Improving variance forecasts: The role of Realized Variance features

In this paper, we effectively extend the Realized-EGARCH (R-EGARCH) framework by allowing the conditional variance process to incorporate exogenous variates related to different observable features of Realized Variance (RV). The choice of these features is well motivated by recent studies on the Heterogeneous Autoregressive (HAR) class of models. We examine several specifications nested within our augmented R-EGARCH representation, and we find that they perform significantly better than the standard REGARCH model. These specifications incorporate realized semi-variances, heterogeneous long-memory effects of RV, and jump variation. We also show that the performance of our framework further improves if we allow for skewness and excess kurtosis for asset return innovations, instead of assuming normality. This can better filter the true distribution of the return innovations, and thus can more accurately estimate their effects on the variance process. This is also supported by a Monte Carlo simulation exercise executed in the paper.& COPY; 2022 Published by Elsevier B.V. on behalf of International Institute of Forecasters.