Forecasting Covariance Matrices: A Mixed Approach

Abstract

In this article, we introduce a new method of forecasting large-dimensional covariance matrices by exploiting the theoretical and empirical potential of mixing forecasts derived from different information sets. The main theoretical contribution of the article is to find the conditions under which a mixed approach (MA) provides a smaller mean squared forecast error than a standard one. The conditions are general and do not rely on distributional assumptions of the forecasting errors or on any particular model specification. The empirical contribution of the article regards a comprehensive comparative exercise of the new approach against standard ones when forecasting the covariance matrix of a portfolio of thirty stocks. The implemented MA uses volatility forecasts computed from high-frequency-based models and correlation forecasts using realized-volatility-adjusted dynamic conditional correlation models. The MA always outperforms the standard methods computed from daily returns and performs equally well to the ones using high-frequency-based specifications, however at a lower computational cost.