A Machine Learning Ensemble Framework for Carbon Price Prediction and Decision Support Under Information Structure Heterogeneity in Regional Carbon Markets in China

Unofficial AI-generated summary based on the public title and abstract. Not an official translation.

🔬研究者:Provides a robust ensemble ML framework for carbon price forecasting that accounts for market heterogeneity, useful for researchers in carbon finance and applied machine learning.

🏢実務担当者:Corporate sustainability teams can use the forecasting approach for allowance trading strategies and cost impact assessments of carbon prices on projects like offshore wind.

🏛政策担当者:Policymakers can see how ML techniques enhance transparency and predictability in fragmented carbon markets, informing the design of unified national systems.

Reliable prediction of carbon allowance prices plays a crucial role in emissions trading systems, particularly for market participation, regulatory compliance, and long-term cost planning. In China, regional carbon markets differ markedly in trading activity, price formation mechanisms, and responsiveness to external signals, which limits the effectiveness of conventional single-model forecasting approaches. This study develops a unified machine learning framework designed to accommodate such cross-market heterogeneity. The framework incorporates a diverse set of explanatory variables, including historical price-based indicators, trading volume information, inter-market linkage signals, and macroeconomic factors. Three ensemble-based learning algorithms-XGBoost, LightGBM, and Random Forest—are implemented, and their outputs are further integrated using a weighted aggregation scheme to improve generalization across markets. The empirical evaluation across seven pilot markets shows that, while LightGBM consistently performs well as a standalone model, the proposed ensemble framework achieves superior stability and adaptability under varying market conditions. The forecasting accuracy is high across all cases, with coefficients of determination above 0.74 and reaching values greater than 0.92 in most markets. Further investigation through feature ablation highlights the heterogeneous role of external information, indicating that predictor importance varies significantly between markets and that no universal feature combination yields optimal performance. Leveraging the forecast outputs, the study also demonstrates practical applications in decision support, including timing strategies for allowance sales and dynamic cost assessment in offshore wind engineering scenarios. By systematically evaluating the marginal contribution of different information groups to predictive uncertainty, the framework offers a flexible tool for managing information-structure uncertainty in fragmented carbon markets. The proposed framework offers an integrated solution that connects predictive modeling with operational and engineering decision on processes, providing a flexible tool for managing uncertainty in fragmented carbon markets.

• openalex https://doi.org/10.3390/e28060656first seen 2026-06-13 05:04:17