Forecasting large covariance matrix with high-frequency data using factor approach for the correlation matrix

Abstract

We apply the factor approach to the correlation matrix to forecast large covariance matrix of asset returns using high-frequency data, using the principal component method to model the underlying latent factors of the correlation matrix. The realized variances are separately forecasted using the Heterogeneous Autoregressive model. The forecasted variances and correlations are then combined to forecast large covariance matrix. Our proposed method is found to perform better in reporting smaller forecast errors than some selected competitors. Empirical application to a portfolio of 100 NYSE and NASDAQ stocks shows that our method provides lower out-of-sample realized variance in selecting global minimum variance portfolio.