Improved VaR forecasts using extreme value theory with the Realized GARCH model

Abstract

Purpose This study aims to forecast daily value-at-risk (VaR) for international stock indices by using the conditional extreme value theory (EVT) with the Realized GARCH (RGARCH) model. The predictive ability of this Realized GARCH-EVT (RG-EVT) model is compared with those of the standalone GARCH models and the conditional EVT specifications with standard GARCH models. Design/methodology/approach The authors use daily data on returns and realized volatilities for 13 international stock indices for the period from 1 January 2003 to 8 October 2014. One-step-ahead VaR forecasts are generated using six forecasting models: GARCH, EGARCH, RGARCH, GARCH-EVT, EGARCH-EVT and RG-EVT. The EVT models are implemented using the two-stage conditional EVT framework of McNeil and Frey (2000). The forecasting performance is evaluated using multiple statistical tests to ensure the robustness of the results. Findings The authors find that regardless of the choice of the GARCH model, the two-stage conditional EVT approach provides significantly better out-of-sample performance than the standalone GARCH model. The standalone RGARCH model does not perform better than the GARCH and EGARCH models. However, using the RGARCH model in the first stage of the conditional EVT approach leads to a significant improvement in the VaR forecasting performance. Overall, among the six forecasting models, the RG-EVT model provides the best forecasts of daily VaR. Originality/value To the best of the authors’ knowledge, this is the earliest implementation of the RGARCH model within the conditional EVT framework. Additionally, the authors use a data set with a reasonably long sample period (around 11 years) in the context of high-frequency data-based forecasting studies. More significantly, the data set has a cross-sectional dimension that is rarely considered in the existing VaR forecasting literature. Therefore, the findings are likely to be widely applicable and are robust to the data snooping bias.