Forecasting crude oil price volatility utilizing multivariate decomposition and deep learning approaches: Considering the role of influencing factors

This paper attempts to identify the main influencing factors of the realized volatility (RV) for West Texas Intermediate (WTI) crude oil spot price from three categories: commodity attribute factors, alternative energy factors, and macro uncertainty factors, so as to develop a multi-factor hybrid forecasting model. We integrated extremely randomized trees (ERT) with multivariate variational mode decomposition (MVMD) and temporal convolutional network (TCN) to construct an innovative ERT-MVMD-TCN model. Firstly, the ERT method is employed to select essential factors from candidate influencing variables as the main input features of the forecasting model. Secondly, MVMD is utilized to simultaneously decompose both the RV and the screened input variables into subsequences that reflect different scale features. Subsequently, the TCN with stable forecasting ability is applied to forecast each subsequence separately to effectively capture the time dependence. All individual forecasts are compiled to form the final prediction of RV. Finally, the forecasting performance of the model is evaluated through four loss functions: mean absolute error (MAE), mean square error (MSE), heteroscedastic adjusted MAE (HMAE), heteroscedastic adjusted MSE (HMSE) and modified Diebold and Mariano test (MDM). The empirical findings not only demonstrate the superior accuracy of the proposed ERT-MVMD-TCN model over the comparative single and hybrid models, but also maintain a robust predictive advantage across different forecast horizons and volatility periods. Hence, in the context of the current complex crude oil market, the proposed ERT-MVMD-TCN model presents a wide range of application prospects in crude oil price volatility prediction, which offers valuable insights for decision-making to both market participants and policymakers.