Exploring long-run CO2 emission patterns and the environmental kuznets curve with machine learning methods

This study utilizes machine learning methods to forecast long-term CO 2 emission trends and employs interpretable machine learning techniques to reexamine the Environmental Kuznets Curve (EKC) hypothesis. All methods we experimented with, including Random Forest, XGBoost, Support Vector Regression, and Artificial Neural Networks (ANN), demonstrated remarkable forecasting power, with ANN outperforming all others. Using SHapley Additive exPlanations (SHAP), we identified an EKC relationship within the ANN model, consistent with linear regression findings. Moreover, long-run forecasts for the next decade indicate that CO 2 emissions growth in high-income economies tends to approach zero. In contrast, low-income countries may maintain relatively high emissions, but such predictions exhibit uncertainty due to significant discrepancies among the algorithms. • Use five machine learning algorithms for long-term CO2 emissions forecasting. • Identify ANN's excellent performance with a limited dataset. • Utilize SHAP to explore model predictions across different country categories. • Validate the EKC hypothesis from a machine learning perspective. • Identify heterogeneous relationships among countries.