Liquidity and realized covariance forecasting: a hybrid method with model uncertainty

This paper investigates the realized covariance forecasting and liquidity effects on the covariance. The realized covariance is calculated based on the high frequency data of CSI 300 stock index and futures, and nonlinear support vector regression (SVR) approach is employed to evaluate the out-of-sample forecasting ability of HAR-type models. Then, we propose the hybrid method, named the weighted average windows (WAveW) method based on both OLS and SVR forecasts, to accommodate model uncertainty. The empirical results find that the performance of the WAveW method based on SVR forecasts obtains more accurate forecasting than the OLS and SVR methods, and the incorporation of liquidity helps to improve the forecasting ability. From the portfolio selection perspective, we show that our new method achieves higher economic value, which further confirms the effectiveness of our proposed hybrid method. The results are robust under alternative rolling windows, liquidity, covariance and cojumps.