The GARCH(1,1) and GARCH(1,1)-t models lead to highly volatile quantile forecasts, while historical simulation, Variance–Covariance, adaptive generalized Pareto distribution and non-adaptive generalized Pareto distribution models provide more stable quantile forecasts. In general, GARCH(1,1)-t, generalized Pareto distribution models and historical simulation are preferable for most quantiles.
Mahdavi G, Majedi Z. The application of extreme value theory in value at risk estimation:The case of liability insurance claims in Iran insurance company. JSS 2010; 4 (1) :59-76 URL: http://jss.irstat.ir/article-1-87-en.html