“APPLYING HYBRID APPROACH TO CALCULATING VAR IN CROATIA“

M.Sc. Saša Žiković Faculty of Economics University of Rijeka/Assistant Ivana Filipovića 4, Rijeka, Croatia Phone: +385 51 355 111 Fax: +385 51 212 268 E-mail: sasa.zikovic@efri.hr Key words: Value at Risk, historical simulation, hybrid approach, Croatia

1. INTRODUCTION The most prominent of risks present in trading operations of a bank is the market risk, since it reflects the potential economic loss caused by the decrease in the market value of a bank’s portfolio of securities. According to the 1996 Market Risk Amendment to the Basel Accord, besides using the standardized approach, banks can set their capital requirements for market risk of their trading positions based on the ten-day 1-percent VaR. The 1996 Amendment allows ten-day 1percent VaR to be measured as a multiple of one-day 1-percent VaR by using a simplistic square root of time rule1. Although VaR is a conceptually simple measure of risk, computing VaR in practice can be very difficult due to a simple reason that position risk of a bank's portfolio depends on the joint distribution of all of the securities composing that particular portfolio. Fortunately, portfolio level risk measurement requires only a univariate, portfoliolevel model2 thus drastically reducing the computational burden of multivariate models. If interest centres on the distribution of the portfolio returns, then this distribution can be modelled directly from portfolio returns rather than via aggregation based on a larger and almost inevitably less-well-specified multivariate model. In this paper the author examines the theoretical background of two nonparametric approaches to calculating VaR, historical simulation and hybrid approach developed by Boudoukh, Richardson and Whitelaw, and examines their performance in a transitional capital market such as Republic of Croatia. The paper evaluates and analyses the out-of-sample forecasting accuracy of both methods on two Croatian indexes, CROBEX – the official index of Zagreb Stock Exchange and VIN - the official index of Varazdin Stock Exchange. In section 2 of the paper, an overview of Value at Risk as a methodology for measuring market risk is presented. Section 3, present the methodology of calculating VaR via historical simulation. In section 4, a hybrid nonparametric approach to calculating VaR is presented and its advantages and disadvantages are discussed. Section 5 analyses the characteristics of Croatian financial market. In section 6, out-of-the-sample performance evaluation of historical simulation and hybrid BRW approach, with different observation windows and decay factors is performed on two Croatian stock indexes. The conclusions are summarized in Section 7.

1

2

Basel Committee on Banking Supervision (1996): ‘Amendment to the Capital Accord to incorporate Market Risks. Bank for International settlements’, BIS, January 1996 Benson, P. and Zangari, P. (1997): ‘A general approach to calculating VaR without volatilities and correlations’, RiskMetrics Monitor, (second quarter 1997): 19-23

2. MEASURING MARKET RISK USING VAR METHODOLOGY Value at Risk (VaR) has become the standard measure that financial analysts use to quantify market risk. It is defined as the maximum potential loss in value of a portfolio of financial instruments with a given probability over a certain horizon. It indicates how much can a bank expect to lose with probability C for a given time horizon3. The main advantage of VaR as a risk measure over other risk measures is that it is theoretically simple. VaR can be used to summarize the risk of individual positions, or a risk of large portfolio of an internationally active bank. VaR reduces the risk associated with any portfolio to just one number, the expected loss associated with a given probability over a defined holding period. VaR for a given probability C can be expressed as: VaRc = F-1(C) where F–1(C) denotes the inverse of cumulative probability distribution of the changes in the market value of a portfolio. Thus, losses greater than the estimated VaR should only occur with the probability 1-C. For example, if a VaR calculated at the C% confidence level is accurate, then losses greater than the VaR measure, so-called “tail events”, should on average only occur C*N times in every N trading days. While VaR is a very easy and intuitive concept, its measurement is a very challenging statistical problem. Although the existing models for calculating VaR employ different methodologies, some are entire parametric, while other are semi parametric or nonparametric in their nature, they all follow a common general structure: 1) Marking to market of a portfolio of securities, 2) Estimation of distribution of portfolio returns (using either parametric or nonparametric methods), 3) Calculating the VaR of the portfolio. The main difference among numerous VaR methods is related to the estimation of distribution that adequately describes the portfolio returns, i.e. the way the problem of how to estimate the possible changes in the value of the portfolio is dealt with. Research papers dealing with VaR calculation or volatility forecasting in the financial markets of Croatia or EU new member states are very rare. Croatia as an EU member candidate state offers large profits for investors, and represents a very interesting opportunity for foreign investors, primarily banks. Most of the banking sector in Croatia, about 92%, is foreign owned, and banks that operate in Croatia employ the same risk measurement models in forming of provisions as they do in developed markets. This means that risk managers in Croatian banks de facto presume similar or even equal market characteristics and behaviour in Croatian market, as they would expect in developed markets. This is a dangerous assumption, which is not realistic. Employing VaR models in forming of bank’s provisions that are not suited to financial markets that they are used on, can have serious consequences for banks, resulting in significant losses in trading portfolio that could pass undetected by the employed risk measurement models, leaving the banks unprepared for such events.

3

Jorion, P. (2001): Value at Risk, The New Benchmark for Managing Financial Risk, Second Edition, New York: McGraw Hill

Most of the widespread VaR methods used in the financial industry do not capture the following characteristics of financial markets that are known since the pioneering work of Mandelbrot (1963)4: 1. Equity returns are usually skewed to the left. 2. Distribution of financial return is leptokurtotic, i.e. it has fatter tails and a higher peak than described under Gaussian distribution. 3. There is significant autocorrelation in financial time series, i.e. periods of high and low volatility tend to cluster. The last assumption is a very important characteristic of financial returns, since it means that market volatility can be consider as being quasistable, subject to change in longer periods, but stable in the short run. This gives some justification to the assumption that is the basis for most of the VaR models (historical simulation among them), that recent past will be similar to near future. 3. MEASURING VAR USING HISTORICAL SIMULATION One of the most common methods for VaR estimation is the historical simulation (HS VaR). This approach drastically simplifies the procedure for computing VaR, since it doesn’t make any direct distributional assumption about portfolio returns. Banks often rely on VaR figures calculated by historical simulations. The value of VaR is calculated as the 100C’th percentile or the (T+1)C’th order statistic of the set of pseudo portfolio returns. In principle it is easy to construct a time series of historical portfolio returns using current portfolio holdings and historical asset returns. The returns on the indexes in this paper are calculated as: P rt = ln(1 + Rt ) = ln t Pt −1
The problem is that historical asset prices for the assets held at some point in time may not be available. In such a case “pseudo” historical prices must be constructed using either pricing models, factor models or some ad hoc consideration. The current assets without historical prices can for example be matched to “similar” assets by capitalization, industry, leverage, and duration. Historical pseudo asset prices and returns can then be constructed using the historical prices on these substitute assets: rw,t = ∑ wi ,T ri ,t ≡ W 'T Rt , t = 1, 2,…, T
i =1 N

Historical simulation VaR can than be expressed as:
C HS − VaRT +1|T ≡ rw ((T + 1)C )

where rw ((T + 1)C ) is taken from the set of ordered pseudo returns {rw (1), rw (2),..., rw (T )}. If (T+1)C is not an integer value then the two adjacent observations can be interpolated to
4

Mandelbrot, B. (1963): 'The Variation of Certain Speculative Prices', The Journal of Business, Vol. 36, No. 4, (October 1963): 394-419

calculate the VaR. Historical simulation has some serious problems, which have been well documented5. Perhaps most importantly, it does not properly incorporate conditionality into the VaR forecast. The only source of dynamics in the HS VaR is the fact that the observation window is updated with the passing of time. This source of conditionality is minor in practice. Historical Simulation is based on the concept of rolling windows. The process of calculating VaR by historical simulation begins by choosing a length of the window of observations, which usually ranges from two months to two years. Calculated portfolio returns within the observation window are sorted and the desired quantile is given by the return xi that satisfies the condition that j out of n observations do not exceed it. The probability that j out of n observations do not exceed some fixed value of observed variable x follows a binomial distribution: n P{ j _ observations ≤ x} =  j {F ( x)} j {1 − F ( x)}n − j   It follows that the probability of at least i observations in the selected sample do not exceed x also follows a binomial distribution:
n n Gi ( x) = ∑  j {F ( x)} j {1 − F ( x)}n − j j =i  

Gi(x) is the distribution function of the order statistic and thus also of the VaR. To compute the VaR the following day, the whole window is moved forward by one observation and the entire procedure is repeated. Historical simulation method assigns equal probability weight of 1/N to each observation. This means that the historical simulation estimate of VaR at the C confidence level corresponds to the N(1-C) lowest return in the N period rolling sample. Because a crash is the lowest return in the N period sample, the N(1-C) lowest return after the crash, turns out to be the (N(1-C)-1) lowest return before the crash. If the N(1-C) and (N(1-C)-1) lowest returns happen to be very close in magnitude, the crash actually has almost no impact on the historical simulation estimate of VaR for the long positions in a portfolio of securities. From the equation for historical simulation it can be seen that HS VaR changes significantly only if the observations around the order statistic rw ((T + 1)C ) change significantly. Although historical simulation makes no explicit assumptions about the distribution of portfolio returns, an implicit assumption is hidden behind the procedure: the distribution of portfolio returns doesn’t change within the window. From this implicit assumption several problems may arise in using this method in practice. From the assumption that all the returns within the observation window used in historical simulation have the same distribution, it follows that all the returns of the time series also have the same distribution: if yt-window,...,yt and yt+1-window,...,yt+1 are independently and identically distributed (IID), then also yt+1 and ytwindow has to be IID, by the transitive property. Another serious problem of the historical simulation is the fact that for the empirical quantile estimator to be consistent, the size of observation window must go to infinity. The length of the observation window hides another serious problem. Forecasts of VaR under historical simulation are meaningful only if the
5

Pritsker, M. (2001): ‘The Hidden Dangers of Historical Simulation’, Board of Governors of the Federal Reserve System, Economics Discussion Series, Working paper, No. 27, April 2001.

historical data used in the calculations have the same distribution. The length of the window must satisfy two contradictory properties: it must be large enough, in order to make statistical inference significant, and it must not be too large, to avoid the risk of taking observations outside of the current volatility cluster. Clearly, there is no easy solution to this problem.6 If the market is moving from a period of low volatility to a period of high volatility, VaR forecasts based on the historical simulation will under predict the true risk of a position since it will take some time before the observations from the low volatility period leave the observation window. Finally, VaR forecasts based on historical simulation may present predictable jumps, due to the discreteness of extreme returns. If VaR of a portfolio is computed using a rolling window of N days and that today’s return is a large negative number, it is easy to predict that the VaR estimate will jump upward, because of today’s observation. The same effect (reversed) will reappear exactly after N days, when the large observation drops out of the observation window.

4. EXPONENTIAL MODIFICATION OF HISTORICAL SIMULATION
When relaxing the assumption that returns are IID, it might be reasonable to assume that simulated returns from the recent past better represent today portfolio's risk than returns from the distant past. Boudoukh, Richardson, and Whitelaw, BRW hereafter, used this idea to introduce a generalization of the historical simulation and assign a relatively higher amount of probability weight to returns from the more recent past.7 The BRW approach combines RiskMetrics and historical simulation methodologies, by applying exponentially declining weights to past returns of the portfolio. Each of the most recent N returns of the portfolio, yt, yt-1, ..., yt-N+1, is associated a weight, 1− λ  1− λ   1 − λ  N −1 respectively8. After the probability weights are assigned, λ ,...,  , λ N  N  N  1− λ  1− λ  1− λ  VaR is calculated based on the empirical cumulative distribution function of returns with the modified probability weights. The basic historical simulation method can be considered as a special case of the more general BRW method in which the decay factor - λ is set equal to 1. The BRW method involves a simple modification of the historical simulation. However, the modification makes a large difference. The most recent return in the BRW methods receives probability weights of just over 1% for λ = 0,99 and of just over 3% for λ = 0,97. In both cases, this means that if the most recent observation is the worst loss of the N days, then it will be the VaR estimate at the 1% confidence level. Hence, the BRW methods appear to remedy the main problems with the historical simulation methods because very large losses are immediately reflected. The simplest way to implement BRW's approach is to construct a history of N hypothetical returns that the portfolio would have earned if held for each of the previous N days, rt-1,…, rt-N and then assign exponentially declining probability weights wt6

Manganelli, S. and Engle F.R. (2001): ‘Value at Risk models in Finance’, ECB working paper series, Working paper, No. 75, August 2001 7 Boudoukh, J., Richardson, M. and Whitelaw, R. (1998): ‘The best of both worlds’, Risk, Vol. 11, No. 5, (May 1998): 64-67 1− λ 8 The role of the term is to ensure that the weights sum to 1. 1 − λN

wt-N to the return series9. Given the probability weights, VaR at the C percent confidence level can be approximated from G(.; t;N), the empirical cumulative distribution function of r based on return observations rt-1,…, rt-N.
1,…,

G ( x; t , N ) = ∑ 1{rt−i ≤ x} wt −i
i =1

N

Because the empirical cumulative distribution function, unless smoothed, for example via kernel smoothing as suggested by Butler and Schachter (1998)10, is discrete, the solution for VaR at the C percent confidence level will typically not correspond to a particular return from the return history. Instead, the BRW solution for VaR at the C percent confidence level can be between a return that has a cumulative distribution that is less than C, and one which has a cumulative distribution that is more than C. These returns can be used as estimates of the BRW method's VaR estimates at confidence level C. The estimate that understates the BRW estimate of VaR at the C percent confidence level (upper limit) is given by11:
BRW u (t | λ , N , C ) = inf(r ∈ {rt −1 ,...rt −1− N } | G (r ; t , N ) ≥ C ) and the estimator of lower limit is given by: BRW o (t | λ , N , C ) = sup( r ∈ {rt −1 ,...rt −1− N } | G (r ; t , N ) ≤ C ) where λ is the exponential weight factor, N is the length of the history of returns used to compute VaR, and C is the VaR confidence level. BRW u (t | λ , N , C ) is the lowest return of the N observations whose empirical cumulative probability is greater than C, and BRW o (t | λ , N , C ) is the highest return whose empirical cumulative probability is less than C. The main issue in evaluation of BRW based VaR, as a risk measure, is the extent to which VaR forecasts based on the BRW method respond to changes in the underlying risk factors. It is important to know under what circumstances risk estimates increase when using the BRW u (t | λ , N , C ) estimator. The result is provided in the following proposition: If rt > BRW u (t , λ , N ) then BRW u (t + 1, λ , N ) ≥ BRW u (t , λ , N ) . When the VaR estimate using the BRW method is estimated for returns during time period t+1, the return at time t−N is dropped from the sample, the return at time t receives weight 1− λ and the weight on all other returns are λ times their earlier values. 1 − λN

Consequently, r(C) is defined as:
9

The weights sum to 1 and are exponentially declining at rate λ (0 < λ ≤ 1)
N t −i

∑w
i =1
10

=1

wt −i −1 = λwt −i
Butler, J.S. and Schachter, B. (1998): ‘Estimating Value-at-Risk with a Precision Measure by Combining Kernel Estimation with Historical Simulation’, Review of Derivatives Research, No. 1, (1998): 371-390 11 Pritsker, M. (2001)

r (C ) = {rt −1 , i = 1,...N | G (rt −1 ; t , N ) ≤ C}

To verify the proposition, it suffices to examine how much probability weight the VaR estimate at time t+1 places below BRW u (t , λ , N ) . There are two cases to consider: Case 1: rt − N ∉ r (C ) In this case, since by assumption, rt ∉ r (C ) then
G ( BRW u (t , λ , N ); t + 1, λ , N ) < λG ( BRW u (t , λ , N )) . Therefore, BRW u (t + 1, λ , N ) = inf(r ∈ {rt ,...rt −1− N } | G (r ; t + 1, λ , N ) ≥ C ) ≥ BRW u (t , λ , N )

Case 2: rt − N ∈ r (C ) . In this case, since rt ∈ r (C ) by assumption, then
G ( BRW o (t , λ , N ); t + 1, λ , N ) < λG ( BRW o (t , λ , N ))

Therefore, BRW o (t + 1, λ , N ) = sup( r ∈ {rt ,...rt −1− N } | G (r ; t + 1, λ , N ) ≤ C ) ≤ BRW o (t , λ , N ) The proposition shows that when losses at time t are bounded below the BRW VaR estimate at time t, the BRW VaR estimate for time t+1 will indicate that risk at time t+1 is no greater than it was at time t. To understand the importance of this proposition, it suffices to examine the case when today's BRW VaR estimate for tomorrow's return is conditionally correct, but that risk changes with returns, so that tomorrow's return will influence risk for the day after tomorrow. Under these circumstances, an important question is what is the probability that a VaR estimate that is correct today will increase tomorrow. The answer provided by the proposition is that tomorrow's VaR estimate will not increase with probability 1−C. So, for example, if C is equal to 1%, then a VaR estimate that is correct today will not increase tomorrow with probability 99%. Although the BRW approach suffers from the explained logical inconsistency, this approach still represents a significant improvement over the historical simulation, since it drastically simplifies the assumptions needed in the parametric models and it incorporates a more flexible specification than the historical simulation approach. To better understand the assumptions behind the BRW approach and its connection to historical simulation, BRW quantile estimator can be expressed as12:

ˆ qt +1,C =

j = t − N +1

∑ y I (∑
t j

N

i =1

f i (λ ; N ) I ( yt +1−i ≤ y j ) = C

)

where f i (λ ; N ) are the weights associated with return yi and I(●) is the indicator function. If
f i (λ ; N ) = 1 / N BRW quantile estimator equals the historical simulation estimator. The main difference between BRW approach and historical simulation is in the specification of the quantile process. With historical simulation each return is given the same weight, while with the BRW approach returns have different weights, depending on how old the observations are. Strictly speaking, none of these models is nonparametric, since a parametric specification is proposed for the quantile. Boudoukh, Richardson and Whitelaw in their original paper set λ

12

Manganelli, S. and Engle F.R. (2001)

equal to 0,97 and 0,99, as in their framework no statistical method is available to estimate this unknown parameter.

5. ANALYSIS OF CROATIAN FINANCIAL MARKET
For a transitional economy, with a short history of market economy such as Croatia the main problem for a statistically significant analysis is the short history of active trading in the financial markets and their low liquidity. Because of the short time series of returns of individual stock and their highly variable liquidity it is practical to analyse the stock indexes of the two stock exchanges in Croatia. Although a small country, Croatia has two stock exchanges, Zagreb Stock Exchange (ZSE) and Varazdin Stock Exchange (VSE). This situation reduces further the liquidity of an already illiquid market, so choosing to analyse the portfolio of securities that form the official stock indexes of Zagreb Stock Exchange (CROBEX) and Varazdin Stock Exchange (VIN), and by definition present the most liquid stocks on these stock exchanges presents itself as a logical choice. The analysis of the two stock indexes is performed in period 04.01.2000. – 04.01.2006. In this observation period the obtained sample of CROBEX index consists of 1469 observations, and for VIN index, the sample consists of 1483 observations. The data is sourced from ZSE and VSE. Since the period from the beginning of year 2000 onward was very turbulent for Croatian financial market, the analysis of the selected indexes consists of three parts (entire period, first and second half of the period). This procedure is implemented to detect any structural changes in the characteristics of selected indexes, i.e. test if the assumption of stationarity of time series can be applied, while at the same time no generality is lost due to the statistically significant number of observations in analysed half-periods. In case of CROBEX index, half-periods consist of 734 observations, and the first half-period covers the period from 04.01.2000 to 19.12.2002, for VIN index half periods consist of 741 observations, and the first half period, covers the same period as CROBEX index. The two analysed indexes show a strong positive trend. This was expected since Croatian securities are currently trading at a discount compared to other surrounding market, especially EU new member states. Due to the process of accession of Republic of Croatia to the European Union, and the adjustment of legislation and business climate to European standards, the growth of stock indexes is a natural consequence. Croatia is subject to an evergrowing inflow of foreign direct and portfolio investments that is further boosting the appreciation of Croatian securities. From figures 1 to 4 it is visible that there is significant volatility clustering and presence of extreme positive and negative returns. Looking at a measure of linear dependence between two variables i.e. correlation coefficient which equals 0,9714 for the entire analysed period, suggest that the CROBEX index and VIN index are strongly positively correlated. First half-period sample correlation coefficient between CROBEX index and VIN index equals 0,8954. Second half period sample correlation coefficient between CROBEX index and VIN index equals 0,9782. These results indicate that in both half-periods there is significant linear dependence between these two indexes, as could be expected because although there are composed of completely different securities, they both represent the same market. The fact that there is a strong linear dependence between these two indexes, that has strengthened even further in the last three years, and the fact that the presence of two stock exchanges in such a small country reduces the transparency and liquidity of the whole financial market gives a strong argument to those

that advocates the merger of two Croatian stock exchanges. In spite of the strong linear dependence of the two indexes, their descriptive statistics, in particular higher moments around the mean, tell a different story. Descriptive statistics for the CROBEX and VIN index are presented in table 1.
Table 1 - Summary descriptive statistics for CROBEX and VIN index in the period 04.01.2000 - 04.01.2006.
Descriptive statistics Mean Median Mode Minimum Maximum St.Dev. Skewness Kurtosis CROBEX CROBEX CROBEX VIN index VIN index VIN index index (whole index (first index (whole (second (first half) period) half) (second half) period) half) 0,00067 0,00059 0,00074 0,00128 0,00096 0,00160 0,00023 0,00046 0,00009 0,00053 0,00 0,00100 0,00 0,00 0,00 0,00 0,00 0,00 -0,09032 -0,08854 -0,09032 -0,15670 -0,15670 -0,06547 0,11399 0,11399 0,07355 0,10186 0,10186 0,05837 0,01359 0,01596 0,01071 0,01280 0,01493 0,01025 0,23920 0,39708 -0,30973 -0,68183 -0,94255 0,35245 9,86525 7,59346 11,71601 20,76386 20,83433 5,59279

Both CROBEX and VIN index have significant positive means and medians in the entire analysed period as well in both half-periods. This clearly points to the conclusion that securities composing these two indexes had a steady positive mean, resulting in considerable capital gains for the investors. For both indexes the highest positive means are present in the second half-period which covers the period from 20.12.2002 to 04.01.2006, and can be easily explained by the positive influence of the accession negotiations between Croatia and EU. Mean, median and mode of the stock indexes are not equal for individual indexes, which is assumed under normality of the distribution. Standard deviation of both indexes is quite high during the entire analysed period, and is equal to 13,59% for CROBEX index, and 12,80% for VIN index. Both indexes were more volatile in the first analysed half-period, equalling 15,96% for CROBEX index, and 14,93% for VIN index. Based on the standard deviation of the two indexes, it turns out that CROBEX index was the more volatile index, and thus riskier. Combining this conclusion with the realized daily mean return, during the entire period, for CROBEX index (0,067%) and for VIN index (0,128%), it turns out that VIN index, in the entire observed period, was at the same time less volatile and more profitable, especially in the second half period when its daily mean return reached 0,160%. Third moment around the mean (skewness) for both indexes, and all the analysed periods, is significantly different from the zero, which is assumed under normal distribution. In the entire analysed period VIN index had negative skewness of –0,68183, while CROBEX index had positive skewness of 0,2392. This fact is very important for the investors meaning that the probability of positive returns occurring is far greater when investing in CROBEX index than in VIN index. The interesting phenomenon is the significant positive skewness in the second half period of the VIN index and negative skewness of CROBEX index. This shows that in the last three years the skewness of the analysed indexes has completely changed, VIN index has become positively skewed, while CROBEX index has become negatively skewed. The fourth moment around the mean (kurtosis) for both indexes, and all the analysed periods is different from the zero, as presumed under normality. Both indexes experienced extreme daily returns in the observed period. The high values of kurtosis for these indexes, especially VIN index, indicate to the investors investing in either stock exchange that they can expect unusually high, both positive and negative returns on their investments. Combining the third and fourth moment of the VIN index with the previously described mean and standard deviation, one may conclude that, although in the entire observation period, negative returns were more frequent than positive returns, the magnitude of the positive

returns was significantly higher than the magnitude of loses resulting in a strong positive trend and continually growing index. In case of the CROBEX index, for the whole period, the standard deviation was higher compared to the VIN index, and also the mean return was lower, meaning that from the classical portfolio theory perspective CROBEX index can be viewed as an inferior portfolio compared to VIN index. Skewness and kurtosis, during the entire period, of CROBEX index indicate that from higher moments perspective was in fact not as inferior to the VIN index as could be expected from looking at just the first two moments of the distribution of the returns of these two indexes. To determine if the daily returns of CROBEX and VIN indexes are normally distributed, normality of distribution is tested in several ways. For CROBEX index, the tests and analysis are performed on 1469 observations for the entire period and 734 observations for halfperiods. For VIN index, the tests and analysis are performed on 1483 observations for the entire period and 741 observations for half-periods. The simplest test of normality is the analysis of the third and fourth moment around the mean of the distribution. Third moment around the mean, asymmetry, in the case of normal distribution should be zero. Negative asymmetry means that the distribution is skewed to the left, which implies that there is a greater chance of experiencing negative returns, and vice versa. Fourth moment around the mean, kurtosis, in the case of normal distribution, when modified, should also equal zero. Kurtosis higher than zero means that the distribution has fatter tails than normal distribution, meaning that more extreme events occur more frequently in the sample data than can be expected under normal distribution. More sophisticated tests of normality of distribution used in this paper are Lilliefors test, Shapiro-Wilks test and Jarque-Bera test. Normality tests for the CROBEX and VIN index are presented in table 2.
Table 2 – Tests of normality of distribution for CROBEX and VIN index returns in the period 04.01.2000 04.01.2006.
Normality tests Lilliefors (p value) Shapiro Wilk (p value) CROBEX CROBEX CROBEX index (whole index (first index period) half) (second half) VIN index (whole period) VIN index (first half) VIN index (second half)

0,09406
0,00 0,89222 0,00

0,10067
0,00 0,90728 0,00

0,08178
0,00 0,89983 0,00

0,10510
0,00 0,86270 0,00

0,11095
0,00 0,84574 0,00

0,08961
0,00 0,92738 0,00

1.741,80 26.470,00 13.241,00 Jarque-Bera 5.900,70 4.118,80 0,00 0,00 0,00 0,00 0,00 (p value) Critical value for Lilliefors test (whole period = 0,023124, half period = 0,032703) Critical value for Jarque-Bera test = 5,9915

957,33
0,00

All normality tests show that the hypothesis of the normality of returns for CROBEX and VIN indexes, over all the analysed periods, should be rejected at 5% significance level. Probability values of distribution of returns being normal, for both indexes, under all of the normality test are zero, strongly indicating that there is no possibility that the returns on these indexes are normally distributed. The distribution of CROBEX and VIN index returns are leptokurtotic and are not symmetrical i.e. they skew to the left and to the right, as can be seen from figure 5. Since the assumption of IID returns underlies the logic behind the historical simulation it is necessary to test whether returns of the analysed time series are indeed IID. Returns on financial assets themselves are usually not dependent (correlated), otherwise traders could forecast daily returns. Returns squared are usually dependent; meaning that volatility of the returns can be forecasted, but not the direction of the change of a variable.

A widely accepted approach to detecting volatility clusters, which is actually autoregression in the data, is the Ljung-Box Q-statistic calculated on the squared returns and Engle’s Archtest. Ljung-Box Q-statistic is the lth autocorrelation of the T-squared returns, and calculates whether the size of the movement at time t has any useful information to predict the size of the movement at time t+l. Engle's Archtest for the presence of autoregressive conditional heteroskedasticity (ARCH) effects tests the hypothesis that a time series of sample residuals consists of IID Gaussian disturbances, i.e., that no ARCH effects exist. Ljung-Box Q-statistic and Archtest for CROBEX and VIN index are presented in table 3. The Ljung-Box Q-statistic and Engle’s Archtest confirm that there is significant autocorrelation and ARCH effects present in both indexes i.e. that the volatilities tend to cluster together (periods of low volatility are followed by further periods of low volatility and vice versa), and for all the analysed periods, with the exception of CROBEX index in the second-half period, meaning that the returns on CROBEX and VIN index are not IID. The results are that much more indicative when considering that the hypothesis of IID was rejected for all the tested time lags (5, 10, 15 and 20 days) and all of the indexes, with the exception of CROBEX index in the second half-period. This is very indicative for risk managers, because these tests prove that the elementary assumption of historical simulation is not satisfied, and that the VaR figures obtained from it cannot be trusted and can at best provide only unconditional coverage. Because of the existence of volatility clustering BRW approach could provide a better alternative to historical simulation, due to the fact it assigns more probability weight to recent events, thus providing an updating scheme that is more responsive to changes in the market.

6. PERFORMANCE EVALUATION OF HISTORICAL SIMULATION AND BRW APPROACH ON CROATIAN STOCK INDEXES
The performance of the historical simulation and BRW approach is tested in two ways. First test is the Kupiec test13, a simple expansion of the failure rate, which is also prescribed by Basel Committee on Banking Supervision. The set-up for this test is the classic framework for a sequence of successes and failures, also known as Bernoulli trials. The second test used is the test of temporal unpredictability of extreme returns i.e. tail events. While extreme returns do happen, a good VaR estimator should not allow them to cluster together. An efficient VaR estimator should react in such a way that once a tail event occurs, it increases in such a way that, given this new estimate, the next tail event has the same probability of occurring i.e. it is temporally independent. The testing of temporal independence of tail events is done by examining the sample autocorrelation function of tail events, which in the case of temporal independence should not be significant. For the period of 1.000 days realized daily returns of CROBEX and VIN index are compared with the VaR forecasts based on historical simulation and BRW approach at 95%, 97,5% and 99% confidence level. Four historical simulation models with rolling windows of (50, 100, 250 and 400 days), and two BRW models with decay factors of 0,99 and 0,97, are shown in figures 6 to 17. Backtesting results of Kupiec test are presented in table 5, and backtesting results of test of temporal independence of tail events is presented in table 4. Backtesting results based on Kupiec procedure, for both indexes indicate that historical simulation based on shorter rolling windows (50 and 100 days) performs poorly at both 95%,
13

Jorion, P. (2001)

97,5% and at 99% confidence level. Historical simulation based on longer observation periods (250 and 400 days) was accepted at 5% significance level, as being unconditionally correct for both indexes at 95% and 97,5% confidence level but failed at 99% confidence level, with the exception of historical simulation based on 400 day rolling window which provided unconditionally correct coverage for VIN index at all the tested confidence levels. A fair performance of historical simulation models based on longer observation periods can be attributed to the fact that more extreme losses were present in the observation period, and that is why the binomial test accepted these models as being unconditionally correct, but unfortunately these models are also the slowest to react to the changes. Backtesting results for historical simulation based on test of temporal independence of tail events show that there exists temporal dependence in the tail events for all of the tested historical simulation models. When comparing the results of test for temporal independence between the indexes, they are mixed, and show that for CROBEX index historical simulation performed quite well while for VIN index it performed poorly. This results show that historical simulation is not an efficient VaR estimator and does not react adequately to the changes in the market and occurrence of extreme events i.e. it does not adapt adequately and timely to the changes, and thus does not reflect the true risk of a position. Backtesting results for BRW approach based on Kupiec procedure, for both indexes show that BRW approach performed quite good, with BRW model with decay factor set at 0,97 provided unconditional coverage for 95% and 97,5% confidence levels for both indexes but failed at 5% significance level for 99% confidence level. BRW model with decay factor set at 0,99 provided unconditional coverage for both indexes, at all of the tested confidence levels. Backtesting results for BRW approach based on test of temporal independence of tail events show mixed results. The existence of temporal dependence in the tail events under BRW approach for VIN index can be confirmed for both BRW models. In case of the CROBEX index, the existence of temporal dependence in the tail events under BRW approach can be rejected. This shows that when it comes to temporal independence, BRW approach should not be taken for granted since it also showed that it to does not always react adequately to the changes in the market and can also be misleading about the true level of risk. On a sample of two Croatian stock indexes, CROBEX and VIN, BRW approach proved to be a far better performer according to both the Kupiec test and test of temporal independence. BRW model with decay factor set at 0,99, according to the Kupiec test provided unconditional coverage for both indexes, at all of the tested confidence levels and as such was the best performer. According to the test of temporal independence BRW models were also better performers than any historical simulation models. The results point to the conclusion that even though historical simulation with longer observation periods provided correct unconditional coverage for 95% and 97,5% confidence levels banks and other investors should be very careful when using it. Historical simulation should not be used for high confidence level estimates (above 95%), especially models based on shorter rolling windows. The obtained results show that although BRW approach also has its flaws, especially when testing for temporal dependence in the tail events, it brings significant improvement to historical simulation with minimal additional computational effort and could prove to be an interesting alternative to historical simulation.

7. CONCLUSION
Banks operating in Croatia and other EU new member and member candidate states employ the same risk measurement models in forming of provisions as they do in developed markets. This means that risk managers de facto presume similar or even equal market characteristics and behaviour in these developing markets, as they would expect in developed markets. This is a dangerous assumption, which is not realistic. Employing VaR models in forming of bank’s provisions that are not suited to financial markets that they are used on, can have serious consequences for banks, resulting in significant losses in trading portfolio that could pass undetected by the employed risk measurement models, leaving the banks unprepared for such events. In this paper the author evaluates the performance of a widely spread VaR risk measure – historical simulation against a hybrid model developed by Boudoukh, Richardson, and Whitelaw, in transitional surroundings. The performance of the VaR models is evaluated out-of-the-sample for two Croatian stock indexes based on a time series of 1.000 observations by using Kupiec test and test of temporal independence of tail events. The author finds that the basic assumption underlying the implementation of historical simulation is violated in case of two tested indexes i.e. the assumption that analysed returns are IID was rejected. Based on the performed tests it can be concluded that historical simulation should not be used for high confidence level estimates (above 95%), especially models based on shorter rolling windows. The obtained results show that although BRW approach also has its flaws, especially when testing for temporal dependence in the tail events, it brings significant improvement to historical simulation with minimal additional computational effort. BRW approach should be further studied and tested in other transitional and emerging economies, because based on these obtained results it prove to be a far better alternative to historical simulation.

BIBLIOGRAPHY:
1. Basel Committee on Banking Supervision (1996): ‘Amendment to the Capital Accord to incorporate Market Risks. Bank for International settlements’, BIS, January 1996 Benson, P. and Zangari, P. (1997): ‘A general approach to calculating VaR without volatilities and correlations’, RiskMetrics Monitor, (second quarter 1997): 19-23 Boudoukh, J., Richardson, M. and Whitelaw, R. (1998): ‘The best of both worlds’, Risk, Vol. 11, No. 5, (May 1998): 64-67 Butler, J.S. and Schachter, B. (1998): ‘Estimating Value-at-Risk with a Precision Measure by Combining Kernel Estimation with Historical Simulation’, Review of Derivatives Research, No. 1, (1998): 371-390 Jorion, P. (2001): Value at Risk, The New Benchmark for Managing Financial Risk, Second Edition, New York: McGraw Hill Mandelbrot, B. (1963): 'The Variation of Certain Speculative Prices', The Journal of Business, Vol. 36, No. 4, (October 1963): 394-419 Manganelli, S. and Engle F.R. (2001): ‘Value at Risk models in Finance’, ECB working paper series, Working paper, No. 75, August 2001 Pritsker, M. (2001): ‘The Hidden Dangers of Historical Simulation’, Board of Governors of the Federal Reserve System, Economics Discussion Series, Working paper, No. 27, April 2001.

2.

3.

4.

5.

6.

7.

8.

Table 3 - Test of independency for CROBEX and VIN index returns in the period 04.01.2000 - 04.01.2006.
Ljung-Box-Pierce Q-test (CROBEX index - whole period) Period Critical H p-value Statistic (days) value 5 10 15 1 1 1 Ljung-Box-Pierce Q-test (CROBEX index - first half) Period Critical H p-value Statistic (days) value 5 10 15 1 1 1

1 20 ARCH test (CROBEX index - whole period) Period H p-value Statistic (days) 5 10 15 20 1 1 1 1

0 0 0 0

156,830 169,500 184,490 199,270

11,070 18,307 24,996 31,410
Critical value

1 20 ARCH test (CROBEX index - first half) Period H p-value Statistic (days) 5 10 15 20 1 1 1 1

0 0 0 0

105,460 109,870 119,710 128,680

11,070 18,307 24,996 31,410
Critical value

0 0 0 0

122,500 129,660 144,310 144,370

11,070 18,307 24,996 31,410

0 0 0 0

86,947 91,913 103,550 104,120

11,070 18,307 24,996 31,410

Ljung-Box-Pierce Q-test (CROBEX index - second half) Period Critical H p-value Statistic (days) value 5 10 15 0 0 0

Ljung-Box-Pierce Q-test (VIN index - whole period) Period Critical H p-value Statistic (days) value 5 10 15 1 1 1

0 20 ARCH test (CROBEX index - second half) Period H p-value Statistic (days) 5 10 15 20 0 0 0 0

0,55914 0,68312 0,88390 0,97845

3,933 7,442 8,876 9,353

11,070 18,307 24,996 31,410
Critical value

1 20 ARCH test (VIN index - whole period) Period H p-value Statistic (days) 5 10 15 20 1 1 1 1

0 0 0 0

140,200 140,860 150,860 156,510

11,070 18,307 24,996 31,410
Critical value

0,62209 0,77846 0,93332 0,98110

3,509 6,424 7,752 9,150

11,070 18,307 24,996 31,410

0 0 0 0

147,560 148,560 157,560 159,130

11,070 18,307 24,996 31,410

Ljung-Box-Pierce Q-test (VIN index - first half) Period Critical H p-value Statistic (days) value 5 10 15 1 1 1

Ljung-Box-Pierce Q-test (VIN index - second half) Period Critical H p-value Statistic (days) value 5 10 15 1 1 1

20 1 ARCH test (VIN index - first half) Period H p-value (days) 5 10 15 20 1 1 1 1

0 0 0 0

74,249 74,516 78,853 80,807
Statistic

11,070 18,307 24,996 31,410
Critical value

20 1 ARCH test (VIN index - second half) Period H p-value Statistic (days) 5 10 15 20 1 1 1 1

0 0 0 0

104,120 118,710 140,520 163,490

11,070 18,307 24,996 31,410
Critical value

0 0 0 0

81,225 81,460 85,560 85,970

11,070 18,307 24,996 31,410

0 0 0 0

72,142 75,961 94,324 99,334

11,070 18,307 24,996 31,410

Table 4 - Test of independency for CROBEX and VIN index tail events in the period 04.01.2000 - 04.01.2006. CROBEX index
Ljung-Box-Pierce Q-test (Tail events for HS 50 95%) Period Critical H p-value Statistic (days) value 5 10 15 0 0 0 Ljung-Box-Pierce Q-test (Tail events for HS 250 95%) Period Critical H p-value Statistic (days) value 5 10 15 1 0 0

0 20 Ljung-Box-Pierce Q-test (Tail events for HS 100 95%) Period Critical H p-value Statistic (days) value 5 10 15 20 1 0 1 0

0,35268 0,39100 0,20365 0,43852

5,548 10,582 19,227 20,313

11,070 18,307 24,996 31,410

0 20 Ljung-Box-Pierce Q-test (Tail events for HS 400 95%) Period Critical H p-value Statistic (days) value 5 10 15 20 1 0 0 0

0,01962 0,10062 0,11886 0,32669

13,435 15,965 21,595 22,258

11,070 18,307 24,996 31,410

0,03837 0,09236 0,04948 0,14492

11,751 16,263 25,035 26,666

11,070 18,307 24,996 31,410

0,01108 0,05531 0,05376 0,18561

14,837 17,979 24,725 25,426

11,070 18,307 24,996 31,410

Ljung-Box-Pierce Q-test (Tail events for BRW λ=0,97 95%) Period Critical H p-value Statistic (days) value 5 10 15 0 0 0

0 20 Ljung-Box-Pierce Q-test (Tail events for BRW λ=0,99 95%) Period Critical H p-value Statistic (days) value 5 10 15 20 1 0 0 0

0,68598 0,91650 0,93993 0,98885

3,091 4,596 7,568 8,402

11,070 18,307 24,996 31,410

0,04230 0,16225 0,17775 0,42272

11,501 14,242 19,850 20,572

11,070 18,307 24,996 31,410

VIN index
Ljung-Box-Pierce Q-test (Tail events for HS 50 95%) Period Critical H p-value Statistic (days) value 5 10 15 1 1 1 Ljung-Box-Pierce Q-test (Tail events for HS 250 95%) Period Critical H p-value Statistic (days) value 5 10 15 1 1 1

1 20 Ljung-Box-Pierce Q-test (Tail events for HS 100 95%) Period Critical H p-value Statistic (days) value 5 10 15 20 1 0 1 1

1,14E-05 0,00024 1,41E-05 6,33E-07

30,571 33,354 49,582 66,661

11,070 18,307 24,996 31,410

1 20 Ljung-Box-Pierce Q-test (Tail events for HS 400 95%) Period Critical H p-value Statistic (days) value 5 10 15 20 1 1 1 1

0,00142 0,02010 0,00203 1,50E-05

19,710 21,146 35,589 57,905

11,070 18,307 24,996 31,410

0,01018 0,08009 0,00346 1,26E-06

15,0430 16,7500 33,9510 64,7860

11,070 18,307 24,996 31,410

2,12E-06 0,00012 1,77E-05 4,60E-08

34,254 35,074 48,978 73,633

11,070 18,307 24,996 31,410

Ljung-Box-Pierce Q-test (Tail events for BRW λ=0,97 95%) Period Critical H p-value Statistic (days) value 5 10 15 0 0 1

1 20 Ljung-Box-Pierce Q-test (Tail events for BRW λ=0,99 95%) Period Critical H p-value Statistic (days) value 5 10 15 20 1 0 1 1

0,16974 0,46721 0,019947 0,0009445

77,639 96,999 28,269 45,497

11,070 18,307 24,996 31,410

0,009488 0,080374 0,0046158 1,45E-05

15,214 16,738 33,053 57,984

11,070 18,307 24,996 31,410

Table 5 – Backtesting results for 1.000 historical simulation and BRW VaR forecasts for CROBEX and VIN index CROBEX index
HS 50 (95%) Number of failures 65 Probability value 0,01493 Frequency of failures 0,065 0,05052 Confidence interval 0,0821 Accept model NO

Model

HS 50 HS 50 HS 100 HS 100 (97,5%) (99%) (95%) (97,5%) 41 22 53 35 0,00101 0,00027 0,30017 0,02104 0,041 0,022 0,053 0,035 0,02958 0,01384 0,03995 0,0245 0,05521 0,03312 0,06876 0,04834
NO NO YES NO

HS 100 (99%) 22 0,00027 0,022 0,01384 0,03312
NO

HS 250 (95%) 46 0,68847 0,046 0,03387 0,06088
YES

HS 250 (97,5%) 26 0,36964 0,026 0,01705 0,03787
YES

HS 250 (99%) 14 0,08241 0,014 0,00767 0,02338
NO

BRW λ=0,97 (95%) Number of failures 44 24 14 53 Probability value 0,78532 0,52761 0,08241 0,30017 Frequency of failures 0,044 0,024 0,014 0,053 0,03215 0,01544 0,00767 0,03995 Confidence interval 0,05862 0,0355 0,02338 0,06876 Accept model YES YES NO YES

Model

HS 400 (95%)

HS 400 HS 400 (97,5%) (99%)

BRW BRW BRW BRW BRW λ=0,97 λ=0,97 λ=0,99 λ=0,99 λ=0,99 (97,5%) (99%) (95%) (97,5%) (99%) 27 14 45 27 11 0,29801 0,08241 0,73904 0,29801 0,30265 0,027 0,014 0,045 0,027 0,011 0,01787 0,00767 0,03301 0,01787 0,0055 0,03904 0,02338 0,05975 0,03904 0,0196
YES NO YES YES YES

VIN index
HS 50 (95%) Number of failures 58 Probability value 0,11056 Frequency of failures 0,058 0,04433 Confidence interval 0,07434 YES Accept model

Model

HS 50 HS 50 HS 100 HS 100 (97,5%) (99%) (95%) (97,5%) 38 26 50 31 0,00514 5,57E-06 0,46247 0,09729 0,038 0,026 0,05 0,031 0,02703 0,017053 0,03734 0,02116 0,05179 0,037865 0,06539 0,04372
NO NO YES NO

HS 100 (99%) 14 0,08241 0,014 0,00767 0,02338
NO

HS 250 (95%) 50 0,46247 0,05 0,03734 0,06539
YES

HS 250 (97,5%) 30 0,13381 0,03 0,02033 0,04255
YES

HS 250 (99%) 14 0,08241 0,014 0,00767 0,02338
NO

Model

HS 400 (95%)

HS 400 (97,5%)

Number of failures 49 25 Probability value 0,52026 0,44707 Frequency of failures 0,049 0,025 0,03647 0,01624 Confidence interval 0,06427 0,03669 Accept model YES YES

BRW HS 400 λ=0,97 (99%) (95%) 11 46 0,30265 0,68847 0,011 0,046 0,0055 0,03387 0,0196 0,06088
YES YES

BRW BRW BRW BRW BRW λ=0,97 λ=0,97 λ=0,99 λ=0,99 λ=0,99 (97,5%) (99%) (95%) (97,5%) (99%) 29 16 43 23 9 0,17933 0,02639 0,82671 0,60806 0,5427 0,029 0,016 0,043 0,023 0,009 0,01951 0,00917 0,03129 0,01464 0,00412 0,04139 0,02585 0,05749 0,03431 0,01702
YES NO YES YES YES

Figure 1 – Daily values of CROBEX index in the period 04.01.2000 - 04.01.2006. (1469 observations)
2 2 0 0 2 0 0 0 1 8 0 0 1 6 0 0 1 4 0 0 1 2 0 0 1 0 0 0 8 0 0 6 0 0 J a n

2 0 0 0

J a n

2 0 0 1

J a n

2 0 0 2

J a n

2 0 0 3

J a n

2 0 0 4

J a n

2 0 0 5

J a n

2 0 0 6

Figure 2 – Daily returns on CROBEX index in the period 04.01.2000 - 04.01.2006. (1468 observations)
0 . 1 5 0 . 1

0 . 0 5 Return

0

- 0 . 0 5

- 0 . 1 J a n

2 0 0 0

J a n

2 0 0 1

J a n

2 0 0 2

J a n

2 0 0 3

J a n

2 0 0 4

J a n

2 0 0 5

J a n

2 0 0 6

Figure 3 – Daily values of VIN index in the period 04.01.2000 - 04.01.2006. (1483 observations)
2 2 0 0 2 0 0 0 1 8 0 0 1 6 0 0 1 4 0 0 1 2 0 0 1 0 0 0 8 0 0 6 0 0 4 0 0 2 0 0 J a n 2 0 0 0 J a n 2 0 0 1 J a n 2 0 0 2 J a n 2 0 0 3 J a n 2 0 0 4 J a n 2 0 0 5 J a n 2 0 0 6

Figure 4 – Daily returns on VIN index in the period 04.01.2000 - 04.01.2006. (1482 observations)
0 . 1 5 0 . 1 0 . 0 5 0 - 0 . 0 5 - 0 . 1 - 0 . 1 5 - 0 . 2 J a n

Return

2 0 0 0

J a n

2 0 0 1

J a n

2 0 0 2

J a n

2 0 0 3

J a n

2 0 0 4

J a n

2 0 0 5

J a n

2 0 0 6

Figure 5 - Probability plot for CROBEX and VIN index, period 04.01.2000 - 04.01.2006

0.99 0.95 0.9 0.75 0.5 0.25 0.1 0.05 0.01

0.99 0.95 0. 9 0.75 0. 5 0.25 0. 1 0.05 0.01

Probability

-0.05

0 Return

0.05

0.1

Probability

-0.15

-0.1

-0. 05 Return

0

0.05

0.1

CROBEX index

VIN index

Figure 6 – Historical simulation VaR for 1.000 observations at confidence level of 1, 2.5 and 5 percent with 50 days observation window for CROBEX index
0 . 1 5 C H H H R O S 5 S 5 S 5 B E 0 9 0 9 0 9 X 5 % 7 , 5 % 9 %

0 . 1

0 . 0 5 Return

0

-0 . 0 5

-0 . 1 J a n

2 0 0 2

J a n

2 0 0 3

J a n

2 0 0 4

J a n

2 0 0 5

J a n

2 0 0 6

Figure 7 – Historical simulation VaR for 1.000 observations at confidence level of 1, 2.5 and 5 percent with 100 days observation window for CROBEX index
0 . 1 5 C H H H R O S 1 S 1 S 1 B 0 0 0 E X 0 9 5 % 0 9 7 , 5 % 0 9 9 %

0 . 1

0 . 0 5 Return

0

-0 . 0 5

-0 . 1 J a n

2 0 0 2

J a n

2 0 0 3

J a n

2 0 0 4

J a n

2 0 0 5

J a n

2 0 0 6

Figure 8 – Historical simulation VaR for 1.000 observations at confidence level of 1, 2.5 and 5 percent with 250 days observation window for CROBEX index
0 .1 5 C H H H R O S 2 S 2 S 2 B 5 5 5 E X 0 9 5 % 0 9 7 ,5 % 0 9 9 %

0 .1

0 .0 5 Return

0

-0 .0 5

-0 .1 J a n

2 0 0 2

J a n

2 0 0 3

J a n

2 0 0 4

J a n

2 0 0 5

J a n

2 0 0 6

Figure 9 – Historical simulation VaR for 1.000 observations at confidence level of 1, 2.5 and 5 percent with 400 days observation window for CROBEX index
0 . 1 5 C H H H R O S 4 S 4 S 4 B 0 0 0 E X 0 9 5 % 0 9 7 , 5 % 0 9 9 %

0 . 1

0 . 0 5 Return

0

-0 . 0 5

-0 . 1 J a n

2 0 0 2

J a n

2 0 0 3

J a n

2 0 0 4

J a n

2 0 0 5

J a n

2 0 0 6

Figure 10 – BRW VaR (λ=0,97) for 1.000 observations at confidence level of 1, 2.5 and 5 percent for CROBEX index
0 . 1 5 C B B B R R R R O B E X W H S W H S W H S 9 5 % 9 7 , 5 % 9 9 %

0 . 1

0 . 0 5 Return

0

-0 . 0 5

-0 . 1 J a n

2 0 0 2

J a n

2 0 0 3

J a n

2 0 0 4

J a n

2 0 0 5

J a n

2 0 0 6

Figure 11 – BRW VaR (λ=0,99) for 1.000 observations at confidence level of 1, 2.5 and 5 percent for CROBEX index
0 . 1 5 C B B B R R R R O B E X W H S W H S W H S 9 5 % 9 7 , 5 % 9 9 %

0 . 1

0 . 0 5 Return

0

-0 . 0 5

-0 . 1 J a n

2 0 0 2

J a n

2 0 0 3

J a n

2 0 0 4

J a n

2 0 0 5

J a n

2 0 0 6

Figure 12 – Historical simulation VaR for 1.000 observations at confidence level of 1, 2.5 and 5 percent with 50 days observation window for VIN index
0 . 0 6 0 . 0 4 0 . 0 2 0 - 0 . 0 2 - 0 . 0 4 - 0 . 0 6 - 0 . 0 8 J a n V H H H 2 0 0 2 J a n 2 0 0 3 J a n IN S S S 5 0 5 0 5 0 9 5 % 9 7 , 5 % 9 9 % J a n 2 0 0 5 J a n 2 0 0 6

Return

2 0 0 4

Figure 13 – Historical simulation VaR for 1.000 observations at confidence level of 1, 2.5 and 5 percent with 100 days observation window for VIN index
0 . 0 6 0 . 0 4 0 . 0 2 0 - 0 . 0 2 - 0 . 0 4 - 0 . 0 6 - 0 . 0 8 J a n V H H H 2 0 0 2 J a n 2 0 0 3 J a n IN S S S 1 0 0 1 0 0 1 0 0 9 5 % 9 7 , 5 % 9 9 % J a n 2 0 0 5 J a n 2 0 0 6

Return

2 0 0 4

Figure 14 – Historical simulation VaR for 1.000 observations at confidence level of 1, 2.5 and 5 percent with 250 days observation window for VIN index
0 . 0 6 0 . 0 4 0 . 0 2 0 -0 . 0 2 -0 . 0 4 -0 . 0 6 -0 . 0 8 J a n V H H H 2 0 0 2 J a n 2 0 0 3 J a n IN S S S 2 5 0 2 5 0 2 5 0 9 5 % 9 7 , 5 % 9 9 % J a n 2 0 0 5 J a n 2 0 0 6

Return

2 0 0 4

Figure 15 – Historical simulation VaR for 1.000 observations at confidence level of 1, 2.5 and 5 percent with 400 days observation window for VIN index
0 .0 6 0 .0 4 0 .0 2 0 -0 .0 2 -0 .0 4 -0 .0 6 -0 .0 8 J a n V H H H 2 0 0 2 J a n 2 0 0 3 J a n IN S 4 0 0 S 4 0 0 S 4 0 0 2 0 0 4 9 5 % 9 7 ,5 % 9 9 % J a n 2 0 0 5 J a n 2 0 0 6

Figure 16 – BRW VaR (λ=0,97) for 1.000 observations at confidence level of 1, 2.5 and 5 percent for VIN index
0 .0 6 0 .0 4 0 .0 2 0 -0 .0 2 -0 .0 4 -0 .0 6 -0 .0 8 Ja n V B B B 2 0 0 2 Ja n 2 0 0 3 IN R W R W R W H S H S H S 9 5 % 9 7 ,5 % 9 9 % Ja n 2 0 0 5 Ja n 2 0 0 6

Return

Return

Ja n

2 0 0 4

Figure 17 – BRW VaR (λ=0,99) for 1.000 observations at confidence level of 1, 2.5 and 5 percent for VIN index
0 .0 6 0 .0 4 0 .0 2 0 -0 .0 2 -0 .0 4 -0 .0 6 -0 .0 8 Ja n V B B B 2 0 0 2 Ja n 2 0 0 3 Ja n IN R W R W R W H S H S H S 9 5 % 9 7 ,5 % 9 9 % Ja n 2 0 0 5 Ja n 2 0 0 6

Return

2 0 0 4

Sign up to vote on this title
UsefulNot useful