We adopt the semicovariance decomposition method and machine-learning models to forecast the realized correlation and realized volatility of oil and gold futures markets. Forecasting models are fitted with step-forward crossvalidation method, and evaluated using out-of-sample accuracy and robustness tests. In predicting the risk and correlation of oil and gold, the semicovariance decomposition method can achieve higher accuracy and stable results toward the various periods of training data compared to the benchmark DCC-GARCH model. In particular, a combination of random forest and the semicovariance decomposition method can significantly increase the out-of-sample accuracy. These results show a possibility that the semicovariance decomposition combined with machine-learning techniques can be effcient in predicting the realized correlation and realized volatility.​

Keywords: Realized Volatility, Forecasting, Machine Learning, Semicovariance decomposition
JEL classification: C32, C53, Q02