Sensitivity Analysis of Two-Step Multinomial Backtests for Evaluating Value-at-Risk

Nowadays, the measurement of the risk of the marketplace has a significant effect on investments; however, the inadequate evaluation of this risk will cause a financial crisis and possible bankruptcy. One of the typical approaches to measure this risk is the probability-based risk measurement method, known as Value-at-Risk (VaR), for estimating and backtesting of which there are various methods. The purpose of this paper is to put forward a comprehensive test for backtesting and analyzing the sensitivity of VaR based on the number of samples (n) and confidence levels (N).

First, the VaR of Tehran Stock Exchange data was estimated by applying GARCH-Copula, DCC, and EVT. Next, by using the multinomial backtesting in two steps the accuracy of VaR estimation and ranked the models were tested. Thereafter, considering the number of samples (n) and the confidence levels (N), the sensitive analysis of the backtesting result demonstrated the accuracy of the estimated VaR by selecting the most appropriate parameters.

Sensitive analysis findings indicated that in all three models, increasing the parameter "N" will result in an increase in the error rate. On the other hand, sensitive analysis of parameter "n" proved that its value depends on the technique used to estimate VaR, but generally, any increase in it leads to validation of VaR estimation models. The results also showed that according to the EVT method, at least 29% of the data is required to be used as a test sample in VaR estimation; however, the amount is equal to 22% in the DCC and GARCH-Copula methods.

The result of the sensitivity analysis indicated that the reliability of different estimating VaR techniques relies on "n" and "N" parameters and different amounts of these two parameters can generate inaccurate and uncertain outcomes for each model. In addition, ranking these methods by using the loss function, GARCH-Copula, EVT and DCC methods ranked first to third, respectively.