Comparative analysis of sigma-based, quantile-based and time series VaR estimators

Abstract

Since its inception at the end of the XX century, VaR risk measure has gained massive popularity. It is synthetic, easy in interpretation and offers comparability of risk levels reported by different institutions. However, the crucial idea of comparability of reported VaR levels stays in contradiction with the differences in estimation procedures adopted by companies. The issue of the estimation method is subject to the internal company decision and is not regulated by the international banking supervision.

The paper was dedicated to comparative analysis of the prediction errors connected with competing VaR estimation methods. Four methods, among which two stationarity-based – variance-covariance and historical simulation – and two time series methods – GARCH and RiskMetrics^TM – were compared through the Monte Carlo study. The analysis was conducted with respect to the method choice, series length and VaR tolerance level.

The study outcomes showed the superiority of the sigma-based method of variance-covariance over the quantile-based historical simulation. Furthermore the comparison of the stationarity-based estimates to the time series results showed that allowing for time-varying parameters in the estimation technique significantly reduces the estimator bias and variance.