Predicting corporate bankruptcy through a scalable and interpretable knowledge distillation framework

Abstract

Abstract

Corporate bankruptcies trigger severe systemic market disruptions. Accurate risk identification remains a critical economic priority. However, existing prediction models struggle with extreme class imbalance. They also face structural noise and high computational costs. This research presents a knowledge distillation framework for bankruptcy prediction. The scope of the present study is strictly limited to structured, tabular financial ratios; integration of unstructured textual data, such as management discussion and analysis and earnings call transcripts, falls outside the purview of this framework and is deferred to future multimodal extensions. The architecture compresses a high-capacity teacher model into a lightweight student network. The methodology utilizes Taiwanese and Polish multivariate datasets. A hybrid mechanism combining LASSO and Mutual Information extracts optimal features. Advanced resampling strategies resolve severe topological class imbalance. These include SMOTE, SMOTEENN, and ADASYN. A blended distillation objective transfers the predictive patterns. This loss formulation mathematically combines temperature-scaled soft probabilities with ground-truth hard labels. The student then employs Histogram Gradient Boosting for rapid inference. Empirical results demonstrate superior predictive fidelity and efficiency against seventeen baselines. On the Taiwanese dataset using SMOTE, the student model achieves 93.77% Accuracy, 69.05% F1, and 93.17% AUC. On the Polish dataset with SMOTE, it delivers 86.79% Accuracy, 63.05% F1, and 88.74% AUC. Across these evaluations, the student model yields merely 0.67% to 1.51% less in F1 score than the teacher model. Crucially, the student model reduces overall computational training time by over 80%. Training consistently completes in under just four seconds. SHAP and LIME algorithms provide essential global and local interpretability. The distilled framework successfully balances predictive accuracy with significantly reduced algorithmic complexity. Ultimately, it offers a scalable, transparent, and highly efficient solution for real-world corporate financial risk assessment.