Microsoft Word - 00_tresc.docx DYNAMIC ECONOMETRIC MODELS Vol. 10 – Nicolaus Copernicus University – Toruń – 2010 Joanna Górka Nicolaus Copernicus University in Toruń The Sign RCA Models: Comparing Predictive Accuracy of VaR Measures† A b s t r a c t. Evaluating Value at Risk (VaR) methods of predictive accuracy in an objective and effective framework is important for both efficient capital allocation and loss prediction. From this reasons, finding an adequate method of estimating and backtesting is crucial for both the regulators and the risk managers’. The Sign RCA models may be useful to obtain the accurate forecasts of VaR. In this research one briefly describes the Sign RCA models, the Value at Risk and backtesting. We compare the predictive accuracy of alternative VaR forecasts obtained from different models. Empirical example is mainly related to the PBG Capital Group shares on the Warsaw Stock Exchange. K e y w o r d s: Family of Sign RCA Models, Value at Risk, backtesting, loss function. 1. Introduction Nowadays, accurate modelling of risk is very important in risk management. This is a result of the globalisation of financial market, the evolution of the derivative markets and the technological development. Value at Risk (VaR) has become the standard measure to quantify market risk1. This measure can be used by the financial institutions to assess their risks or by a regulatory committee to set margin requirements. In the literature, many parametric VaR models and many forecasting accuracy assessments for VaR methods exist. The important representation of the parametric VaR models are the generalized autoregressive conditional heteroskedasticity models (GARCH) (Bollerslev, 1986; Engle, 1982). These models describe non-linear dynamics of financial time series. A different, alternative approach to the description of financial time series represent the † This work was financed from the Polish science budget resources in the years 2008-2010 as the research project N N111 434034. 1 It was introduced by JP Morgan in 1996. Joanna Górka 62 random coefficient autoregressive models (RCA) (which were proposed by Nicholls, Quinn, 1982). Thavaneswaran et al. proposed a number of expansions of the random coefficient autoregressive model order one. The new models, such as Sign RCA(1), RCAMA(1,1), Sign RCAMA(1,1), RCA(1)-GARCH(1,1) and Sign RCA(1)-GARCH(1,1) can be used to obtain Value-at-Risk measure. The aim of this paper is to use the family of Sign RCA models to obtain the VaR forecasts and compare the results obtained from Sign RCA models with other selected VaR models. 2. The Family of Sign RCA Models Random coefficient autoregressive models (RCA) are straightforward generalization of the constant coefficient autoregressive models. A full description of this class of models including their properties, estimation methods and some applications can be found in Nicholls and Quinn (1982). Thavaneswaran, Appadoo and Bector (2006) proposed a first order random coefficient autoregressive model with a first order moving average component, i.e. RCAMA(1,1). In another paper Thavaneswaran and Appadoo (2006) proposed to add the sign function to RCA(1) and RCAMA(1,1) models. The last modification is based on assumption that residuals from the RCA model or the Sign RCA model can be described by a GARCH model. In this way, the RCA(1)-GARCH(1,1) model and Sign RCA(1)-GARCH(1,1) model were created. All these modifications influence the increase of variance and kurtosis of processes2. In Table 1 equations of individual models from the family of Sign RCA models and their names are presented. To ensure the existence of the I-VI models (Table 1) the following assumptions must be satisfied: ⎟ ⎟ ⎠ ⎞ ⎜ ⎜ ⎝ ⎛ ⎟ ⎟ ⎠ ⎞ ⎜ ⎜ ⎝ ⎛ ⎟⎟ ⎠ ⎞ ⎜⎜ ⎝ ⎛ ⎟⎟ ⎠ ⎞ ⎜⎜ ⎝ ⎛ 2 2 0 0 , 0 0 ~ ε δ σ σ ε δ iid t t , (1) 122 <+ δσφ . (2) The sign function, described by the following formula 1 for 0, 0 for 0, 1 for 0, t t t t y s y y >⎧ ⎪ = =⎨ ⎪− <⎩ (3) 2 Theoretical properties of the family of Sign RCA models can be found in articles, i.e.: Appadoo, Thavaneswaran, Singh (2006), Aue (2004), Górka, (2008), Thavaneswaran, Appadoo, Bector (2006), Thavaneswaran, Appadoo (2006), Thavaneswaran, Appadoo, Ghahramani, (2009), Thavaneswaran, Peiris, Appadoo (2008). The Sign RCA Models: Comparing Predictive Accuracy of VaR Measures 63 has the interpretation: if Φ>+ tδφ , the negative value of Φ means that the negative (positive) observation values at time 1−t correspond to a decrease (increase) of observation values at time t . In the case of stock returns it would suggest (for returns) that after a decrease of stock returns, the higher decrease of stock returns occurs than expected, and in the case of the increase of stock returns the lower increase in stock returns occurs than expected. Table 1. The family of Sign RCA models (without conditions) Model Model equations No. RCA(1) ( ) tttt yy εδφ ++= −1 I Sign RCA(1) ( ) ttttt ysy εδφ +Φ++= −− 11 II RCAMA(1,1) ( ) 11 −− +++= ttttt yy θεεδφ III Sign RCAMA(1,1) ( ) 111 −−− ++Φ++= tttttt ysy θεεδφ IV RCA(1)-GARCH(1,1) ( ) tttt yy εδφ ++= −1 , ttt zh=ε 11 2 110 −− ++= ttt hh βεαα V Sign RCA(1)-GARCH(1,1) ( ) ttttt ysy εδφ +Φ++= −− 11 , ttt zh=ε 11 2 110 −− ++= ttt hh βεαα VI Note: ts – sign function is described by equation (3); φ , θ , Φ , iα , 1β – model parameters. Condition (2) is necessary and sufficient for the second-order stationarity of process described by equation I, however conditions (1)-(2) ensure strict stationarity of this process. If conditions (1)-(2) are satisfied, then processes described by equations II-IV are stationary in mean. If residuals from the RCA model are described by a GARCH model, then the RCA(1)-GARCH(p,q) model described by equation V, where ( )2,0~ zt Nz σ , 00 >α , 0≥iα and 0≥jβ , is obtained. If the sign function is added to the RCA-GARCH model, then the process described by equation VI is obtained. The conditions ensuring the positive value of conditional variance of this process are the following: ( )2,0~ zt Nz σ , 00 >α , 0≥iα , 0≥jβ , 0α≤Φ . Predictors of the conditional mean and conditional variance of Sign RCA models are presented in Table 2 and 3 respectively. Joanna Górka 64 Table 2. Conditional mean predictors Models Conditional mean RCA(1), RCA(1)-GARCH(1,1) ( )11P t t tt ty E y F yϕ++ = = Sign RCA(1), Sign RCA(1)-GARCH(1,1) ( ) ( )11 P t t t tt ty E y F s yϕ++ = = +Φ RCAMA(1,1) ( )11P t t t tt ty E y F yϕ θε++ = = + Sign RCAMA(1,1) ( ) ( )11P t t t tt ty E y F s yϕ θε++ = = +Φ + Table 3. Conditional variance predictors Models Conditional variance RCA(1), Sign RCA(1), RCAMA(1,1), Sign RCAMA(1,1) ( ) 2 2 2 2 2 11 t t tt t E u F yε δσ σ σ++ = = + RCA(1)-GARCH(1,1), Sign RCA(1)-GARCH(1,1) ( ) 2 2 2 2 2 11 ( )t t z t tt t E u F E h yδσ σ σ++ = = + 3. Value-at-Risk Value-at-Risk (VaR) is used as a tool for measuring market risk. It is defined as „the maximum potential loss that a portfolio can suffer within a fixed confidence level during a holding period”. Formal definition of VaR is following (Artzner, Delbaen, Eber, Heath, 1999): ( ) ( ){ } ( ){ }VaR inf : inf : 1XX x F x x P X xα α α= ≥ = > ≤ − , (4) where ( )0,1α ∈ is a particular confidence level, XF – the cumulative density function. Consider a time series of daily ex post returns ( ( )1100 ln lnt t tr P P−= − where tP is the share price at time t) and corresponding time series of ex ante VaR forecasts ( VaRα ), the formula (4) takes the form: ( )1 VaRtP r α α+ ≤ − = . (5) The negative sign arises from the convention of reporting VaR as a positive number. One-step-ahead conditional forecasts of Value-at-Risk are calculated by the formula: ( )1 1 1VaR , l t t t t t zαα μ σ+ + += + (6) where tt |1+μ , tt |1+σ are one-step-ahead conditional forecasts of mean and volatility respectively. The Sign RCA Models: Comparing Predictive Accuracy of VaR Measures 65 3.1. Estimation Methods for VaR This section briefly describes the alternative models that we use for estimating VaR forecasts in this paper. The following models are used in the research to obtain VaR forecasts: − The historical simulation (HS)3. The VaR is estimated as the α-th quantile of the empirical distribution of returns. HS is based on the assumption that returns are iid time series of an unknown distribution. − The equally weighted moving average (EWMA) model, i.e. 2 2 1 1 1 t it t i t k r k σ + = − + = ∑ , (7) where k – size of window, 2ir – returns. The returns are assumed to be normally distributed. − The RiskMetrics (RM) model, i.e. ( ) ( )2 2 2 21 1 1 1 t t i i t tt t i t k r rσ λ λ λσ λ−+ = − + = − = + −∑ , (8) where ( )0,1λ ∈ is known as the decay factor, 2tλσ is the previous volatility forecast weighted by the decay factor, and ( ) 21 trλ− is the latest squared returns weighted by ( )1 λ− . The VaR is estimated under the assumption that returns are normally distributed (as in the case of EWMA). − The AR(1)-GARCH(1,1) model, i.e. 1t t tr rφ ε−= + , (9) where t t tzε σ= , )1,0(~ Nzt , 2 2 2 1 1t t tσ ω αε βσ− −= + + . (10) In this case, returns series is assumed to be conditionally normally distributed. − Models from the family of Sign RCA models4. 3 HS is the oldest and still very popular estimator of the VaR. 4 They were presented in previous section. Joanna Górka 66 3.2. Backtesting VaR Estimates Backtesting is based on testing whether the VaR estimates are statistically accurate. The ,,failure process” is defined as: ( )1 VaR , 1, ...,lt t tI r t T T N= < − = + + , (11) where ( )*1 denotes the indicator function returning a unit if the argument is true, and zero otherwise; T is the size of the sample used to estimate parameters of the model; N is the number of one-step-ahead VaR forecasts computed. The VaR forecasts are accurate if the { }tI series is iid with mean α , i.e. | 1t tE I α−⎡ ⎤ =⎣ ⎦ . To test the statistical accuracy we used the standard likelihood ratio tests: 1. The proportion of failures test – LRpof (Kupiec, 1995) 5: [ ] [ ]0 1: . :t tH E I vs H E Iα α= ≠ , 1 2 ln ˆ ˆ1 N n n pofLR α α α α −⎡ ⎤−⎛ ⎞ ⎛ ⎞ = − ⎢ ⎥⎜ ⎟ ⎜ ⎟−⎝ ⎠ ⎝ ⎠⎢ ⎥⎣ ⎦ ~ 21χ , (12) where n is the number of failures VaR, α̂ is the MLE of α , i. e. nN . 2. The Christoffersen independence test – LRind (Christoffersen, 1998): 0 01 11:H α α= , ( ) ( ) ( ) 00 10 01 11 00 1001 11 01 01 11 11 1 2 ln 1 1 T T T T ind T TT T LR α α α α α α + +− = − − − ~ 21χ , (13) where: 0 1 ˆ ijij i i T T T α = + , 11011000 1101 TTTT TT +++ + =α , ijT – number of i values followed by a j value in the tI series ( ), 0,1i j = . 3. The time between failures test – LRtbf (Haas, 2001) 6: 1 1 1 2 ln 1 ivN tbf i i i LR α α α α − = ⎡ ⎤⎛ ⎞− ⎢ ⎥= − ⎜ ⎟ −⎢ ⎥⎝ ⎠⎣ ⎦ ∑ ~ 2Nχ , (14) 5 Similar, the LR test of unconditional coverage by Christoffersen (1998) was proposed. Other symbol of this test is the LRuc. 6 Haas extended the Kupiec’s time until first failure test (TUFF test) by adding test for every exception (second and next). The Sign RCA Models: Comparing Predictive Accuracy of VaR Measures 67 where i i v 1 =α , 1v – time until first failure, iv – time between exception ( )1−i and exception i for 2, ...,i N= . If, in above tests the null hypothesis is not rejected, then a particular model gives accurate forecasts of VaR. However, if more than one model is deemed adequate, we cannot conclude which of VaR model should be selected. Lopez (1998) suggested measuring the accuracy of VaR forecasts on the basis of distance between observed returns and forecasted VaR values. This approach does not give any formal statistical selection of model adequacy but it allows to rank the models. Let 1 N t t f f = = ∑ means a total loss function. A model which minimizes the total loss function is preferred over the other models. In the literature, different loss functions were proposed (see Lopez, 1998, 1999; Blanco and Ihle, 1998; Sarma, Thomas and Shah, 2003, Caporin, 2003; Angelidis, Benos and Degiannakis, 2004). In this paper, the loss functions used to compare the accurate VaR forecasts are as follows: − The regulatory loss function – RL (Lopez, 1999)7: ( ) 1 , 21 1 , 1 , 0 VaR , 1 VaR VaR . t r t t t r t t r t r f r r + + + + > −⎧⎪ = ⎨ + + ≤ −⎪⎩ (15) − The firm’s loss function – FL (Sarma, Thomas, Shah, 2003): ( ) , 1 , 21 1 , 1 , VaR VaR , 1 VaR VaR . r t t r t t t r t t r t c r f r r + + + + > −⎧⎪ = ⎨ + + ≤ −⎪⎩ (16) where c is a measure of cost of capital opportunity. Sarma, Thomas and Shah (2003) proposed testing for superiority of a model vis-á-vis another in terms of the loss function. They suggested a two-stage VaR model selection procedure. The first stage consists in testing the statistical accuracy for the competing VaR models. In the second stage of the VaR model selection procedure, the firm’s loss function is used to evaluate statistically VaR models8. 7 This name comes from Sarma, Thomas and Shah (2003) who explain that (16) is able to express the regulatory concerns in model evaluation. However, no score is attached in case if exception does not occur. 8 Only that VaR model for which the average number of failures was equal to the expected and these failures are independently distributed is included in the second stage. Joanna Górka 68 Consider two VaR models, i and j. The hypotheses are: 0 1: 0 . : 0H vs Hθ θ= < , where θ is the median of the distribution of , ,t i t j tz f f= − , where ,i tf and ,j tf are the values of loss function generated by model i and model j respectively. Negative values of tz indicate a superiority of model i over j. The testing procedure is as follows: 1. Define an indicator variable ( )0t tzψ = ≥1 and the number of non-negative tz ’s, as 1 T N ij t t T S ψ + = + = ∑ . 2. Calculate the statistics as: ( ) 0.5 ~ 0,1 0.25 ij ij S N STS N N − = asymptotically, (17) ijSTS is based on assuming that the tz is iid 9. Alternatively, we can compare competing VaR models using the predictive quantile loss function (see Giacomini and Komunjer, 2005; Bao et al., 2006). The expected loss function is given by: ( ) ( ) 1 1 1 VaR VaR N i i i i i Q r r Nα α = ⎡ ⎤= − < − +⎣ ⎦∑ . (18) The selected model is the VaR model which has the minimum of Qα . 4. Empirical application The data used in the empirical application are daily prices of twenty Polish firms’ shares from the WIG20 portfolio on the Warsaw Stock Exchange (WSE). The data were obtained from bossa.pl for the period from 23-rd September 2005 to 18-th February 2009, which yields 852 observations. However, one of shares was excluded because it was not quoted on 23 September 2005. To analyze daily percentage log returns of each share were used. This empirical study was composed of two parts. The first part (Analysis I) was carried out with regard to all of twenty shares from WSE. The research procedure was the following: 1. For the first 500 observations of each returns series the descriptive statistics and some tests were calculated. Next, returns series with 9 For details on the sign test see Diebold and Mariano (1995). The Sign RCA Models: Comparing Predictive Accuracy of VaR Measures 69 autocorrelation and kurtosis bigger than for normal distribution were chosen10. 2. Parameters of six models from the family of Sign RCA were estimated for the first 500 observations of time series selected in step one. Next, only models with statistically significant parameters were used. 3. The estimation of parameters for models selected in step 2 was performed for rolling window of 100, 150, 200, 250, 300, 400, 500 observations. In the same way the estimation of AR(1)-GARCH(1,1) models was obtained. 4. For all models from step 3 and for the historical simulation (HS), the equally weighted moving average (EWMA) model, the RiskMetrics (RM) models (with 0.95λ = and 0.99λ = ) VaR measures were calculated11. One-step-ahead forecasts of VaR (that is 751, 701, 651, 601, 551, 451, 351 forecasts, respectively) were calculated on the basis of these models. 5. The traditional VaR tests and loss functions for each model and window were calculated. 6. The obtained results in above step were compared. In the second part (Analysis II) only the PBG shares (PBG Capital Group) was chosen. All presented models of VaR for the last 250 observations were calculated12. For obtained VaR forecasts the two-stage VaR model selection procedure was applied. All model parameters (Analysis I and II) were estimated using maximum likelihood (MLE) with the BFGS algorithm. Calculations were carried out in the Gauss program. 4.1. Results of the Analysis I Selected results of the descriptive statistics and some tests are given in Table 4. All series have a mean between -0.052 and 0.561, kurtosis bigger than for normal distribution. The standard deviations are different, ranging from 1.955 for PGNIG to 5.354 for BIOTON. The skewness and kurtosis differ among all series. Only 8 of 19 returns series have autocorrelation. The LBI test rejects the null hypothesis of non random coefficient to four stock returns. 10 This method of the elimination of initially selected companies can impact on the results. It would be worth to check out which results might be obtained for the whole set of companies. However, such analysis was omitted in this paper. 11 The returns series were assumed either to be normally distributed or conditionally normally distributed, respectively. 12 The set of 250 observations corresponds to roughly one year of trading days and according to the Basel II Accord requirement the minimum of 250 VaR forecasts should be used to the backtesting approach. Therefore, one-step-ahead forecasts of VaR at the same period (250 observations) were calculated. Parameters were estimated for rolling windows of 125, 250, 375 observations each. The returns series were assumed either to be conditionally normally distributed or normally distributed respectively. Joanna Górka 70 Next, the 7 different models were estimated for 8 returns series. Further, only models with statistically significant parameters were chosen. In this way models like RCA and Sign RCA were chosen. To present backtesting results for VaR forecasts of the PBG shares was chosen because for that share the autoregressive parameter in the RCA models for all returns series has been the biggest. It is very important because we can expect the Sign RCA model to be better than other models. The traditional VaR tests and loss functions for the PBG for all models are presented in Table 5 and the 5% at significance level. One can see that the accuracy test rejects the null hypothesis for windows size of 500, 400 observations for HS, EWMA model, AR(1)-GARCH(1,1) model, RCA model and Sign RCA model. For example, for window size 250 the regulatory loss function is the smallest for RM ( 0.95λ = ). Next position in this ranking have AR(1)-GARCH(1,1) model, EWMA model, RCA model, Sign RCA model and the last position has RM ( 0.99λ = ). The HS method is not taken into consideration because the accuracy test rejects the null hypothesis for windows size of 250 observations. On the other hand, the firm’s loss function is the smallest for RM ( 0.99λ = ) and the next positions in ranking have Sign RCA model, RCA model, EWMA model, AR(1)-GARCH(1,1) model and RM ( 0.95λ = ). The differences between values of the firm’s loss function are small for estimated models. To compare these results, the tests for superiority of a model vis-á-vis another were used only for models included into the second stage at Sarma, Thomas and Shah procedure. The results are presented in Table 6. For the window size 300 we can see that the Sign RCA model is significantly better than other models, i.e. the null hypothesis is rejected in the test of superiority between the Sign RCA model and the other models presented in subsection 3.1. However, as the size of windows decreases the RM model ( 0.99λ = ) outperforms the Sign RCA model. RCA and Sign RCA models are statistically the same for the window size 100. In cases when results with HS are compared one can see that HS is almost everywhere significantly better than others. The Table 7 includes the results of the VaR tests and the loss function at the 2,5% significance level which are similar to the results obtained at the 5% significance level. Only for HS with the window size 250 and for RCA model with the window size of 300 observations, some differences can be noticed, i. e. In the case of HS the accuracy at the 2.5% is better than at the 5% significance level (except RCA model). For the loss function conclusions are the same with one exception, i. e. the HS has the last rank for regulatory loss function and the first rank for firm’s loss function. At the 1% significance level we obtained more differences (see Table 8). Firstly, Risk Metrics models are accurate only for windows size 500 and 500, 400 for 0.95λ = , 0.99λ = , respectively. The RCA, Sign RCA and EWMA The Sign RCA Models: Comparing Predictive Accuracy of VaR Measures 71 models are accurate for small windows (size 200, 150, 100). The regulatory loss function is the smallest for HS. The firm’s loss function has the lowest values for Sign RCA models for the window size 200. Very strange results were obtained for HS and therefore we are not able to find any rules for accuracy and value of the regulatory loss function. 4.2. Results of the Analysis II Firstly, we calculated the 250 one-step-ahead forecasts of VaR of the PBG share using all models of VaR (presented in 3.1)13. The VaR forecasts were received from different models estimated for the different window sizes, i.e, T =125, 250 and 375. Secondly, the competing VaR models were testing for statistical accuracy. For the established period of forecasting, only Sign RCAMA(1,1) models (for T = 375 and all significance level, for T = 250 and α = 2.5%, 1%), Sign RCA(1)-GARCH(1,1) models (for α = 1% and rolling window size T = 375, 125) and Risk Metrics models (for λ= 0,99 and α = 1% and T = 125) did not fulfill the conditions used at first stage of Sarma, Thomas and Shah procedure (the null hypothesis was rejected at least for one test, see (12)-(14)). For other models, the firm’s loss function (see the Table 9), the STS test and the predictive quantile loss function (see the Table 10) were calculated. Lower values of the firm’s loss function for VaR forecasts were received from RCAMA(1,1), RCA(1) and Sign RCAMA(1,1) (if it was included at second stage) models (with the exception of the HS for α = 5% and T = 375, 250 and with the exception of the RM(λ= 0,99) for T = 125 and α = 5%, 2.5%). The test for superiority of a model vis-á-vis another indicates that: 1. At the 5% significance level, for different rolling window sizes, each of models having first rank is superior over other models. 2. At the 2.5% significance level, for rolling windows size of 125 observations, the RM (λ= 0,99) is superior over other models. The RCAMA(1,1) model is better than almost all other models (with the exception of HS method and RCA(1) model for T = 375 and with the exception of the RCA(1)-GARCH(1,1) model for T = 250, for which the predictive ability is equal). 3. For the α = 1%, for different rolling window sizes, each of models having first rank is superior over other models (with the exception of RCAMA(1,1) and RCA(1) models for T = 375 that have equal predictive ability). Other conclusions are formulated based on the predictive quantile loss function (Table 10), which yields different position in the ranking. For VaR forecasts of the PBG share, for established forecasting period, the choice of the 13 One-step ahead forecasts on the period 19.02.2008-18.02.2009 were computed. Joanna Górka 72 best model from the competing models depends on the significance level and rolling window sizes. For the Sign RCA models the rolling window size of 125 observations seemed too small. This conclusion is similar to one from Analysis I. 5. Conclusions Evaluating forecasts based solely on one criterion yield the limited information regarding the accuracy method. Thus, in the literature is commonly accepted that results of each evaluation criterion are presented separately and then best performing method is selected. However, it can be noticed that the different evaluation criteria give the different choice of the best estimation method of VaR. Therefore, it is difficult to make general remarks, nevertheless the empirical results showed that: 1. None of the presented methods gave a satisfactory VaR estimates. 2. The results showed no domination of either forecasting methods of VaR. 3. Bigger sample did not lead to the better results. 4. It seems that the family of Sign RCA models should be used for the sample size of 150 to 300 observations. 5. In terms of the firm’s loss function the Sign RCA model was significantly better than the AR-GARCH model, RM ( 0.95λ = ) model and EWMA model. The Sign RCA model was not worse than the standard RCA model. 6. One should treat every share individually and use different methods and models for obtaining a good forecast of VaR. 7. The historical simulation gave better results (in terms of accuracy) at the 1% significance level than for other significance levels. It seems that the minimum window size should be 250 observations but smaller than 500 observations. 8. The RCAMA(1,1) model can be competitive to other VaR measures from the firm’s loss function point of view. 9. The Sign RCA models with GARCH errors did not give better forecasts of VaR for the PBG share. References Angelidis, T., Benos, A., Degiannakis, S. (2004), The Use of GARCH Models in VaR Estimation, Statistical Methodology, 1, 105–128. Appadoo, S. S., Thavaneswaran, A., Singh, J. (2006), RCA Models with Correlated Errors, Applied Mathematics Letters, 19, 824–829. Artzner, P., Delbaen, F., Eber, J.-M., Heath, D. (1999), Coherent Measures of Risk, Mathematical Finance, 9, 203–228. Aue, A. (2004), Strong Approximation for RCA(1) Time Series with Applications, Statistics & Probability Letters, 68, 369–382. The Sign RCA Models: Comparing Predictive Accuracy of VaR Measures 73 Bao, Y., Lee, T.-H., Saltoglu, B. (2006), Evaluating Predictive Performance of Value-at-Risk Models in Emerging Markets: A Reality Check, Journal of Forecasting, 25,101–128. Blanco, C., Ihle, G. (1998), How good is your VaR? Using Backtesting to Assess System Performance, Financial Engineering News, August, 1–2. Bollerslev, T. (1986), Generalized Autoregressive Conditional Heteroscedasticity, Journal of Econometrics, 31, 307–327. Caporin, M. (2003), Evaluating Value-at-Risk Measures in Presence of Long Memory Conditional Volatility, Working Paper 05.03, GRETA. Christoffersen, P. F. (1998), Evaluating Interval Forecasts, International Economic Review, 39, 841–862. Diebold, F. X., Mariano, R. S. (1995), Comparing Predictive Accuracy, Journal of Business & Economic Statistics, 13, 253–263. Engle, R. F. (1982), Autoregressive Conditional Heteroscedasticity with Estimates of the Variance of United Kingdom Inflation, Econometrica, 50, 987–1006. Giacomini, R., Komunjer, I. (2005), Evaluation and Combination of Conditional Quantile Forecasts, Journal of Business and Economic Statistics, 23, 416–431. Górka, J. (2008), Description the Kurtosis of Distributions by Selected Models with Sing Function, Dynamic Econometric Models, 8, 39–49. Haas, M. (2001), New Methods in Backtesting, Financial Engineering, Working Paper, Bonn. Lopez, J. (1998), Methods for Evaluating Value-at-Risk Estimates, FRBNY Economic Policy Review. Lopez, J. (1999), Regulatory Evaluation of Value-at-Risk Models, FRBNY Economic Policy Review, 4, 119–124. Nicholls, D., Quinn, B. (1982), Random Coefficient Autoregressive Models: An Introduction, Springer, New York. Sarma, M., Thomas, S., Shah, A. (2003), Selection of Value-at-Risk Models, Journal of Forecasting, 22, 337–358. Thavaneswaran, A., Appadoo, S. S. (2006), Properties of a New Family of Volatility Sing Models, Computers & Mathematics with Applications, 52, 809–818. Thavaneswaran, A., Appadoo, S. S., Bector, C. R. (2006), Recent Developments in Volatility Modeling and Application, Journal of Applied Mathematics and Decision Sciences, 2006, 1–23. Thavaneswaran, A., Appadoo, S. S., Ghahramani, M. (2009), RCA Models with GARCH Innovations, Applied Mathematics Letters, 22, 110–114. Thavaneswaran, A., Peiris, S., Appadoo, S. (2008), Random Coefficien Volatility Models, Statistics & Probability Letters, 78, 582–593. Modele Sign RCA: Porównanie trafności prognoz VaR Z a r y s t r e ś c i. Obiektywna i skuteczna ocena trafności prognozowania wartości narażonej na ryzyko (Value at Risk – VaR) jest bardzo ważna zarówno dla efektywnego zarządzania kapitałem jak i do prognozowania strat. Z tego powodu znalezienie odpowiednich metod estymacji i weryfikacji VaR jest kluczowe zarówno dla instytucji nadzorujących jak i dla menadżerów. Modele Sign RCA mogą być użyteczne do otrzymywania trafnych prognoz VAR. W artykule, pokrótce przedstawione są modele Sign RCA, wartość narażona na ryzyko i weryfikacja prognoz VaR. Porównana jest trafność prognoz VaR otrzymanym z różnych alternatywnych modeli. Przykład empiryczny skoncentrowany jest głównie na cenach akcji spółki PBG notowanej na Giełdzie Papierów Wartościowych w Warszawie. S ł o w a k l u c z o w e: Modele klasy Sign RCA, Value at Risk, testowanie wsteczne, funkcja strat. Table 4. Results of the descriptive statistics, Box-Ljung tests and locally best invariant test Company Mean Std. Dev. Skewness Kurtosis B-L (1) B-L (2) LBI AGORA -0,052 2,451 -0,204 4,853 8,925*** 9,029*** 1,672 ASSECOPOL 0,169 2,493 -0,582 13,015 6,953*** 8,357*** 2,848** BIOTON -0,036 5,354 -8,286 138,621 1,673 2,111 -0,028 BRE 0,256 1,972 0,263 4,055 3,915** 4,025 2,378** BZWBK 0,175 2,442 -0,135 3,472 1,478 2,738 1,034 CERSANIT 0,247 2,361 0,567 6,312 0,156 1,887 1,639 GETIN 0,231 2,646 0,523 11,370 0,008 0,837 0,954 GTC 0,255 2,737 0,461 5,383 1,510 8,046*** 0,461 KGHM 0,212 3,011 -0,591 5,303 0,001 4,766 1,156 LOTOS 0,036 2,174 -0,329 4,835 1,596 2,249 -0,078 PBG 0,363 2,094 0,095 5,344 3,466* 3,468 1,909 PEKAO 0,071 2,160 0,219 3,616 0,005 0,044 0,273 PGNIG 0,068 1,955 0,192 4,413 0,284 4,870* 2,929** PKNORLEN -0,021 2,170 -0,069 3,853 0,017 3,680 0,508 PKOBP 0,117 2,055 0,324 3,912 3,625* 3,647 0,002 POLIMEXMS 0,366 2,420 -0,172 6,835 2,402 3,945 1,449 POLNORD 0,561 5,290 -1,387 28,269 2,085 2,489 -0,047 TPSA -0,022 1,978 -0,161 3,775 0,310 1,757 1,109 TVN 0,145 2,242 -0,083 3,716 3,004* 3,250 3,218** Note: *, **, *** indicate rejection of H0 at the 10% ,5% and 1% significant level, respectively. B-L (1) – estimates of the Box-Ljung test statistics of order 1. B-L (2) – estimates of the Box-Ljung test statistics of order 2. LBI – estimates of the locally best invariant test statistics. Table 5. Results of the VaR tests (95% VaR for PBG) and the loss function Model α̂ LRpof LRind LRtbf RL FL SH 500 10,54% 17,451*** 1,370 41,329 205,55 1366,59 400 9,09% 12,929*** 1,139 33,525 240,15 1722,11 300 7,62% 6,923*** 0,603 34,955 279,97 2106,98 250 7,15% 5,213** 0,494 34,272 281,66 2246,19 200 5,84% 0,914 0,026 32,805 261,26 2462,33 150 5,56% 0,453 0,016 40,548 263,62 2668,63 100 4,79% 0,068 0,394 35,815 224,13 2913,60 EWMA 500 9,12% 10,179*** 1,967 32,075 186,89 1427,20 400 7,98% 7,207*** 0,350 28,098 208,58 1818,34 300 6,35% 1,961 0,027 26,143 236,04 2235,86 250 5,82% 0,817 0,723 28,900 229,42 2428,40 200 5,07% 0,007 0,348 31,243 223,04 2628,37 150 4,42% 0,512 0,121 32,300 222,40 2856,85 100 4,26% 0,907 0,117 35,007 215,48 3079,17 Table 5. Continued Model α̂ LRpof LRind LRtbf RL FL RM (λ= 0,95) 500 7,12% 2,959* 3,850* 25,133 117,94 1629,97 400 6,43% 1,788 0,544 25,727 148,73 2036,69 300 5,99% 1,070 0,657 27,728 206,95 2394,60 250 5,49% 0,296 0,481 29,504 206,95 2546,11 200 5,07% 0,007 0,348 33,443 206,96 2686,53 150 4,85% 0,033 0,326 36,837 210,83 2856,07 100 4,66% 0,186 0,310 37,665 213,93 3052,66 RM (λ= 0,99) 500 6,55% 1,630 3,238* 24,787 137,97 1576,26 400 5,99% 0,872 0,306 25,381 172,62 1948,88 300 6,17% 1,484 0,796 28,060 230,84 2272,02 250 6,16% 1,581 0,041 28,334 240,02 2397,50 200 6,14% 1,678 0,104 35,573 255,05 2502,31 150 6,56% 3,293 0,000 46,305 287,75 2571,93 100 7,19% 6,720*** 0,253 55,251 346,13 2559,78 AR(1)-GARCH(1,1) 500 8,26% 6,628** 5,247** 25,450 168,05 1497,28 400 7,54% 5,331** 1,411 29,023 185,14 1884,31 300 6,72% 3,095* 0,117 31,296 233,15 2267,60 250 5,66% 0,525 0,596 27,342 219,27 2460,78 200 5,38% 0,190 0,550 31,366 214,92 2629,52 150 5,14% 0,027 0,514 36,663 215,14 2868,68 100 4,26% 0,907 0,117 36,932 221,00 3138,36 RCA 500 8,83% 8,924*** 1,693 26,278 187,36 1442,89 400 7,76% 6,238** 1,630 28,136 204,92 1828,93 300 6,53% 2,498 0,065 27,313 237,79 2223,04 250 5,82% 0,817 0,723 26,199 230,09 2410,01 200 5,07% 0,007 0,348 25,277 221,17 2594,08 150 4,99% 0,000 0,415 37,995 227,54 2812,48 100 4,39% 0,604 0,171 34,631 220,41 3025,38 Sign RCA 500 8,83% 8,924*** 1,693 26,278 186,61 1438,09 400 7,54% 5,331** 5,564** 27,559 204,61 1838,01 300 6,53% 2,498 0,065 27,312 239,14 2219,78 250 5,82% 0,817 0,723 26,199 230,74 2404,58 200 5,07% 0,007 0,348 25,277 221,40 2586,80 150 4,99% 0,000 0,038 39,013 227,35 2801,30 100 4,79% 0,068 0,864 37,365 228,79 3003,25 Note: *, **, *** indicate rejection of H0 at the 10% ,5% and 1% significant level, respectively, LRpof – the values of the proportion of failures test statistics, LRind – the values of the independence test statistics, LRtbf – the values of the time between failures test statistics, RL – regulatory loss function, FL – firm’s loss function. Table 6. The test for superiority of a model vis-á-vis another Sample: 300 ↓ better → Sign RCA RCA AR-GARCH RM(0.99) RM(0.95) EWMA HS Sign RCA x -7,455* -8,052* -10,863* -9,330* -12,397* RCA 7,455 x -3,962* -10,182* -9,159* -3,451* AR-GARCH 8,052 3,962 x -2,343* -9,245* 0,724 RM(0.99) 10,863 10,182 2,343 x -8,989* 8,904 RM(0.95) 9,330 9,159 9,245 8,989 x 8,563 EWMA 12,397 3,451 -0,724 -8,904* -8,563* x HS x Sample: 250 ↓ better → Sign RCA RCA AR-GARCH RM(0.99) RM(0.95) EWMA HS Sign RCA x -5,262* -4,691* 1,020 -8,199* -6,323 RCA 5,262 x -4,691* 1,999 -7,954* -5,099 AR-GARCH 4,691 4,691 x 4,854 -7,954* -0,612 RM(0.99) -1,020 -1,999 -4,854* x -9,994* -6,159 RM(0.95) 8,199 7,954 7,954 9,994 x 6,078 EWMA 6,323 5,099 0,612 6,159 -6,078* x HS x Sample: 200 ↓ better → Sign RCA RCA AR-GARCH RM(0.99) RM(0.95) EWMA HS Sign RCA x -5,369* -5,056* 11,640 -4,194* -8,662* 12,895 RCA 5,369 x -3,253* 12,581 -3,880* -5,683* 13,992 AR-GARCH 5,056 3,253 x 14,619 -1,842 -2,234* 13,522 RM(0.99) -11,640* -12,581* -14,619* x -10,308* -16,971* 4,586 RM(0.95) 4,194 3,880 1,842 10,308 x 1,999 11,092 EWMA 8,662 5,683 2,234 16,971 -1,999 x 14,619 HS -12,895* -13,992* -13,522* -4,586* -11,092* -14,619* x Sample: 150 ↓ better → Sign RCA RCA AR-GARCH RM(0.99) RM(0.95) EWMA HS Sign RCA x -0,567 -2,984* 19,905 -2,002* -8,120* 10,462 RCA 0,567 x -3,059* 21,113 -1,775 -7,063* 12,502 AR-GARCH 2,984 3,059 x 20,887 2,379 -2,757* 13,257 RM(0.99) -19,905* -21,113* -20,887* x -14,315* -24,135* -6,761* RM(0.95) 2,002 1,775 -2,379* 14,315 x -0,944 10,613 EWMA 8,120 7,063 2,757 24,135 0,944 x 15,448 HS -10,462* -12,502* -13,257* 6,761 -10,613* -15,448* x Sample: 100 ↓ better → Sign RCA RCA AR-GARCH RM(0.99) RM(0.95) EWMA HS Sign RCA x -1,715 -5,218* -2,153* -9,159* 4,634 RCA 1,861 x -5,729* -1,861 -6,386* 6,313 AR-GARCH 5,218 5,729 x 4,415 2,007 9,086 RM(0.99) x RM(0.95) 2,153 1,861 -4,415* x 0,401 6,240 EWMA 9,159 6,386 -2,007* -0,401 x 9,597 HS -4,634* -6,313* -9,086* -6,240* -9,597* x Note: * indicate rejection of H0 at the 10% and 5% significant level. Table 7. Results of the VaR tests (97.5% VaR for PBG) and the loss functions Model α̂ LRpof LRind LRtbf RL FL SH 500 5,98% 12,642*** 2,683 35,191** 123,33 1628,86 400 5,10% 9,659*** 0,031 30,168 135,69 2074,56 300 4,54% 7,587*** 0,019 27,865 171,95 2467,40 250 3,66% 2,912* 0,047 17,641 155,85 2764,23 200 3,69% 3,289* 0,015 24,976 154,39 2950,27 150 3,14% 1,086 0,130 20,965 158,94 3208,25 100 3,06% 0,911 0,117 23,765 157,49 3400,58 EWMA 500 4,84% 6,234** 1,736 28,034** 114,91 1661,45 400 4,88% 8,226*** 0,006 29,047 130,49 2111,82 300 3,81% 3,358* 0,049 25,024 151,85 2597,32 250 3,33% 1,533 0,156 20,585 145,23 2833,12 200 3,07% 0,816 0,217 21,926 144,02 3070,24 150 2,85% 0,343 0,281 23,676 144,58 3332,11 100 2,53% 0,003 0,455 25,225 138,45 3610,65 RM (λ = 0,95) 500 3,13% 0,536 0,714 12,178 60,26 1932,02 400 3,10% 0,628 0,582 10,837 80,65 2404,45 300 3,27% 1,214 0,256 13,397 129,14 2810,04 250 3,00% 0,569 0,337 12,575 129,14 2990,57 200 2,76% 0,181 0,419 13,916 129,15 3157,89 150 2,85% 0,343 0,281 21,614 132,36 3352,30 100 2,80% 0,261 0,255 21,282 134,38 3584,68 RM (λ = 0,99) 500 3,99% 2,710 1,167 23,185* 80,16 1836,63 400 3,77% 2,586 0,186 21,844 104,95 2268,16 300 3,63% 2,537 0,099 21,407 148,64 2644,88 250 3,33% 1,533 0,156 20,585 153,45 2799,71 200 3,23% 1,291 0,143 25,544 164,68 2925,99 150 3,71% 3,668* 0,001 31,614 189,66 3000,01 100 4,93% 14,208*** 0,018 54,397** 240,05 2947,31 AR(1)-GARCH(1,1) 500 5,41% 9,215*** 2,182 26,878 103,79 1732,20 400 4,21% 4,517*** 1,676 22,917 109,42 2199,89 300 3,63% 2,537 0,099 21,407 148,91 2650,81 250 3,33% 1,533 0,156 20,585 139,01 2877,29 200 3,07% 0,816 0,217 21,926 136,56 3081,41 150 3,14% 1,086 0,130 27,158 134,82 3359,94 100 3,06% 0,911 0,117 22,612 142,15 3662,46 Table 7. Continued Model α̂ LRpof LRind LRtbf RL FL RCA 500 4,84% 6,234** 1,736 24,499* 115,61 1674,19 400 4,66% 6,888*** 2,057 29,051 127,91 2125,85 300 3,99% 4,277** 0,017 25,205 154,07 2583,65 250 3,33% 1,533 0,156 20,585 146,76 2814,50 200 3,07% 0,816 0,217 21,926 142,87 3032,33 150 2,85% 0,343 0,281 23,676 144,83 3292,28 100 2,80% 0,261 0,255 24,688 141,09 3540,01 Sign RCA 500 5,13% 7,666*** 1,953 26,390 115,69 1665,44 400 3,99% 3,494* 1,500 23,079 126,41 2142,75 300 3,81% 3,358* 0,049 21,490 154,02 2582,48 250 3,33% 1,533 0,156 20,585 147,33 2807,82 200 3,23% 1,291 0,143 24,923 143,88 3020,79 150 3,00% 0,665 1,299 27,904 145,29 3276,84 100 2,93% 0,539 0,179 27,440 145,11 3517,04 Note: *, **, *** indicate rejection of H0 at the 10% ,5% and 1% significant level, respectively, LRpof – the values of the proportion of failures test statistics, LRind – the values of the independence test statistics, LRtbf – the values of the time between failures test statistics, RL – regulatory loss function, FL – firm’s loss function. Table 8. Results of the VaR tests (99% VaR for PBG) and the loss functions Model α̂ LRpof LRind LRtbf RL FL SH 500 2,56% 6,056** 0,475 17,577** 42,00 2163,96 400 1,77% 2,218 0,290 13,426* 56,23 2775,29 300 1,27% 0,375 0,180 4,422 77,63 3452,54 250 2,16% 6,162** 0,576 17,518 91,70 3464,14 200 1,08% 0,036 0,152 1,888 67,93 4456,03 150 2,00% 5,459** 0,571 24,181** 105,19 4229,94 100 0,93% 0,036 0,132 3,400 60,19 5611,84 EWMA 500 3,13% 10,313*** 0,714 23,320** 66,52 1933,39 400 2,66% 8,633*** 0,658 25,503** 74,00 2476,85 300 2,00% 4,285** 0,449 13,755 90,20 3046,29 250 2,00% 4,676** 1,436 11,420 86,18 3320,07 200 1,54% 1,624 0,313 7,374 86,15 3617,66 150 1,14% 0,135 0,185 3,069 84,75 3936,79 100 1,07% 0,032 0,173 3,535 81,86 4268,12 Table 8. Continued Model α̂ LRpof LRind LRtbf RL FL RM (λ = 0,95) 500 1,99% 2,719* 0,286 9,927 26,51 2271,53 400 2,22% 5,014** 1,581 12,666 39,05 2820,60 300 2,36% 7,441*** 1,053 14,764 77,73 3286,66 250 2,16% 6,162** 1,181 12,415 77,73 3500,93 200 2,00% 5,067** 1,303 12,230 77,73 3699,53 150 2,00% 5,459** 1,184 14,802 78,92 3935,00 100 1,86% 4,516** 1,288 12,587 79,58 4213,58 RM (λ = 0,99) 500 1,71% 1,472 0,209 5,627 38,83 2171,52 400 1,55% 1,188 0,221 7,810 54,51 2679,38 300 2,00% 4,285** 0,449 14,208 89,53 3103,32 250 2,00% 4,676** 0,490 12,402 93,77 3280,42 200 2,15% 6,547** 1,074 15,053 104,02 3421,39 150 2,85% 16,200*** 0,281 39,564 124,04 3485,92 100 3,06% 20,831*** 0,117 48,907 157,58 3431,58 AR(1)-GARCH(1,1) 500 2,28% 4,259** 0,374 11,647 55,22 2048,00 400 2,22% 5,014** 0,455 15,475 58,34 2590,51 300 1,81% 2,977* 0,370 11,299 89,91 3118,01 250 1,83% 3,360* 0,411 9,867 81,74 3386,59 200 1,69% 2,592 0,379 7,917 81,97 3635,11 150 1,71% 2,958 0,419 14,533 79,46 3965,85 100 1,33% 0,755 0,270 9,869 79,37 4329,54 RCA 500 3,13% 10,313*** 0,714 23,320** 67,06 1947,01 400 2,88% 10,707*** 0,774 25,957** 73,93 2487,60 300 2,18% 5,778** 0,535 18,055 92,98 3031,22 250 2,00% 4,676** 0,490 13,800 88,61 3302,35 200 1,69% 2,592 0,379 11,307 86,68 3571,25 150 1,43% 1,138 0,290 9,825 86,71 3881,87 100 1,20% 0,281 0,219 6,357 82,19 4187,14 Sign RCA 500 3,13% 10,313*** 0,714 23,320** 65,98 1940,98 400 2,88% 10,707*** 0,774 25,957** 75,11 2494,86 300 2,00% 4,285** 0,449 13,755 92,42 3029,72 250 2,00% 4,676** 0,490 13,800 88,94 3293,79 200 1,69% 2,592 0,379 11,307 86,81 3560,67 150 1,43% 1,138 0,290 11,433 85,97 3867,81 100 1,33% 0,755 0,270 11,177 84,15 4156,81 Note: *, **, *** indicate rejection of H0 at the 10% ,5% and 1% significant level, respectively, LRpof – the values of the proportion of failures test statistics, LRind – the values of the independence test statistics, LRtbf – the values of the time between failures test statistics, RL – regulatory loss function, FL – firm’s loss function. Table 9. Results of the firm’s loss function Model T = 375 T = 250 T = 125 FL rank FL rank FL rank α = 5% Sym, Hist, 1075,956 1 1141,764 2 1263,7779 10 EWMA 1109,013 5 1192,913 9 1257,8114 8 RM (λ= 0,95) 1229,387 9 1251,9575 10 1263,0397 9 RM (λ= 0,99) 1189,228 8 1191,2823 8 1101,7305 1 AR(1)-GARCH(1,1) 1154,158 7 1169,4068 6 1207,2458 4 RCA 1101,740 3 1153,1165 4 1212,1169 6 Sign RCA 1103,836 4 1172,3837 7 1239,823 7 RCAMA 1098,295 2 1149,6521 3 1204,0552 3 Sign RCAMA - - 1102,4462 1 1180,7179 2 RCA GARCH 1132,959 6 1158,4801 5 1209,4749 5 Sign RCA GARCH 1296,929 10 1339,8815 11 1404,0839 11 α = 2,5% Sym, Hist, 1284,699 5 1364,1283 5 1521,6798 10 EWMA 1282,948 4 1377,3036 7 1454,5161 8 RM (λ= 0,95) 1447,782 9 1473,5854 9 1481,442 9 RM (λ= 0,99) 1379,883 8 1378,9041 8 1283,0513 1 AR(1)-GARCH(1,1) 1334,391 7 1366,5462 6 1426,3841 6 RCA 1272,254 2 1341,8047 2 1406,9397 5 Sign RCA 1274,395 3 1358,9357 4 1433,1749 7 RCAMA 1270,903 1 1337,7536 1 1397,165 3 Sign RCAMA - - - - 1361,558 2 RCA GARCH 1305,188 6 1349,0442 3 1405,9143 4 Sign RCA GARCH 1549,821 10 1577,6011 10 1630,4145 11 α = 1% Sym, Hist, 1754,0863 9 1754,7449 9 2234,0469 9 EWMA 1506,0564 4 1623,2342 7 1723,7757 7 RM (λ= 0,95) 1697,2481 8 1719,5178 8 1727,7362 8 RM (λ= 0,99) 1625,3806 7 1619,3369 6 - - AR(1)-GARCH(1,1) 1569,7932 6 1611,8139 5 1688,9632 5 RCA 1494,4265 2 1573,8864 2 1663,8987 3 Sign RCA 1496,858 3 1593,0414 4 1698,2509 6 RCAMA 1492,2845 1 1568,9729 1 1650,9211 2 Sign RCAMA - - - - 1600,3206 1 RCA GARCH 1535,3552 5 1589,0235 3 1664,72 4 Sign RCA GARCH - - 1849,3472 10 - - Note: T denotes the rolling window size, FL– the firm’s loss function. Table 10. Results of the the predictive quantile loss function Model T = 375 T = 250 T = 125 Qα rank Qα rank Qα rank α = 5% Sym. Hist 0,3183 7 0,3159 7 0,3202 7 EWMA 0,3169 5 0,3141 5 0,3181 5 RM (λ= 0,95) 0,3191 8 0,3182 9 0,3189 6 RM (λ= 0,99) 0,3163 4 0,3153 6 0,3213 9 AR(1)-GARCH(1,1) 0,3215 9 0,3167 8 0,3122 1 RCA 0,3156 1 0,3131 3 0,3173 3 Sign RCA 0,3161 3 0,3122 1 0,3169 2 RCAMA 0,3174 6 0,3130 2 0,3206 8 Sign RCAMA - - 0,3363 11 0,3253 10 RCA GARCH 0,3158 2 0,3137 4 0,3179 4 Sign RCA GARCH 0,3421 10 0,3283 10 0,3657 11 α = 2,5% Sym. Hist 0,1945 5 0,1908 2 0,1938 6 EWMA 0,1943 4 0,1910 3 0,1914 2 RM (λ= 0,95) 0,1916 1 0,1914 5 0,1915 3 RM (λ= 0,99) 0,1923 2 0,1931 9 0,1995 9 AR(1)-GARCH(1,1) 0,1985 9 0,1913 4 0,1845 1 RCA 0,1947 6 0,1925 6 0,1956 7 Sign RCA 0,1943 3 0,1896 1 0,1927 4 RCAMA 0,1957 8 0,1931 8 0,1982 8 Sign RCAMA - - - - 0,2016 10 RCA GARCH 0,1953 7 0,1926 7 0,1928 5 Sign RCA GARCH 0,2108 10 0,2048 10 0,2358 11 α = 1% Sym. Hist 0,1015 9 0,0992 9 0,0957 6 EWMA 0,0978 3 0,0947 1 0,0941 4 RM (λ= 0,95) 0,0973 2 0,0962 5 0,0958 7 RM (λ= 0,99) 0,0957 1 0,0960 3 - - AR(1)-GARCH(1,1) 0,1000 7 0,0961 4 0,0909 1 RCA 0,0997 6 0,0963 6 0,0950 5 Sign RCA 0,0988 5 0,0950 2 0,0934 2 RCAMA 0,1004 8 0,0967 7 0,0969 8 Sign RCAMA - - - - 0,1022 9 RCA GARCH 0,0980 4 0,0967 8 0,0938 3 Sign RCA GARCH - - 0,1143 10 - - Note: T denotes the rolling window size, Qα – the predictive quantile loss function.