A context-aware dynamic ensemble interval-valued crude oil price forecasting system

Crude oil price forecasting is crucial for economic stability and strategic planning; however, traditional point-value and static models fail to capture the market’s volatility and complexity. Existing approaches often rely on simple data processing, limited frequency models, and static ensembles; moreover, these methods typically lack interpretability. To overcome these limitations, this study proposes an interval-valued ensemble forecasting system that effectively leverages multi-source mixed-frequency interval-valued information. The proposed system introduces three key innovations: a multi-stage feature engineering process that enriches information from diverse sources and overcomes the limitations of individual feature-processing algorithms; a multi-source mixed-frequency model that enables the joint processing of multiple high-frequency data streams; and a context-aware dynamic ensemble module powered by a dual-attention mechanism, which adaptively adjusts base model contributions according to prevailing market conditions while enhancing the interpretability of the developed system. Experimental evaluations show that the proposed system outperforms several advanced ensemble and state-of-the-art models, achieving Interval Root Mean Squared Errors (IRMSE) of 4.1763 and 4.5703 on West Texas Intermediate (WTI) and Brent crude oil datasets, respectively. Compared with the state of the art models, it delivers improvements ranging from 23.6% to 81.2% on the WTI dataset and from 14.2% to 76.8% on the Brent dataset. Interpretability analysis reveals the dynamic ensemble adapts its strategy intelligently, favoring different models during high volatility and strong market trends. This study thus offers a robust, transparent forecasting system for more informed and reliable decision-making.