Robust Value at Risk Prediction

We propose a general robust semiparametric bootstrap method to estimate conditional predictive distributions of GARCH-type models. Our approach is based on a robust estimator for the parameters in GARCH-type models and a robustified resampling method for standardized GARCH residuals, which controls the bootstrap instability due to influential observations in the tails of standardized GARCH residuals. Monte Carlo simulation shows that our method consistently provides lower VaR forecast errors, often to a large extent, and in contrast to classical methods never fails validation tests at usual significance levels. We test extensively our approach in the context of real data applications to VaR prediction for market risk, and find that only our robust procedure passes all validation tests at usual confidence levels. Moreover, the smaller tail estimation risk of robust VaR forecasts implies VaR prediction intervals that can be nearly 20% narrower and 50% less volatile over time. This is a further desirable property of our method, which allows to adapt risky positions to VaR limits more smoothly and thus more efficiently.