As a pivotal global commodity, crude oil price volatility directly impacts economic stability and strategic security. Being the most widely traded asset worldwide, it also serves as a key financial barometer and a critical transition fuel in the shift towards renewable energy. Nevertheless, accurate forecasting of crude oil prices remains challenging due to three persistent challenges: (1) the lack of a systematic method to filter out redundant and noisy features for deep learning models; (2) the limited ability of existing models to simultaneously capture both local bidirectional dependencies and global periodic patterns; and (3) the non-adaptive nature of conventional attention mechanisms, which restricts their capacity to dynamically focus on the most informative historical periods. To bridge these gaps, this study introduces a novel forecasting framework with three key contributions. First, we introduce a hierarchical feature selection paradigm based on LightGBM to systematically eliminate data redundancy and noise, thereby constructing an optimal feature subset for subsequent deep modeling. Second, an improved Autoformer encoder, integrated with Bidirectional GRUs, is designed to simultaneously capture local bidirectional dependencies and global periodic patterns, enabling a more comprehensive multi-scale temporal representation. Third, a dynamic fusion mechanism is incorporated to adaptively recalibrate the significance of historical timesteps. This enables the model to focus on periods rich in information, enhancing contextual awareness in predictions. Future research aims to enhance forecasting capabilities by achieving a deeper integration of local and global temporal representations, potentially through exploring advanced gating or sparse attention mechanisms.
Zhang et al. (Fri,) studied this question.