SW-HyDEC: A structure-aware hybrid decomposition and ensemble learning framework for carbon price forecasting

Abstract Accurate carbon price forecasting is essential for market efficiency, effective climate policy design, and strategic decision-making by firms managing carbon assets. Carbon prices reflect a dynamic interplay of emission quotas, economic conditions, and policy expectations, directly influencing emission reduction incentives and investment in green technologies. However, existing models often treat data input variables independently and struggle with the non-stationary, multi-scale nature of carbon price signals. This paper proposes SW-HyDEC, a structure-aware hybrid framework that integrates Correlation-based Feature Selection (CFS), Recursive Feature Elimination (RFE), Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN), and a Hybrid-Kernel Extreme Learning Machine (Hybrid-KELM). To enhance the training convergence and robustness, model parameters are optimized using a Sobol-sequence enhanced Whale Optimization Algorithm (Sobol-WOA). Experimental results on real-world carbon market datasets demonstrate that SW-HyDEC significantly outperforms mainstream statistical, machine learning, and deep learning baselines in terms of accuracy, generalization, and stability.