A Multiscale Decomposition-Ensemble Framework with Explainable AI for Carbon Price Forecasting and Driver Analysis

Abstract

Accurate carbon price prediction and fluctuation analysis are essential for carbon market risk management and achieving carbon neutrality goals. However, carbon prices exhibit complex multi-scale nonlinear dynamics intertwined with time-varying external factors, hindering reliable forecasting. This study constructs a hybrid prediction framework integrating CEEMDAN, Sample Entropy (SE) reconstruction, and the Coefficient of Variation (VC) ensemble algorithm. SHapley Additive exPlanations (SHAP) quantifies scale-specific nonlinear factor contributions, while TVP-SV-VAR captures dynamic carbon price-driver correlations. Empirical results indicate that the CEEMDAN-SE preprocessing strategy significantly improves the prediction accuracy of baseline models. The CEEMDAN-SE-GRU model achieves optimal performance, with an R2 of 0.96 and over 60% reductions in both MSE and MAE relative to the baseline GRU model. Meanwhile, the VC ensemble outperforms single models and alternative fusion strategies. SHAP identifies scale-heterogeneous drivers: short-run prices follow sentiment and macro outlooks, medium-run trends tie to industrial costs and global markets, long-run paths align with energy transition and global climate governance. The TVP-SV-VAR model uncovers significant time-varying spillover effects on raw carbon prices. Based on these findings, we recommend establishing multi-layered dynamic monitoring and early warning systems, constructing a differentiated and adaptive policy toolkit, refining cross-market risk isolation and buffering mechanisms, and advancing institutional improvements through gradual implementation.