Evaluating the performance of volatility forecasts with the aid of statistical criteria

This paper focuses on the use of statistical criteria for evaluating the forecasting performance of volatility models. The empirical analysis included eight volatility models, ranging from the IGARCH to stochastic volatility, and produced out-of-sample forecasts for five stock indices considering two distinct time intervals: the crisis of 2007-2009 and the recovery period of 2012-2014. Individual rankings of the forecasts showed that evaluation results are heavily impacted by the choice of criteria, the choice of volatility proxies and the considered time intervals. On the other hand, the average rank indicated the superiority of asymmetric models in the case of stock indices, as well as the superiority of models based on heavy-tailed distributions relative to those with Gaussian errors. Aggregating the results, EGARCH and stochastic volatility emerged as the most accurate forecasts but the statistical criteria employed in this study were not able to delimit clearly the best of the two models.