Do high-frequency-based measures improve conditional covariance forecasts?

In this paper we investigate the possible benefits from using ex-post highfrequency based (realized) measures of volatility and correlation in conditional covariance forecasting. For this, we combine the (Robust) Realized GARCH framework with time varying conditional copulas and compare their forecasting abilities with those of multivariate Realized GARCH models and wellestablished competing models from the literature, i.e. the GJR-GARCH copula and the corrected DCC. The one-step-ahead forecasting abilities of the models are assessed in an empirical illustration on three pairs of financial assets by relying on the Model Confidence Set test. Our findings indicate that the proposed specifications relying on realized measures significantly improve the quality of covariance matrix forecasts.

JEL Codes: C32, C53, C58