Mixed-frequency data-driven forecasting of thermal coal price: A novel hybrid model

As the global energy crisis and the acceleration of energy transition become the focus worldwide, it is crucial to analyze and forecast energy commodity prices accurately to maintain energy economic market security. The forecasting of thermal coal prices poses substantial challenges due to disparities in sampling frequencies of various influencing factors. This paper proposes a hybrid model, termed CVM-Transformer, that integrates Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN), Vector Auto Regression (VAR), Mixed Data Sampling (MIDAS), and Transformer to achieve accurate, responsive, and interpretable thermal coal price forecasts by utilizing mixed-frequency data throughout the entire coal supply chain. Empirical and experimental results demonstrate periodic fluctuation patterns of thermal coal prices across different time scales, with short-term dependence on price analysis, medium-term dependence on supply and demand dynamics, and long-term dependence on the total social inventory level. The application of CEEMDAN, VAR, and MIDAS enables responsive and interpretable forecasts by using mixed-frequency data up to the current moment, and contributes to accuracy enhancement by 29.91 %, 30.72 %, and 21.60 %, respectively. The proposed CVM-Transformer model achieves a comprehensive improvement in forecasting accuracy by 66.42 %, providing a dependable basis for coal procurement decision-making and valuable insights for stakeholders in the coal industry.