CSA-Net: a cost-sensitive adaptive hybrid framework for imbalanced financial distress prediction

Financial distress prediction is a severely imbalanced tabular-learning problem in which distressed firms are rare and false-negative errors can carry disproportionate economic consequences. We propose CSA-Net, a cost-sensitive adaptive hybrid framework that treats LightGBM as a strong tabular prior and learns a lightweight neural residual whose sample-specific amplitude limits unnecessary deviation from the tree predictor. An error-aware objective places additional weight on distressed observations that the prior scores poorly. Eleven statistical, ensemble, deep-tabular, and neural models were evaluated across five Polish bankruptcy horizons and an independent Taiwanese benchmark using five-fold outer evaluation, three-fold inner model selection, matched 12-candidate tuning budgets, fold-local preprocessing, and corrected paired inference. AUC-PR was prespecified as the primary endpoint. CSA-Net ranked first by mean AUC-PR in five of six settings. Relative to standalone LightGBM, corrected differences were significant at four of five Polish horizons, while the Taiwanese difference was positive but not statistically significant; the remaining Polish horizon was effectively tied. Ablation showed that generic residual stacking contributed little, whereas adaptive and cost-sensitive correction accounted for most of the incremental gain. Reliability analyses further showed competitive calibration, behaviorally validated post-hoc attributions, dataset-dependent robustness, and moderate computational overhead. These findings support targeted neural augmentation of strong tabular boosting rather than universal model superiority.