Improving the accuracy of tail risk forecasting models by combining several realized volatility estimators

The statistical properties of realized volatility estimators critically depend on the sampling frequency of the underlying intra-day returns and on the chosen estimation formula. This gives rise to a substantial model uncertainty when realized volatility is used as a regressor in tail risk forecasting models. In this paper, aiming to mitigate the impact of model uncertainty on the generation of tail risk forecasts, we propose parsimonious extensions of the Realized Exponential GARCH model that combine information from several volatility estimators. Both fixed and time-varying parameter models are considered. An application to the prediction of daily Value-at-Risk and Expected Shortfall for the S&P 500 provides evidence that modelling approaches based on the combination of different frequencies and estimation formulas can lead to significant accuracy gains.