Title: Combining multiple frequencies in multivariate volatility forecasting
Authors: Alessandra Amendola - University of Salerno (Italy) [presenting]
Vincenzo Candila - Sapienza University of Rome (Italy)
Giuseppe Storti - University of Salerno (Italy)
Abstract: In a multivariate volatility framework, several options are available to estimate the conditional covariance matrix of returns. Some models, like the multivariate GARCH (MGARCH) ones, rely on daily returns while others exploit the additional information provided by intra-daily prices, like the realized covariance (RC) specifications. A first question arises in the choice of a superior model, for a given period and dataset. An additional source of uncertainty is related to the selection of the frequency at which the intradaily returns, used to construct the RC matrices, are observed. In order to overcome these issues, we propose a prediction strategy based on the combination of multivariate volatility forecasts coming from different model structures, estimated using information at various frequencies. More specifically, we adopt a Model Confidence Set (MCS) procedure applied on rolling forecasts under different loss functions (LFs). Some of the selected LFs rely on the economic evaluation criteria while other belong to the class of robust statistical LFs. The conditional covariances of those models entering the MCS are then equally or proportionally weighted, in order to obtain the combined predictor. Empirical findings give evidence that in an out-of-sample perspective the combined predictor always belongs to the set of superior models.