Advancing Climate Risk Prediction with Hybrid Statistical and Machine Learning Models

Accurately predicting climate-related risk requires models that balance high predictive accuracy with strong calibration and interpretability. This study evaluates four predictive approaches: Stacked Ensemble (Stack), Extreme Gradient Boosting (XGB), Random Forest (RF), and Generalized Additive Models (GAM) using a simulated climate dataset. Model performance was assessed through AUROC, AUPRC, Brier Score, LogLoss, and calibration metrics (ECE, reliability slope, and intercept). The Stack model achieved the highest AUROC (0.864) and AUPRC (0.559) while maintaining the lowest Brier Score (0.146) and ECE (0.018), demonstrating superior discrimination and probability calibration. Feature importance analysis revealed that temperature anomaly (0.127), precipitation (0.103), and greenhouse gas concentrations (0.089) were the most influential predictors, alongside seasonal and geographic factors. Spatial risk mapping and observed predicted comparisons confirmed the Stack model's ability to capture both the magnitude and distribution of high-risk zones with minimal error (< 0.015). These results highlight the potential of hybrid statistical–machine learning approaches for enhancing climate risk assessment and supporting targeted adaptation strategies.