Bankruptcy prediction using cost-sensitive multi-objective feature selection and imbalance-XGBoost

Bankruptcy prediction remains a central problem in finance and decision analytics, where practitioners require models that are not only accurate but also operationally feasible under asymmetric error costs and severe class imbalance. Existing approaches often address imbalance through re-sampling, treated separately from the classifier, and they frequently rely on high-dimensional sets of financial ratios that increase data acquisition costs and reduce interpretability. This study adopts an operational research perspective and proposes an integrated, cost-sensitive framework that jointly optimises feature selection and classification for bankruptcy prediction. We extend XGBoost to incorporate a cost-sensitive learning objective for imbalanced data and combine it with multi-objective evolutionary feature selection that simultaneously minimises (i) feature subset size and (ii) economically meaningful misclassification cost, defined using loss given default and opportunity cost. Using an imbalanced dataset of US firms spanning multiple time periods and economic cycles, we conduct a comparative evaluation against re-sampling-based ensemble learning benchmarks. The results indicate that the proposed MOEFS + cost-sensitive XGBoost pipeline achieves superior cost-oriented performance and improves minority-class detection while retaining compact, operationally deployable feature sets. Overall, the findings demonstrate that embedding cost objectives within both the selection and learning stages provides a decision-relevant alternative to conventional, metric-driven bankruptcy prediction models.