Combining Multivariate Volatility Models

Forecasting conditional covariance matrices of returns involves a variety of modeling options. First, the choice between models based on daily or intradaily returns. Examples of the former are the Multivariate GARCH (MGARCH) models while models fitted to Realized Covariance (RC) matrices are examples of the latter. A second option, strictly related to the RC matrices, is given by the identification of the frequency at which the intradaily returns are observed. A third option concerns the proper estimation method able to guarantee unbiased parameter estimates even for large (MGARCH) models. Thus, dealing with all these modeling options is not always straightforward. A possible solution is the combination of volatility forecasts. The aim of this work is to present a forecast combination strategy in which the combined models are selected by the Model Confidence Set (MCS) procedure, implemented under two economic loss functions (LFs).