Abstract
Empirical evidence rejects the assumption that the distribution is normal and suggests that the distribution of financial asset returns be heavy-tailed. We compare the performance of extreme value theory in VaR calculations with that of other well-known modeling techniques, such as variance-covariance method and historical simulation. Using GARCH(1,1), EWMA and historical volatility to estimate the volatility. In extreme value method, we choose the peaks over thresholds to derive a natural model for the point process of large losses exceeding a high threshold. Moreover, we use the generalized Pareto distribution (GPD) to describe the samples which exceeding threshold. We provide the historical simulation is the most useful VaR forecast model. But in the extreme things, EVT is better than historical simulation.