Realized Volatility Forecasting for Stocks and Futures Indices with Rolling CEEMDAN and Machine Learning Models

As an essential index for measuring market risk, realized volatility (RV) possesses mixed features and volatility aggregation, which makes it difficult for machine learning (ML) models to identify its features and trends directly for accurate prediction. Hence, this study first uses the rolling CEEMDAN (complete ensemble empirical mode decomposition with adaptive noise) method to decompose the original RV sequence of the major stock market indices as well as the bean and the metal futures indices. Furthermore, the study employs eight ML methods to predict the decomposed sub-sequences. Finally, the prediction outcomes for the sub-sequences are reconstructed. Through 1-step, 5-step, and 22-step predictions, various evaluation criteria, and the Diebold-Mariano test, the study find that the hybrid CEEMDAN-RF model possesses a better RV prediction ability than the non-hybrid models. Taking the Shanghai Stock Exchange as an example, the mean square error (MSE) evaluation criterion of the RF model is reduced by 52.07% after introducing the CEEMDAN decomposition method. Compared with the other hybrid models, the improvement percentages of the MSE for the hybrid CEEMDAN-RF model are 2.56%, 0.72%, 5.03%, 36.89%, 44.14%, 44.15%, and 58.19%, respectively. The empirical findings indicate that the hybrid CEEMDAN method and ML model can provide accurate, robust results for RV prediction of stocks indices and futures indices.