Predicting China's thermal coal price: Does multivariate decomposition-integrated forecasting model with window rolling work?

Coal, as the primary energy source in China, significantly affects the country's energy security and national economic stability. However, the highly nonlinear and non-stationary nature of coal prices poses challenges for accurate forecasting. In this study, we propose the Rolling ICEEMDAN-Methods series model based on the "divide and conquer" approach to predict the Bohai-Rim Steam-Coal Price Index (BSPI), involving the integration of multiple methods, including ANN, CNN, LSTM, GRU, LightGBM, and ERT. Unlike conventional univariate forecasting, we comprehensively summarise the factors influencing coal prices into eight categories, totalling 27 variables, with the aim of capturing more meaningful information. By employing the window-rolling decomposition-ensemble forecasting method, we effectively avoided information leakage and boundary effects, leading to a significant improvement in prediction accuracy. Experimental results demonstrate that the proposed Rolling ICEEMDAN-Methods outperforms other Rolling Methods in terms of accuracy and stability. Novel variables, such as attention, and the other seven categories of influencing factors contribute to enhanced prediction accuracy, among which past coal prices exhibit higher importance in determining forecast results. The findings offer valuable guidance to coal enterprises in making production decisions and provide a basis for the government to formulate macroeconomic energy policies.