Evaluating Risk Forecasts with Central Limits

Abstract

Portfolio risk forecasts are commonly evaluated using test statistics that are sums of random variables. We study the distributional properties of these test statistics for value at risk, expected shortfall, and volatility. For a diverse collection of 74 US equity portfolios, risk forecasts based on an extreme value theory model greatly outperform a conditional normal model with a 23-day halflife. On the other hand, we show that the common assumption of asymptotic normality in test statistics for these risk measures is not always satisfied, especially for test statistics related to volatility.

Keywords: hyptothesis test, value at risk, expected shortfall