Abstract
<jats:p>Road traffic accident remains a major public-health and transportation challenge in Bangladesh, while existing studies often rely on isolated datasets, random validation, point predictions, and descriptive rankings that do not distinguish absolute burden from exposure-adjusted risk. This study develops an integrated, leakage-safe, uncertainty-aware, and policy-oriented framework using four nationwide datasets covering population characteristics, accidents, fatalities, injuries, vehicle involvement, and vehicle-specific fatalities across eight administrative divisions from January 2023 to December 2025. Linked division–month and vehicle–division–month panels were constructed with population-normalized indicators, cyclical seasonality, lagged dynamics, rolling variability, and momentum features. Mean and seasonal-naïve baselines, Ridge, Poisson, Tweedie, Random Forest, Extra Trees, Gradient Boosting, Histogram Gradient Boosting, and a hurdle model were evaluated using expanding-window validation and an independent 2025 holdout year. The framework further incorporated empirical-Bayes vehicle-risk stabilization, split-conformal prediction intervals, permutation importance, feature-family ablation, spatial concentration, hotspot persistence, and a transparent policy-priority index. National fatalities remained persistently high, while exposure-adjusted analysis identified Barishal and Sylhet as emerging-risk divisions despite the larger absolute burden in Dhaka and Chattogram. Motorcycles recorded the highest stabilized fatality risk, followed by auto rickshaws and other lightly protected vehicle categories. Random Forest achieved the lowest holdout fatality RMSE of 15.77, outperforming the seasonal-naïve benchmark by 26.7%, although count and regularized models remained competitive. Ablation analysis showed that the temporal-plus-demographic specification generalized better than the fully integrated feature set. The study contributes a reproducible decision-support architecture for seasonal enforcement, vehicle regulation, emergency-response allocation, infrastructure prioritization, and divisional monitoring. Prospective validation remains necessary before full operational deployment.</jats:p>