Covariance Forecasting Methods for Dynamic Asset Allocation

Abstract

The focus of this work involves a comparison of the performance of a covariance regression model (CovReg) and a vector heterogeneous autoregressive (VHAR) model for the development of a dynamic portfolio allocation framework in equity index strategies. The performance of each method is assessed based on various statistical and financial metrics such as Mean Square Error (MSE), Sharpe ratio, Omega, Sortino, and maximum draw-down. The forecast covariance matrices from each class of model are used to establish a global minimum variance (GVM) class of strategies for the asset allocation in the asset universe of global equity indices. To enhance these dynamic allocation methods further, we also incorporate a machine learning methodology based on early stopping which is used to reduce over-fitting and potentially protect against strategy sensitivities that result in practical shortcomings such as portfolio churn and frequent large re-balancing costs.