Research Explores Explainable AI for Carbon Credit Price Prediction

Forecasting carbon credit prices is challenging, with existing research often focusing on specific markets and simplifying regulatory text. This study identifies ten gaps in current methodologies and proposes a new framework called EPA-CarbonNet. This architecture aims to provide explainable and policy-aware AI for carbon market predictions by integrating market time series data with policy text using cross-attention mechanisms. EPA-CarbonNet is designed to offer calibrated prediction intervals and policy-attributed explanations. The researchers built and tested this model using eleven years of daily S&P carbon index data. However, the initial findings present a mixed picture. While the model achieved 58.6% directional accuracy, outperforming baselines, its Root Mean Squared Error (RMSE) for five-day predictions was worse than a simple random walk. Furthermore, the policy attention mechanisms within the model did not consistently align with documented regulatory events, and SHAP rankings showed only moderate agreement across resampled backgrounds. This suggests that while the concept is valuable, significant challenges remain in effectively integrating and interpreting policy influence on carbon markets.