Predicting non-life insurers’ financial distress: Evidence from Bangladesh

Abstract

This study develops an interpretable early-warning framework to predict financial distress among non-life insurers in Bangladesh. A 2014–2024 firm-year panel is utilized to compare penalized logistic regression, random forest, and gradient-boosted trees, apply class-balance remedies, and map model decisions with SHAP. Gradient-boosted trees deliver the best out-of-time recall; SHAP consistently identifies management expense ratio, lagged underwriting performance, and reinsurance intensity as the strongest predictors. Robustness checks across resampling schemes and feature reductions confirm the stability of these operational signals. The results imply that supervisors in thin-premium markets should prioritize expense and underwriting monitoring and calibrate alarm thresholds to favor sensitivity. This research provides a compact, policy-ready pipeline that balances predictive performance with transparency.