Forecasting Covariance Matrices: A Mixed Frequency Approach

In this paper we introduce a new method of forecasting covariance matrices of large dimensions by exploiting the theoretical and empirical potential of using mixed-frequency sampled data. The idea is to use high-frequency (intraday) data to model and forecast daily realized volatilities combined with low frequency (daily) data as input to the correlation model. The main theoretical contribution of the paper is to derive statistical and economic conditions, which ensure that a mixed-frequency forecast has a smaller mean squared forecast error than a similar pure low-frequency or pure high-frequency specification. The conditions are very general and do not rely on distributional assumptions of the forecasting errors or on a particular model specification. Moreover, we provide empirical evidence that, besides overcoming the computational burden of pure high-frequency specifications, the mixed-frequency forecasts are particularly useful in turbulent financial periods, such as the previous financial crisis and always outperforms the pure low-frequency specifications.