Forecasting realized volatility of Chinese crude oil futures with a new secondary decomposition ensemble learning approach

Abstract

This study proposes a new VMD-ICEEMDAN-LSTM model, which combines secondary decomposition with long short-term memory neural networks (LSTM) to forecast the realized volatility (RV) of Chinese crude oil futures. The RV sequence is first decomposed into subcomponents and residuals through variational mode decomposition (VMD). Then, the iterative complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) is applied to perform a secondary decomposition on the residuals. Finally, we apply LSTM to forecast all decomposed components, and then combine all forecasting results to obtain our final forecast value. Our results show that the VMD-ICEEMDAN-LSTM model significantly outperforms existing individual and combination models.