applicability of the Hybrid approach in emerging markets, Zikovic and Prohaska (2008) developed a procedure to determine the optimal decline factor. They tested the proposed procedure on the example of nine Mediterranean stock markets. The results were very poor. The main reason for this lies in the fact that the Hybrid approach reacts slowly to changes in the basic risk factors, despite its theoretical foundations. Zikovic (2010) suggested the Hybrid Historical Simulation approach (HHS). The HHS approach is based on a combination of a modified recursive bootstrap procedure proposed by Freedman and Peters (1984) and the parametric GARCH approach to volatility forecasting. The HHS approach is easy to implement. It operates with the observed data but is not free of distributional assumption, since the use of nonparametric bootstrapping requests that the observed returns should be identically and independently distributed (IID). The results of Zikovic's (2010) researches show that HHS approach adequately captures market risk in emerging markets of EU new member states. The main flaw of this approach is related to the use of re-sampling methods. An interesting nonparametric approach to VaR estimate has been proposed by Alemany, Bolance and Guillen (2012). The approach is based on the double transformation of kernel estimation of the cumulative distribution function. The authors state that the approach is useful for large data sets and that it improves the quantile estimation compared to other methods in heavy tailed distributions (Alemany, Bolance and Guillen. 2012). Unfortunately, the approach is more useful for measuring the operating rather than the market risk. Some interesting solutions in capturing the asymmetry of the basic data were also proposed by Cener, Baronyana, Menguturk (2012) and Louzis, Xanthopoulos-Sisinis and Refenes (2014). Their solutions are based on the view that the validity of VaR approach depends not only on confidence level, as discussed by Beder (1995) and Christoffersen, Hahn and Inoue (2001), but also on the market conditions. They glorify the EVT approach for dealing with fat tails and extreme returns, which are otherwise typical for the emerging markets. For that reason, Sener, Baronyana and Menguturk (2012) advocate that Conditional Autoregressive Value at Risk by regression quantiles (CAVaiR), proposed by Engle and Manganelli (2004), should be used combined with the EVT approach, but Louzis, Xanthopoulos-Sisinis and Refenes (2014) suggest that Asymmetric Heterogeneous Autoregressive (Asym. HAR) model, proposed by Louzis, Xanthopoulos-Sisinis and Refenes (2012) and Corsi and Reno (2012), should be used together with the EVT approach. The results of their researches show that the proposed solutions provide more statistically accurate VaR estimates, which minimize capital charges and allow more efficient capital allocations. However, when using these solutions, it should be considered that they are computationally more demanding when compared to the most commonly used approaches for VaR estimate.

In order to capture both characteristics of the financial data, heavy tails and heteroscedasticity, Bee (2012) presented the Dynamic Fat Tailed approach to VaR estimate. More precisely, he offered three dynamic VaR models: the Dynamic VaR model with Student t innovations, the Dynamic VaR model with Generalized Error Distribution (GED) innovations and the Dynamic Historical Simulation model (DHS). The first model is based on the use of a standardized Student t random variable as a model for the stochastic component of the GARCH process. The second model is based on modeling the residuals by means of the GED. DHS is semi-parametric. This model is very similar to the filtered HS model proposed by Fernandez (2003). The results of these models' validity tests testify in favor of DHS, which performs very-well, for an extremely high level of confidence. The research covers the developed markets, mainly the EU and the US, so there is no data on the application validity on the models in emerging markets.

3. THE THEORETICAL BACKGROUND OF THE NEW HYBRID VAR APPROACH BASED ON THE EVT

As announced in the introduction, a new hybrid approach to the estimation of VaR and ES was developed in this paper. The proposed approach is based on the EVT. The starting point in the development of this approach was the fact that heteroscedasticity and the presence of autocorrelation are common features in series of financial data from the emerging markets, as well as the fact that extreme returns in these markets are more likely to appear than the presumption of elliptical distribution implies. Therefore, the basic idea on which the approach is based is that the dynamics of the returns of stock indexes in the emerging markets can successfully be captured by a simple AR(p)-GARCH(p, q) model.

The aim is to generate the standardized (IID) residuals, in order to obtain a stationary, or the IID series of returns which will be suitable for updating the volatility, according to the approach proposed by Hull and White (1998). The approach provides a coherent measure of risk. The final outcome of the approach is the ES-EVT. One disadvantage of the use of the most popular and widespread approaches to VaR estimation in the emerging markets is that their VaR estimates do not satisfy all the properties of a coherent risk measure. The reason for this is emphasized in the introduction. The final outcome of this approach is the ES-EVT, which is a coherent risk measure.

The approach is easy to understand and implement in practice. The number of parameters that have to be estimated is relatively small, and this number is determined by the GARCH specific structure and by the number of parameters of the extreme value distribution, because the approach is based on the assumption that extreme returns over a defined threshold (u) follow the Generalised Pareto Distribution (GPD) with the tail index ([xi]) over 0. This assumption is relevant to the financial data because it suits fat tails.

The implementation of the approach begins from the basic specification of the autoregressive model that the return can be predicted by its past values and process innovations, which follows a white noise process:

(1) [r.sub.i] = [[alpha].sub.o] + [p.summation over (i=1)][[alpha].sub.i][r.sub.t-i] + [[epsilon].sub.t]

The first step in the implementation of the approach involves fitting the AR(p) model into a series of historical returns, in order to ensure IID residuals:

(2) [[sigma].sup.2.sub.t] = [omega] + [q.summation over (i=1)] [[alpha].sub.i][[epsilon].sup.2.sub.t-i] + i [p.summation over (i=1)][[beta].sub.i][[sigma].sup.2.sub.t-i]

In the second step, a GARCH(p,q) model is fitted into the obtained residuals:

(3) [[sigma].sup.2.sub.t] = [omega] + [q.summation over (i=1)][[alpha].sub.i][[epsilon].sup.2.sub.t-i][[beta].sub.i][[sigma].sup.2.sub.t-i]

In the third step, the residuals ([[epsilon].sub.t]) obtained by applying the AR(p) model are divided by a corresponding conditional GARCH(p,q) volatility forecast ([[sigma].sub.t]), to obtain the standardized residuals ([z.sub.t]):

(4) [z.sub.t] = [[epsilon].sub.t]/[[sigma].sub.t]

The next step implies that the standardized residuals ([z.sub.t]) are multiplied with the latest GARCH volatility forecast ([[??].sub.t+1]) to obtain a series of historical residuals which have been updated for the latest volatility forecast in order to get a series of residuals which reflects the current market conditions {[[??].sub.t+1]}.

(5) [[??].sub.t+i] = ([z.sub.t]) x ([[??].sub.t+1])

After this, the simulated returns are obtained by using the updated historical residual ([[??].sub.t+1]) in the Equation (1):

(6) [[??].sub.t+l] = [[alpha].sub.0] + [p.summation over (i=1)][[alpha].sub.i][t.sub.t-i+1] + [[??].sub.t+i]

Thus obtained returns are suitable to be used for the VaR and ES estimation by applying the EVT.

In the final step, assuming that tail index ([xi]) is less than 1, an ES-EVT estimate is obtained by using the following equation:

(7) ES - EVT = [VaR.sub.cl]/1 - [xi] + [sigma] -[xi]u/1 - [xi]

noting that a [VaR.sub.cl] estimate can be calculated as:

(8) [VaR.sub.cl] = u + [sigma]/[xi][[(1 - cl)/k/l).sup.[xi]]] -1]

where (k) represents the number of exceedings over the defined threshold (u), ([sigma]) is the scale parameter and (n) the number of observations, or

(9) [MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII]

when the tail index ([xi]) is estimated by the Hill estimator.

Such a specified model provides a coherent market risk measure. The main advantage of this approach is its flexibility. It can be used to obtain the semiparametric VaR estimates. In addition to this feature, the model flexibility can be seen through the possibility of applying various ARCH models. If the obtained standardized residuals, which are calculated in the third step, are not IID, some other ARCH model may be applied (i.e. IGARCH, EGArCh etc). The reason why the approach is based on the basic GARCH(p,q) model is that it is the simplest model able to capture the volatility of clusters and leptokurtosis in the data. The other models, despite the fact that they are able to capture the various empirical features of returns and generate reasonably accurate out-of-sample predictions of the entire distribution of future returns or just particular quantiles, as is needed for VaR forecasting, have the drawbacks because they require a relatively large number of parameters that cannot be solved in a closed, analytical form and can result in negative values, where both problems have a negative influence on the maximum likelihood estimate. The more volatility models get complex, the more estimated parameters become unstable, making such models vulnerable to parameter misspecification and model risk (Zikovic, 2010). A good example of this is the EGARCH(p,q) model suggested by Nelson (1991). Despite its numerous theoretical advantages over the basic GARCH(p,q) model, the EGARCH(p,q) model is known to be very problematic in practice, with the choice of starting values being extremely critical for successful likelihood maximization (Franses and van Dijk (1996), Johnson, 2001). Furthermore, some researches, such as the research conducted by Rossignolo, Fethib and Shaban (2012), showed that the EGARCH(p,q) model did not bring any significant advantage in volatility estimate over the GARCH model in emerging and frontier markets, pointing out that the density assumption was more important than the model specification itself.

The main flaw in the suggested approach comes from the limitations of the application of EVT to VaR estimation. When using the EVT model, we should be aware of its limitations, since it is developed by using the asymptotic arguments, which can create difficulties when applied to finite samples. The critical factor is the choice of the threshold. When applying the EVT, we should be careful about the size of the tail. The choice of the tail size can affect the VaR estimates through its effect on the tail index estimate. This is known as a compromise between the variance and the partiality. See details concerning this issue in the paper presented by Gonzalo and Olmo (2004).

4. DATA AND THE METHODOLOGY OF ANALYSIS

The daily logarithmic returns of the stock indexes of a EU candidate (Serbia) and potential candidate countries (Bosnia and Herzegovina, Macedonia, Montenegro and Turkey) and a EU member (Croatia) were used for the performance analyses of the new hybrid approach. The tested stock indexes are the BIRS (Bosnia and Herzegovina), MONEX20 (Montenegro), MBI10 (Macedonia), BELEXline (Serbia), CROBEX (Croatia) and UX100 (Turkey). For the emerging markets, such as these, an extremely short history of securities trading and the phenomenon of non-synchronous trading can cause serious problems to a reliable statistical analysis. Therefore, to overcome the problem previously mentioned (namely, the short time series of returns of individual stocks and their highly variable liquidity), we use the stock indexes.

The returns are collected from the official stock exchange web sites of these countries for the period between February 2, 2009 and February 2, 2012. These data cover the periods of volatility patterns observed in the EU market. The daily returns of selected stocks are generated using the logarithmic approximation:

(10) [R.sub.i,t] = log [P.sub.i,t]/[p.sub.i,t-1])

where [P.sub.i,t] represents the closing price of asset i on the day t.

The VaR estimates were calculated for one day holding period and for the confidence level of 99 and 95%. The confidence levels were chosen taking into consideration the Basel Accord as well as the basic characteristics of the VaR calculation. The confidence level of 95% is appropriate for application in stable market conditions, while the 99% confidence level is appropriate for application in volatile market conditions. The VaR backtesting period is formed by taking out 253 of the latest observations from each stock index. The rest of the observations are used for volatility model calibration.

At the beginning of the analysis, the characteristics of selected markets for the entire observation period were analyzed. Table 1 gives a summary of the descriptive statistics and normality tests for the entire analyzed sample for all of the stock indexes. The descriptive statistics of the selected stock indexes confirm the results of the recent studies. The stock indexes show a great difference between their maximum and minimum returns. The standard deviations are also high, indicating a high level of fluctuations of the daily returns. The analysis of the selected stock indexes distribution shows that stock indexes have a significantly fatter distribution tails than assumed under normality, ranging from 2,4, in the case of the XU100 index to 9,3, in the case of the MBI10 index. In other words, all the analyzed stock indexes show a significant leptokurtosis. The skewness of all stock indexes is significantly different from zero, which indicates that the stock indexes have asymmetric returns. There is also evidence of negative skewness in the case of XU100, which means that the left tail is particularly extreme. In order to examine formally whether returns follow the normal distribution, we employed the Jarque-Bera test. The value of the Jarque-Bera test indicates that we should reject the null hypothesis of normality providing the evidence that the return series are not normally distributed.

The presence of autocorrelation and the presence of conditional heterosce-dasticity (the ARCH effect) in returns of the selected stock indexes are tested by their sample autocorrelation (ACF) and sample partial autocorrelation function (PAFC), calculating the Ljung-Box Q statistic and the Lagrange Multiplier test to test the presence of the ARCH effect. The results of these tests, for each of the selected stock indexes, are presented in Table 2 and Tables 3 and 3a in the appendix. As expected, the results of these tests confirm that there is a significant autocorrelation and the ARCH effect in returns. The aforesaid leads to the conclusion that classical VaR approaches couldn't estimate the true risk level in these markets.

Since the employed test discovers significant autocorrelation and heterosce-dasticity in returns of the selected stock indexes, the original data should be transformed to the IID. As autocorrelation has been detected in both returns and squared returns, the returns should be modeled as an AR-GARCH process in order to deal successfully with both types of dependence. Assuming that the conditional volatility in these markets can be adequately captured by the simplest GARCH(1,1) model, the original data are transformed by applying an AR(p)-GARCH(U).

The parameters of the AR(p)-GARCH(1,1) model were rated by maximum likelihood estimation. The Levenberg-Marquardt algorithm was used to estimate the parameter of the AR(p) model. The maximum likelihood estimation is used to estimate the parameters of the GARCH(1,1) by choosing the parameters that maximize the Gaussian log-likelihood function:

(11) LR = [T.summation over (i=1)][-ln[square root of 2[pi]] - 1/2 [[epsilon].sup.2.sub.i]/[[sigma].sup.2.sub.i] ([alpha],[beta],[omega]) - 1/2 ln [[sigma].sup.2.sub.i] ([alpha],[beta],[omega])]

The estimated AR(p)-GARCH(1,1) parameters for each of the selected stock indexes are given in Table 4. All estimated parameters are statistically significant.

There are several methods for estimating the tail index of extreme value distribution from the empirical data. In this paper, we used the Hill estimator because it has more desirable properties than the other estimators:

(12) [[??].sup.H] = 1/k [k.summation over (i=1)] ln ([x.sub.n-k+1]) - ln([x.sub.n-k])

The crucial step in estimating the tail index is the determination of a threshold (u). The threshold value for each index is determined by applying the rule of thumb for determining the threshold which was proposed by Christcoffersen (2011). Christoffersen (2011) points out that, for big samples, a good rule of thumb is setting the threshold so that 5% of the greatest observations for estimating the tail index should be found in the distribution tail.

The threshold (u) will then simply be the 95th percentile of the data set. This instruction is applied in the paper. The value of thresholds and the maximum likelihood estimates of the tail index and maximum likelihood estimates of the sigma, for each stock index, are presented in Table 5.

5. THE BACKTESTING RESULTS

In this section of the paper the backtesting results of the suggested approach are presented, analyzed and discussed. The approach is evaluated in terms of its accuracy in estimating VaR over the last 253 days of the observed period. The approach was tested as follows: first, the daily VaR estimates, which were obtained for confidence levels of 95% and 99%, were compared with the actual return movement that occurred in the backtesting period. In the case where the actual loss on a particular day exceeded the VaR estimate for that day, it was concluded that a VaR break had occurred. Then, the number/percentage of the VaR breaks over the backtesting period of the 253 days was established. According to Jorion (2006), in a good model the percentage of VaR breaks should be equal to one minus the level of confidence. In this case, it means that the number of VaR breaks mustn't exceed 3 at a 99% confidence level (1% of VaR estimates total number), i.e. not more than 13 VaR breaks at a 95% confidence level (5% of VaR estimates total number).

The number/percentage of VaR breaks at a 95% and 99% confidence levels over the backtesting period, separately for each of the selected stock indexes, are given in Table 6. As it can be seen in Table 6, percentages of VaR breaks are lower than the theoretical percentage values in all of the six emerging markets. The exception appears in the case of the MBI10 stock index at a 99% confidence level. The approach showed the best performances in the case of the BIRS index, since no VaR breaks were made at a 99% confidence level over the backtesting period of the 253 days and in the case of the CROBEX index at a 95% confidence level. In the case of the MBI10 stock index at a 99% confidence level, the number of breaks is equal to the expected frequencies of VaR breaks. For this reason, the percentage of breaks is higher that the theoretical value.

In order to determine whether the percentage of VaR breaks can be considered as equal to the theoretical value, we employed the unconditional coverage test introduced by Kupiec (1995). The Kupiec's test of the unconditional coverage represents the most widely used model for testing the VaR approach validity. The idea behind this test is that the frequency of VaR breaks should be statistically consistent with the probability level for which VaR is estimated (Samanta et al. 2010). In this paper we used the Kupiec's test at 5% significance level, because a significance level of this magnitude generates clear evidence about the validity of the approach and implies that a model should be rejected only if the evidence against it is reasonably strong. The following likelihood ratio test was employed to test null hypothesis:

(13) [MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII]

where (p) is the tail probability (or the VaR coverage rate).

The backtesting results for VaR at the 95% confidence level are presented in Table 7.

As can be seen from Table 7, the approach satisfied the Kupiec's test at the 5 percent significance level in all of the six emerging markets. Although it is very informative to look at VaR approach performance at different confidence levels, the true test of the VaR model acceptability to regulators is its performance at 99% confidence level, as prescribed by the Basel Committee. The back-testing results for VaR at the 99% confidence level are presented in Table 8.

As can be seen from Table 8, in all of the six emerging markets, the approach satisfied the Kupiec's test at the 5 percent significance level.

However, one of the basic disadvantages of the Kupiec's test is that it considers only the number of VaR breaks and not the time when they occur. In other words, a shortcoming of this test is that it focuses exclusively on the unconditional coverage property of an adequate VaR measure and does not examine the extent to which the independence property is satisfied. The Kupiec's test is based on the assumption that the VaR estimates are efficient, which means that they incorporate all the information known at the time of the forecast. The history of VaR break does not give any information whether the VaR break will happen again or not. As a result it is expected that the probability of occurrence of a new VaR break after the previous one is the same as the probability of its occurrence after the days in which the VaR break did not occur. In other words, the test was based on the expectation that the VaR break would be evenly distributed over the backtesting period. This is equivalent to the assumption that risk forecasts will be independently distributed over time. That is why the time of the VaR break occurrence is not important and it focuses only on the unconditional coverage. The independence property of VaR breaks is nevertheless an essential property because any measure of risk must adapt automatically and immediately to any new information which entails a new evolution in the dynamics of asset returns. If the approach ignores such dynamics then the VaR will react slowly to changing market conditions and VaR breaks will appear clustered in time. The consequence of the exposure to the series of consecutive VaR breaks (clusters) can be just as problematic as the systematic incomplete reporting on exposure to market risks. The risk of bankruptcy is considerably greater than in the situation in which VaR breaks are evenly distributed over time. Hence, the perfect VaR approach needs to satisfy both properties.

This is why we employed the Christoffersen's conditional coverage test in the paper in addition to the Kupiec's test (Christoffersen, 2001):

(14) [LR.sub.cc] = [LR.sub.UC] + [LR.sub.ind]

where [MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII]

The number of days when after a no VaR break day occurred a no VaR break day ([T.sub.00]), i.e. when after a no VaR break day occurred a VaR break day ([T.sub.01]), i.e. when after a VaR break day occurred a no VaR break day ([T.sub.10]), when after a VaR break day occurred a VaR break day ([T.sub.11]), and their probabilities are given in Table 9 in the Appendix. The Christoffersen's conditional coverage test results for VaR at the 95 and 99 percent confidence levels are presented in Table 10. As can be seen from Table 10, in all of the six emerging markets the approach satisfied the Christoffersen's conditional coverage test at the 5 percent significance level.

A significant disadvantage of these tests is reflected in the fact that they have a questionable statistical power for the sample size defined by the Basle Accord. Both of these tests are developed using asymptotic arguments, which can create difficulties when applied to finite samples. Namely, the [LR.sub.uc] test is asymptotically distributed as [chi square] with one degree of freedom under the null hypothesis that the tail probability (p) is the true probability. The [LR.sub.cc] test is asymptotically distributed as [chi square] with two degrees of freedom under the null hypothesis that the hit sequence is IID Bernoulli with the mean equal to the VaR coverage rate. Asymptotically, that is as the number of observations, T, goes to infinity, the [LR.sub.uc] test will be distributed as a [chi square] with one degree of freedom. It is the same with the [LR.sub.cc] test. In large enough samples, the [LR.sub.cc] test will be distributed as a [chi square] with two degree of freedom. Many authors, such as, Christoffersen and Pelletier (2004), Hurlin et al. (2008), Berkowitz, Christoffersen and Pelletier (2008), Ziggel et al. (2013), have shown that when the number of VaR breaks is small, there are substantial differences between asymptotic probability distributions of the considered tests and their finite sample analogues. Hurlin et al. (2008) state that the use of asymptotic critical values based on a [chi square] distribution induces important size distortions even for relatively large sample. Therefore they point out that in case of a small sample size, (as in sample size defined by Basle Accord), i.e. in case of a small number of VaR breaks ([T.sub.1]), which are the informative observations, it is better to rely on Monte Carlo simulated p-values rather than on those from the [chi square] distribution.

The differences between the finite sample critical values and the asymptotic critical values for both test statistics (the [LR.sub.uc] and [LR.sub.cc]) are shown in Table 11 in the Appendix. The finite sample critical values for the both test statistics for the lower 1 percent are based on 10.000 Monte Carlo simulations of sample size T = 253. The percentages shown in brackets represent quantiles that correspond to the asymptotic critical values under the finite sample distribution. When tests tend to be oversized in finite samples, it means their empirical distributions will be moved to the right off the theoretical shape; hence theoretical quantiles tend to be too small, translating into increased rejection rates. The opposite happens when the tests tend to be undersized in finite samples. In such case their empirical distributions will be moved to the left off the theoretical shape, which will give undersized rejection rates.

Due to the differences between the empirical and theoretical distribution quantiles, conclusions based on the results shown in Tables 7, 8 and 10 need to be checked. The Dufour (2006) Monte Carlo testing technique (see Appendix B) (1) was used for this purpose. Dufour (2006) proposed the Monte Carlo test procedure which allowed to obtain the null distribution of tests statistics in finite sample setting. The method has a great advantage of providing exact tests based on any statistics whose finite sample distribution is intractable but can be simulated (Malecka, 2014).

Following the Monte Carlo test procedure: first, 9.999 samples of random IID Bernoulli (p) variables were generated, where the sample size equals the actual sample. Based on these artificial samples, 9.999 simulated [LR.sub.uc] tests were calculated and named [{[L[??].sub.uc] (i)}.sup.9.999.sub.i=1] . Finally, the simulated p-values were calculated as the share of simulated LRuc values which are larger than the actually obtained [LR.sub.uc] test value:

(15) p-value = 1/10.000 {1 + [9.999.summation over (i=1)] I(L[[??].sub.uc] (i)>[LR.sub.uc])}

where I(*) takes on the value one if the argument is true and zero otherwise.

The same procedure was repeated with [LR.sub.cc] test. The cases for which the tests were not feasible were rejected in the simulation. Average feasible rates of tests are from 0,868 and 0,872 for 99% VaR to 0,974 and 0,987 for 95% VaR, (for both tests) the [LR.sub.uc] and [LR.sub.cc], respectively.

The backtesting results rely on the finite sample p-values, they are shown in table 12 and 13.

Based on the results presented in Table 12 and Table 13 we cannot dispute the use of the approach in the capital markets of EU candidate countries and potential candidate countries and Croatia, in terms of the backtesting rules of the Basel Committee. Particularly good results are gained in meeting the Kupiec's test of unconditional coverage. The explanation lies in the fact that the approach is designed in such a way that it can perfectly capture the dinamics in the series of stock returns.

6. CONCLUSION

Given the characteristics of the emerging markets, such as the capital markets of the EU candidate countries and potential candidate countries and Croatia, in this paper we developed, presented and tested a new hybrid approach based on EVT to estimate the market risk of the portfolio of banks and other financial institutions which were operating in these markets. The approach is designed to successfully capture the dynamics in the series of stock returns with emerging markets and to produce the innovations IID. It is based on the AR(p)-GARCH(1,1) model. At the same time, it recognizes the fact that the extreme returns with emerging markets are more likely than the assumption of normality implies.

The nonparametric part of the model enables us to capture successfully the leptokurtosis and the asymmetry, while the parametric part successfully captures the time changeable volatility. Despite the fact that it was designed to successfully capture the strong dynamics in emerging markets returns, this approach isn't computationally intensive as the other approaches which successfully capture the excess kurtosis and the time-varying volatility, and which are based on too many parameters that need to be estimated. The number of parameters which should be estimated in the model is relatively small.

The Kupiec's test of unconditional coverage and Christoffersen's test of conditional coverage were used for testing the validity of the approach. Despite the fact that the traffic light approach was assigned by Basel Accord, the Kupiec's test of unconditional coverage was chosen, because it is equally important for the bank whether the approach overestimates or underestimates the real level of market risk because in that case it additional capital is unnecessarily allocated which has negative impact on its profitability. The difference between the Kupiec's test and the traffic light approach is in the fact that the Kupiec's test is based on a two-sided test and the traffic light approach is based on a one-sided test. This is why we believe that the Kupiec's test is more suitable for the banks. Christoffersen's test of conditional coverage was chosen because it is a test which is simple and easy to implement but at the same time it tests both features which a perfect VaR approach must satisfy simultaneously (both unconditional coverage and independence).

Since these tests are based on certain asymptotic arguments, conclusions that were reached based on them need to be verified. This is why the Dufour (2006) Monte Carlo testing technique was used. Results of conducted simulations suggest that the VaR forecasts obtained by this approach can be trusted and that this approach can be reliably used in the emerging markets in terms of the Basel Committee's rules. Since in the case of market index BIRS for VaR estimates made for the level of trust of 99% no exceedings were detected during the backtesting period it was not possible to conduct a simulation. This is why future researchers need to test the validity of the approach (once again) on this or on similar markets but for a different (or a longer) backtesting period. As the academic community insists on the use of AR(p)-Student-t-GARCH(1,1) model instead of the normal AR(p)-GARCH(1,1) model, particularly for more extreme (1% or less) VaR thresholds, future researchers are left to test the applicability of the specified (improved) approach.

REFERENCES

Alemany, R., Bolance, C. and Guillen, M. (2012). "Nonparametric estimation of VaR". Working Paper No. 19, XREAP.

Brandolini, D. and Colucci, S. (2013). "Backtesting value-at-risk: a comparison between filtered bootstrap and historical simulation", Journal of Risk Model Validation, 6 (4); 3-16.

Bee, M. (2012). "Dynamic value-at-risk models and the peaks-over-threshold method for market risk measurement: an empirical investigation during a financial crisis", Journal of Risk Model Validation, 2 (2); 3-45.

Beder, S. T (1995). "VaR: Seductive but dangerous", Financial Analysts Journal, 51 (5); 12-24.

Berkowitz, J. and O'Brien, J. (2002). "How Accurate Are Value at Risk", Journal of Finance, 57 (3); 1093-1111.

Berkowitz, J., Christoffersen, P.F. and Pelletier, D. (2008). "Evaluating Value-at-Risk Models with Desk-Level Data", Management Science 57, 2213-2227.

Boudoukh, J., Richardson M., Whitelaw, F.R. (1998). "The Best of Both Worlds: A hybrid Approach to Calculating Value at Risk", Risk, 11 (5); 64-67.

Hurlin, C., Colletaz, G., Tokpavi, S. and Candelon, B. (2008). "Backtesting Value-at-Risk: A GmM Duration-Based Test". Working Paper.

Christoffersen, P.F. (1998). "Evaluating interval forecasts", International Economic Review 39; 841-862.

Christoffersen, P.F., Hahn, J. and Inoue, A. (2001). "Testing and Comparing Value at Risk Measures", Journal of Empirical Finance, 8 (3); 325-342.

Christoffersen, P.F. and Pelletier, D. (2004). "Backtesting Value-at-Risk: A Duration-Based Approach", Journal of Financial Econometrics, 2 (1); 84-108.

Christoffersen, P.F. (2011). Elements of Financial Risk Management. Academic Press.

Corsi, F., Reno, R. (2012). "Discrete-time volatility forecasting with persistent leverage effects and link with continuous-time volatility modelling", Journal of Business & Economic Statistics, 30 (3); 368-380.

Cui, X., Zhu, S., Sun, X. and Li, D. (2013). "Nonlinear portfolio selection using approximate parametric Value-at-Risk", Journal of Banking & Finance 37; 2124-2139.

Del Brio, E. B., Mora-Valencia, A. and Javier, P. (2014). "VaR performance during the subprime and sovereign debt crises: An application to emerging markets", Emerging Markets Review 20; 23-41.

Diamandis, F.P., Anastassios A., Kouretas, G. P. and Zarangas, L. (2011). "Value at Risk for long and short trading positions: Evidence from developed and emerging equity markets", International Review of Financial Analysis 20; 165-176.

Dufour, J.M. (2006). "Monte Carlo Tests with Nuisance Parameters: A General Approach to Finite-Sample Inference and Nonstandard Asymptotics in Econometrics", Journal of Econometrics 133; 443-477.

Engle, R.F. and Manganelli, S. (2004). "CaViaR: conditional autoregressive value at risk by regression quantiles", Journal of Business & Economic Statistics 22; 367-381.

Gencay, R. and Selcuk, F. (2004). "Extreme Value Theory and Value at Risk: Relative Performance in Emerging Markets", International Journal of Forecasting 20; 287-303.

Fernandez, V. (2003). "Extreme Value Theory and value at risk", Revista de Analisis Economico: Economic Analysis Review, 18 (1); 57-85.

Franses, P. and van Dijk, D. (1996). "Forecasting stock market volatility using nonlinear Garch model", Journal od Forecasting, 15 (3); 229-235.

Freedman, D. and Peters, S. (1984). "Bootstrapping a regression equation: Some empirical results", Journal of American Statistical Association 79; 97-106.

Gonzalo, J. and Olmo, J. (2004). "Which extreme values are really extreme?", Journal of Financial Econometrics, 2 (3); 349-369.

Harmantzis, F., Miao, L. and Chien, Y. (2006). "Empirical study of value-at-risk and expected shortfall model with heavy tails", Journal of Risk Finance, 7 (2); 117- 135.

Hull, J. and White, A. (1998). "Incorporating volatility updating into the historical simulation method for value-at-risk", Journal of Risk, 1 (1); 1-19.

Johnson, A.C. (2001). "Value at Risk: Teoria y aplicaciones*4, Estudios de Economia , 28 (2); 217-247.

Jorion, P. (2006). Value at Risk - The New Benchmark for Managing Financial Risk. McGraw Hill.

Kupiec, P.H. (2005). "Techniques for verifying the accuracy of risk measurement models", Journal of Derivatives 3, 73-84.

Schittenkopf, C., Lehar, L. and Scheicher, M. (2002). "GARCH vs. stochastic volatility: Option pricing and risk management", Journal of Banking & Finance 26; 323-345.

Lucas, A. (2000). "A note on optimal estimation from a risk management perspective under possibly misspecified tail behavior", Journal of Business and Economic Statistics, 18 (1); 31-39.

Louzis, P.D., Xanthopoulos-Sisinis, S. and Refenes, A.P. (2014). "Realized volatility models and alternative Value-at-Risk prediction strategies", Economic Modelling 40; 101-116.

Louzis, P.D., Xanthopoulos-Sisinis, S. and Refenes, A.P. (2012). "Stock index realized volatility forecasting in the presence of heterogeneous leverage effects and long range dependence in the volatility of realized volatility", Appl. Econ. 44; 3533-3550.

Malecka, M. (2014). "Duration-Based Approach to VaR Independence Backtesting", Statistics in Transition, 15 (4); 627-636.

McNeil, A. and Frey, R. (2000). "Estimation of Tail Related Risk Measure for Heteroscedastic Financial Time Series: An Extreme Value Approach", Journal of Empirical Finance, 7 (3-4); 271-300.

Rossignolo, F.A., Fethib, M. D. and Shaban, M. (2012). "Value-at-Risk models and Basel capital charges Evidence from Emerging and Frontier stock markets", Journal of Financial Stability 8; 303-319.

Rossignolo, F.A., Fethib, M. D. and Shaban, M. (2013). "Market crises and Basel capital requirements: Could Basel III have been different? Evidence from Portugal, Ireland, Greece and Spain (PIGS)", Journal of Banking & Finance 37; 1323-1339.

Samanta, G.P., Jana, P., Hait, A. and Kumar, V. (2010). "Measuring Market Risk --An Application of Value-at-risk to Select Government Bonds in India", Reserve Bank of India Occasional Papers, 31 (1): 2-32.

Cener, E., Baronyana, S. and Menguturk, L. (2012). "Ranking the predictive performances of value-at-risk estimation methods", International Journal of Forecasting 28; 849-873.

Zikovic, S. (2005). "VTWHS (Volatility And Time Weighted Historical Simulation) Model Methodology For Measuring Market Risks", A Bank Vipa, Risk Management Department 10; 45-64.

Zikovic, S. and Prohaska Z. (2009). Optimisation of decay factor in time weighted simulation: VaR Performance in Mediterranen countries. 5th International Finance Conference (IFC5), University of Cergy and ISP Paris, Hammament, Tunisia, 12-14 March.

Zikovic, S. (2010). Market Risk in Transitions Countries: Value at Risk Approach. University of Rijeka.

Ziggel, D., Berens, T., WeiB, G.N.F. and Wied, D. (2013). "A New Set of Improved Value-at-Risk Backtests". Working Paper No. 29, SFB.

Wong, Y.W., Cheng, M. and Wong, C.Y. P. (2003). "Market risk management of banks: Implications from the accuracy of VaR forecasts", Journal of Forecasting, 22 (1): 22-33.
APPENDIX A

TABLE 3 ACF, PACF AND LJUNG-BOX Q TEST FOR THE DAILY LOG RETURNS FOR
TESTED INDEXES IN THE PERIOD 02.02.2009-02.02.2012.

               BIRS index

lag    ACF      PACF     Q-stat.    p-value

1     0,1297   0,1297    12,7090    0,0004
5     0,0147   0,0113    18,6975    0,0022
10    0,0173   0,0135    26,8644    0,0027
15    -0,0221  -0,0415   35,3923    0,0022
20    0,0407   0,0219    37,3823    0,0105

               BELEXline index

lag    ACF      PACF     Q-stat.    p-value

1     0,3440   0,3440    89,4756    0,0000
5     0,0743   0,0060    131,8199   0,0000
10    0,0291   -0,0236   145,8258   0,0000
15    0,0095   -0,0468   171,2537   0,0000
20    0,1088   0,0491    200,5026   0,0000

                MONEX20 index

lag     ACF      PACF     Q-stat.   p-value

1     0,1843    0,1843    25,6828   0,0000
5     0,0407    0,0004    58,4502   0,0000
10    0,0817    0,0685    70,9771   0,0000
15    0,0719    0,0621    83,2689   0,0000
20    0,0143    0,0085    84,1947   0,0000

                CROBEX index

lag     ACF      PACF     Q-stat.   p-value

1     0,1214    0,1214    11,1445   0,0008
5     -0,0078   -0,0138   12,4370   0,0293
10    0,0119    0,0077    14,7255   0,1424
15    0,0740    0,0688    22,3823   0,0982
20    -0,0391   -0,0555   34,0363   0,0259

                MB 110 index

lag     ACF      PACF     Q-stat.   p-value

1     0,1170    0,1170    10,3227   0,0013
5     0,0513    0,0533    19,2940   0,0017
10    0,0954    0,0936    32,0580   0,0004
15    0,0177    0,0097    45,3945   0,0001
20    0,0045    -0,0084   47,6616   0,0005

                XU 100 index

lag     ACF      PACF     Q-stat.   p-value

1     -0,1123   -0,1123   9,5369    0,0020
5     0,0221    0,0101    12,4585   0,0290
10    -0,0120   -0,0249   16,4064   0,0886
15    0,0322    0,0317    21,7823   0,1136
20    -0,0516   -0,0194   35,9534   0,0156

Source: Authors' calculations.

TABLE 3A

ACF, PACF AND LJUNG-BOX Q TEST FOR THE DAILY LOG SQUARED RETURNS FOR
TESTED INDEXES IN THE PERIOD 02.02.2009-02.02.2012.

                BIRS index

lag     ACF      PACF     Q-stat.    p-value

1     0,0647    0,0647     3,1683    0,0751
5     0,0927    0,0884    14,0959    0,0150
10    0,0389    0,0362    16,0515    0,0982
15    0,0908    0,0671    47,6294    0,0000
20    -0,0157   -0,0238   55,0095    0,0000

                BELEXline index

lag     ACF      PACF     Q-stat.    p-value

1     0,4585    0,4585    158,9550   0,0000
5     0,0936    0,0355    284,7290   0,0000
10    0,0762    -0,0160   334,5536   0,0000
15    0,0757    0,0530    351,0855   0,0000
20    0,1461    0,0902    386,6475   0,0000

                MONEX20 index

lag     ACF      PACF     Q-stat.    p-value

1     0,3132    0,3132    74,1755    0,0000
5     0,2753    0,1925    229,9905   0,0000
10    0,1328    0,0524    290,9288   0,0000
15    0,0369    -0,0315   313,5787   0,0000
20    -0,0072   -0,0025   314,7851   0,0000

                CROBEX index

lag     ACF      PACF     Q-stat.    p-value

1     0,1033    0,1033     8,0741    0,0045
5     0,1584    0,0951    91,0099    0,0000
10    0,1853    0,1256    154,9989   0,0000
15    0,0994    0,0144    191,3901   0,0000
20    0,1193    0,0286    252,0330   0,0000

                MB 110 index

lag     ACF      PACF     Q-stat.    p-value

1     0,1173    0,1173    10,3980    0,0013
5     0,1756    0,1529    78,3206    0,0000
10    0,0502    0,0027    148,9301   0,0000
15    0,0590    0,0736    202,7552   0,0000
20    0,0203    0,0325    216,8705   0,0000

                XU 100 index

lag     ACF      PACF     Q-stat.    p-value

1     0,1863    0,1863    26,2504    0,0000
5     -0,0015   -0,0221   36,7695    0,0000
10    0,0181    0,0164    42,8916    0,0000
15    -0,0459   -0,0455   45,0271    0,0001
20    -0,0092   -0,0297   52,4803    0,0001

Source: Authors' calculations.

TABLE 9 THE HIT SEQUENCE OF VAR BREAKS

                 BIRS    M0NEX2U   MBI10    BELEXline   CR0BEX

95%VaR

[T.sub.0]          244       241      243         243      246
[T.sub.1]            9        12      10           11        7
[T.sub.00]         235       229      235         232      239
[T.sub.01]           9        12        8          10        7
[T.sub.10]           9        12        8          10        7
[T.sub.11]           0         0        2           1        0
P               0,0356   0,0474    0,0395     0,0433    0,0277
[P.sub.01]      0,0369   0,0498    0,0329     0,0412    0,0285
[p.sub.11]           0         0     0,2      0,0909         0

                 BIRS    M0NEX20   MBI10    BELEXline   CR0BEX

99%VaR

[T.sub.0]          253       252      249         251      250
[T.sub.1]            0         1        3           1        2
[T.sub.00]         253       251      247         250      248
[T.sub.01]           0         1        2           1        2
[T.sub.10]           0         1        2           1        2
[T.sub.11]           0         0        1           0        0
[pi]                 0    0,004    0,0119     0,0040    0,0079
[[pi].sub.01]        0    0,004    0,0080     0,0040    0,0080
[[pi].sub.11]       /          0   0,3333           0        0

                XU100

95%VaR

[T.sub.0]          243
[T.sub.1]          10
[T.sub.00]         233
[T.sub.01]         10
[T.sub.10]         10
[T.sub.11]           0
P               0,0395
[P.sub.01]      0,0412
[p.sub.11]           0

                XU100

99%VaR

[T.sub.0]          251
[T.sub.1]            1
[T.sub.00]         250
[T.sub.01]           1
[T.sub.10]           1
[T.sub.11]           0
[pi]            0,0040
[[pi].sub.01]   0,0040
[[pi].sub.11]        0

Source: Authors' calculations.

TABLE 11
THE DIFFERENCES BETWEEN THE FINITE SAMPLE CRITICAL VALUES AND THE
ASYMPTOTIC CRITICAL VALUES FOR THE LRUC AND THE LRCC TEST STATISTICS

Significance levels
                               1%           5%          10%

                                       [LR.sub.uc]
                                        Statistic

Asymptotic [chi square](1)   66,348       38,414       27,055
Finite-sample                 5,497       5,025        3,555

                                       [LR.sub.cc]
                                        Statistic

                             (0,49%)     (9,49%)      (12,19%)
Asymptotic [chi square](2)    9,21        59,915       4,605
Finite-sample                 6,007       5,015        5,005
                             (0,20%)     (1,10%)      (11,79%)

Note: The finite sample critical values for the both test statistics
for the lower 1 percent are based on 10.000 Monte Carlo simulations
of sample size T = 253. The percentages shown in the brackets
represent quantiles that correspond to the asymptotic critical values
under the finite sample distribution.


APPENDIX B:

The Dufour (2006) Monte Carlo Testing Technique

Lets take (S) a statistic of a given test of continuous survival function G(.) such as Prob [[S.sub.i] = [S.sub.j]] = 0. Theoretical p-value G(.) can be approximated by its empirical counterpart: [[??].sub.M](x) = 1 /M [M.summation over (i=1)] I([S.sub.i] [greater than or equal to] x) where I(.) is the indicator function. ([S.sub.i]) is the test statistic for a sample simulated under the null hypothesis. Dufour (2006) demonstrated that if (M) is big enough, whatever the value of (S0), theoretical critical region G([S.sub.0]) < [alpha], with (a), the asymptotic nominal size, is equivalent to the critical regio[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII]or when it is possible for a given simulation of the test statistic (under null hypothesis) to find the same value of (S) for two or more times, the empirical survival function can be written as follows:

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII]

0,1 ..., M correspond to realizations of a uniform [0,1] variable.

NIKOLA RADIVOJEVIC **

MILENA CVJETKOVIC ***

SASA STEPANOV ****

* The authors would like to thank the referees and the editor of the journal Estudios de Economia for help, advice and useful comments which have considerably improved the manuscript.

** Technical college of applied studies, Kragujevac, Serbia. E-mail: radivojevic034@gmail. com

*** Technical faculty, University of Novi Sad, Zrenjanin, Serbia. E-mail: caca034@yahoo. com

**** bAS, Belgrade, Serbia. E-mail: sasa.stepanov@gmail.com

(1) See the advantages of applying simulation procedure over bootstrap method in Christoffersen and Pelletier (2004).
TABLE 1 DESCRIPTIVE STATISTICS OF SELECTED EMERGING MARKETS

             BIRS     MONEX20    MBI10    BELEXline   CROBEX

Mean        -0,0159   0,0008    -0,0080     0,0031    0,0073

Standard    0,7492    1,4375    1,5478      1,0558    1,3245
Dev.

Sample      0,5613    2,0665    2,3958      1,1146    1,7544
Variance

Kurtosis    5,2436    7,6201    9,2880      3,7825    6,7776

Skewness    0,1009    1,1441    0,7756      0,4264    0,4836

Jarque-      159,2       834    1312,5          42    477,1
Bera test

p--value    0,0000    0,0000    0,0000      0,0000    0,0000

             XU100

Mean        0,0118

Standard    1,1840
Dev.

Sample      1,4019
Variance

Kurtosis    2,4708

Skewness    -0,1828

Jarque-       12,9
Bera test

p--value    0,0015

Source: Authors' calculations.

TABLE 2 THE ARCH EFFECTS

              BIRS    MONEX20   MBI10    BELEXline   CROBEX   XU100

Lagrange     24,443   106,861   26,997    218,104    28,789   84,656
Multiplier
p-value      0,000     0,000    0,000      0,000     0,000    0,000

Source: Authors' calculations.

TABLE 4 THE ESTIMATES OF OF AR(P)-GARCH(1,1) MODEL PARAMETERS

              BIRS    MONEX20   MBI10    BELEXline   CROBEX
Parameters

C              --       --        --        --         --
AR(p)        0,1727   0,2986    0,1280    0,3896     0,1280
a            0,1666   0,1813    0,1786    0,2955     0,2088
b            0,4367   0,7656    0,8200    0,6219     0,7639
w            0,0000   0,0000    0,0000    0,0000     0,0000

              XU100
Parameters

C            0,0003
AR(p)        -0,0145
a            0,1359
b            0,0160
w            0,0002

Source: Authors'calculations.

TABLE 5 THE MAXIMUM LIKELIHOOD ESTIMATES OF THE TAIL INDEX AND SIGMA,
THRESHOLD VALUE

                  BIRS     MONEX20    MBI10    BELEXline   CROBEX

Parameters       -1,3925   -2,1152   -1,5637    -1,4085    -1,3003
Threshold
value

The tail index   0,4632    0,5227    0,3801     0,3342     0,4010
([xi])

sigma            1,6978    2,7989    0,9137     1,6978     2,1843

                  XU100

Parameters       -1,6146
Threshold
value

The tail index   0,3257
([xi])

sigma            0,8985

Source: Authors'calculations.

TABLE 6 THE NUMBER/PERCENTAGES OF VAR BREAKS AT A 95% AND 99%
CONFIDENCE LEVELS

Stock index            BIRS   MONEX20   MBI10   BELEXline   CROBEX

No. of 95%VaR breaks     9        12      10          11        7
Percent of breaks      3,56     4,74    3,95        4,35     2,77
95%VaR breaks
No. of 99%VaR breaks     0         1       3           1        2
Percent of breaks        0       0,4    1,19         0,4     0,79
99%VaR breaks

Stock index            XU100

No. of 95%VaR breaks     10
Percent of breaks      3,95
95%VaR breaks
No. of 99%VaR breaks      1
Percent of breaks       0,4
99%VaR breaks

Source: Authors' calculations.

TABLE 7 KUPIEC'S TEST BACKTESTING RESULTS AT 5% SIGNIFICANCE LEVEL
FOR VAR AT THE 95% CONFIDENCE LEVEL

Stock index       BIRS    MONEX2C   MBI10    BELEXline   CROBEX

Kupiec's test    1,2274   0,0357    0,6277    0,2504     3,1473
([LR.sub.uc])
p-value          0,2679   0,8500    0,4282    0,6168     0,0761

Stock index      XU100

Kupiec's test    0,6277
([LR.sub.uc])
p-value          0,4282

Source: Authors' calculations.

TABLE 8 KUPIEC'S TEST BACKTESTING RESULTS AT 5% SIGNIFICANCE LEVEL
FOR VAR AT THE 99% CONFIDENCE LEVEL

Stock index      BIRS   MONEX20   MBI10    BELEXline   CROBEX

Kupiec's test     --    1,2129    0,0823    1,2129     0,1208
([LR.sub.uc])

p-value           --    0,2708    0,773     0,2708     0,7281

Stock index      XU100

Kupiec's test    1,2129
([LR.sub.uc])

p-value          0,2708

Source: Authors' calculations.

TABLE 10 CHRISTOFFERSEN'S CONDITIONAL COVERAGE TEST BACKTESTING
RESULTS AT 5% SIGNIFICANCE LEVEL FOR VAR AT THE 95% AND 99%
CONFIDENCE LEVELS

Stock index         BIRS    MONEX20   MBI10    BELEXline   CROBEX

Christoffersen's   1,2274   0,0357    4,4835    0,8245     3,1473
conditional
coverage test
for 95%VaR
([LR.sub.cc])

p-value            0,5414   0,9823    0,1063    0,6622     0,2073

Christoffersen's     --     1,2022    5,4488    1,2022     0,1166
conditional
coverage test
for 99%VaR
([LR.sub.cc])

p-value              --     0,5482    0,0656    0,5482     0,9433

Stock index        XU100

Christoffersen's   0,6277
conditional
coverage test
for 95%VaR
([LR.sub.cc])

p-value            0,7306

Christoffersen's   1,2022
conditional
coverage test
for 99%VaR
([LR.sub.cc])

p-value            0,5482

Notice: In the cases where the sample has [T.sub.11] = 0 (there are
no consecutive VaR breaks), an alternative formula was used in the
paper to calculate the first-order Markov likelihood (see Brandolini
and Colucci, 2013).

Source: Authors' calculations.

TABLE 12 THE BACKTESTING RESULTS BASED ON THE MONTE CARLO P-VALUES
FOR THE [LR.sub.UC] TEST

                             95%VaR

           BIRS    MONEX20   MBI10    BELEXline   CROBEX   XU100

p-value   0,1048   0,4499    0,3274    0,4715     0,1582   0,2714

                             99%VaR

           BIRS    MONEX20   MBI10    BELEXline   CROBEX   XU100

p-value     --     0,1798    0,5910    0,1022     0,3567   0,0987

Notice: Significante level of 5%. Samples where the test cannot be
computed are omitted due to lack of VaR breaks.

Source: Authors' calculations.

TABLE 13 THE BACKTESTING RESULTS BASED ON THE MONTE CARLO P-VALUES
FOR THE [LR.SUB.CC] TEST

                                 95%VaR

           BIRS    MONEX20   MBI10     BELEXline   CROBEX   XU100

p-value   0,3531   0,6221    0,0892     0,3542     0,1735   0,4115

                                 99%VaR

           BIRS    MONEX20   MBI10     BELEXline   CROBEX   XU100

p-value     --     0,4588    0,0549     0,2588     0,1895   0,2588

Notice: Significante level of 5%. Samples where the test cannot be
computed are omitted due to lack of VaR breaks.

Source: Authors' calculations.
COPYRIGHT 2016 Universidad de Chile
No portion of this article can be reproduced without the express written permission from the copyright holder.
Copyright 2016 Gale, Cengage Learning. All rights reserved.

Article Details
Printer friendly Cite/link Email Feedback
Author:Radivojevi, Nikola; Cvjetkovi, Milena; Stepanov, SaUa
Publication:Estudios de Economia
Article Type:Ensayo
Date:Jun 1, 2016
Words:9795
Previous Article:La relacion entre apalancamiento y poder de negociacion de los sindicatos: Evidencia teorica y empirica.
Topics:

Terms of use | Privacy policy | Copyright © 2019 Farlex, Inc. | Feedback | For webmasters