Combining information at different frequencies in multivariate volatility prediction

In the last two decades the literature has been focusing on the development of dynamic models for predicting conditional covariance matrices from daily returns and, more recently, on the generation of co-volatility forecasts by means of dynamic models directly fitted to realized measures. Despite the number of contributions on this topic same open issue still arise. First, are dynamic models based on realized measures able to produce more accurate forecasts than standard MGARCH models based on daily returns? Second, which is the impact of the choice of the volatility proxies on forecasting accuracy? Is it possible to improve the forecasts accuracy by combining forecasts from MGARCH and models for realized measures? Finally, combining information observed at different frequencies can help to improve over the performance of single models? In order to gain some insight about these research questions, in this paper we perform an extensive forecast comparison of different multivariate volatility models considering both MGARCH models and dynamic models for realized covariance measures. Furthermore, we investigate the possibility of increasing predictive accuracy by combining forecasts generated from these two classes of models, using different combination schemes and mixing forecasts based on information sets observed at different frequencies.