Explainable SARIMA Ensemble for Coastal Monthly Rainfall Time-Series Forecasting with Social Choice Integration

Accurate monthly rainfall forecasting is essential for water resources management in climate-vulnerable coastal regions, yet the 710 km long coastal region of Bangladesh, home to over 29% of the country’s population, has received little attention in recent forecasting studies. This study develops a hybrid ensemble learning-based seasonal autoregressive integrated moving average (SARIMA) model, termed XRMBSS, for Cox’s Bazar, Bangladesh. Five machine learning (ML) models—extreme gradient boosting (XGBoost), random forest (RF), multilayer perceptron (MLP), Bayesian ridge (BR), and support vector machine (SVM)—were stacked, and the stacked predictions were combined with SARIMA predictions using four weighting schemes (XRMBSS-I to XRMBSS-IV). Long-term monthly rainfall data (1950–2023) from the Bangladesh Meteorological Department were split chronologically into training (1950–2000) and testing (2001–2023) periods (approximately 7:3). Performance was evaluated using the root mean squared error (RMSE), mean absolute error (MAE), Kling–Gupta efficiency (KGE), and coefficient of determination (R2), and the models were ranked using the Borda count social choice method. XRMBSS-IV outperformed all other models in the testing phase (RMSE = 29.0287 mm, MAE = 17.7612 mm, KGE = 0.9711, and R2 = 0.9939). Conventional and SHapley Additive exPlanations (SHAP)-based sensitivity analyses identified Month_Rainfall_Interaction as the most influential predictor. Re-training the same pipeline at two independent stations (Barisal and Khulna) gave testing R2 values of 0.9919 and 0.9868, suggesting that the framework can be adapted to other stations.