Value at Risk Forecasting

The purpose of this paper is to investigate whether a dynamic Value at Risk model and high frequency realized volatility models can improve the accuracy of 1-day ahead VaR forecasting beyond the performance of frequently used models. As such, this paper constructs 60 conditional volatility forecasting models. Several extensions of the GARCH model are included, such as the nonlinear and asymmetric models. Moreover, several return distributions are assumed for the error term, in order to allow for more flexible modeling in the tails. A rolling Model Confidence Set is subsequently constructed, ensuring that only models with superior out-of-sample forecasting performance remain. A model averaging technique is applied to the remaining superior models, which generates the dynamic VaR forecasts. Moreover, several extensions of the HAR realized volatility model are included in this paper to forecast VaR. In a series of extensive back tests, this paper finds that the dynamic VaR model produces forecasts which are superior to traditional models and HAR models. The result hold for the 95% VaR, but are even more pronounced for the 99% VaR. The traditional models severely underestimate risk at higher confidence levels, whereas the applied dynamic VaR correctly accounts for it