Predicting Carbon Credit Trading Decisions Using Machine Learning

Abstract

Climate change has long been a global threat, this situation necessitated effective emission reduction strategies. The Carbon Emission Trading System (ETS) serves as a core mechanism for enterprises targeting carbon neutrality. Faced with significant challenges from price volatility and changing regulations in the carbon credit market, making the decision to buy or sell becomes more difficult. In this context, machine learning was considered a potential method for identifying market emission factors affecting trading decisions. This study aims to identify the factors influencing BUY or SELL decisions and evaluate the predictive capability of machine learning models. The analysis used a carbon trading dataset containing approximately 5,000 records. Additionally, The target variable has been relabeled based on emission logic and price signals to reflect market behavior better. Key features related to emissions and price volatility have been constructed. The study applies the CRISP-DM process for data normalization, feature selection using RFECV, and training five classification models. The results indicate that tree-based models deliver the highest performance, XGBoost achieved the best stability after hyperparameter tuning, leading in Accuracy, F1- score, and ROC-AUC metrics. The results indicated that machine learning models effectively captured the relationship between emission states, price volatility, and trading behaviors. This capability supported reliable BUY and SELL predictions. Although the achieved performance was high, the model's scalability remains limited due to the singlesource dataset. The data used did not fully reflect actual market conditions. In the future, the model could be assessed using a broader range of data and market simulations. This process would help determine its effectiveness in supporting decisions.