Predicting and Explaining Economic Losses from Natural Disasters: A Post-Hoc Machine Learning and Explainable AI Approach

Abstract

This study develops a post hoc machine-learning and explainable artificial intelligence framework for predicting and interpreting the economic losses associated with natural disasters using 5051 records from the EM-DAT database covering 1975–2025. To address dependence among country-level records belonging to the same multi-country disaster, validation was performed at the disaster-event level rather than through conventional row-level random splitting. Random Forest, XGBoost, LightGBM, and CatBoost models were evaluated alongside statistical and median-based baselines. Under the event-grouped holdout design, tuned CatBoost achieved the strongest predictive performance (RMSLE = 1.890; R2 on the log scale = 0.491), while event-grouped five-fold cross-validation yielded an RMSLE of 1.851 ± 0.056. A strict temporal validation, in which hyperparameters were selected exclusively using pre-2011 data and events from 2011 onward were reserved for final testing, produced an RMSLE of 1.791 and an R2 of 0.464 for tuned CatBoost. Ablation analysis showed that removing geographic identifiers increased RMSLE by 20.9%, whereas removing the non-overlapping human-impact variables increased RMSLE by 16.9%. SHAP analyses consistently identified subregion, country-level location identifiers, total deaths, and total affected population among the most influential predictors. These geographic variables are interpreted as predictive location identifiers that may capture multiple forms of unobserved spatial heterogeneity rather than as direct measurements of structural vulnerability. Leave-one-region-out and leave-one-disaster-type-out experiments further indicated reduced generalisation when entire geographic or hazard domains were withheld. Overall, the results demonstrate that event-aware validation materially strengthens the methodological reliability of post hoc disaster-loss modelling while highlighting persistent limitations in cross-domain transferability and catastrophic-loss prediction.