Using explainable deep learning to improve decision quality: Evidence from carbon trading market

Abstract

To achieve the United Nations Sustainable Development Goals (SDGs), reducing global greenhouse gas emissions is a top priority. Academia and industry have recognized the importance of carbon market management in promoting low-carbon development. However, traditional methods exhibit limitations in balancing accuracy and explainability, thereby reducing trust between users and decision-making models. To address this, we develop a data-driven model to enhance decision quality. Specifically, we evaluate and compare deep learning (DL) algorithms of various structures to explore the most appropriate techniques for modeling high-dimensional nonlinear carbon price data. Furthermore, we incorporate model-agnostic interpretation techniques to infer the contribution of the influencing factors to carbon prices. The results indicate that the predictive performance of the DL algorithm after feature selection and parameter optimization significantly improves. The findings reveal Internet big data and geopolitical risks as key features of carbon prices, complementing traditional indicators such as energy prices, economy, and climate, which exhibit lagged effects, regional heterogeneity, and interaction. These findings deepen our understanding of carbon price formation mechanisms and bolster managers’ ability to utilize artificial intelligence for effective decision-making, thereby supporting the achievement of the SDGs.