An Explainable Machine Learning-Based Intelligent Early Warning System for Corporate Bankruptcy with Data Imbalance Analysis
An Explainable Machine Learning-Based Intelligent Early Warning System for Corporate Bankruptcy with Data Imbalance Analysis
Parisa GhahremanShahraki1
1) Computer Engineering
IOI Code
:
AAHC-ABKKCA
Publication :
National Conference on Management, Commercialization, Knowledge-Based Economy and Investment for Production(impeb.ir)
Abstract :
Corporate bankruptcy has far-reaching consequences for investors, creditors, employees, and economic stability. Therefore, the early identification of firms at risk can support timely decision-making and preventive intervention. This study aimed to develop an intelligent and explainable early warning system for corporate bankruptcy while simultaneously addressing class imbalance, missing data, redundancy among financial ratios, and model interpretability. The primary dataset was derived from the Polish companies bankruptcy dataset. After duplicate removal, its five-year file contained 5,850 companies, 64 financial ratios, and 408 bankrupt firms. The data were stratified into training and test sets. Missing values were imputed using the median values of the training data, and binary missingness indicators were created for variables with substantial missingness. After controlling for high correlations, 58 features, including 47 financial ratios and 11 missingness indicators, were retained for modeling. Logistic Regression, Random Forest, XGBoost, and CatBoost models were compared using class-weighting and SMOTENC strategies. The final weighted XGBoost model, using a decision threshold of 0.65, achieved an accuracy of 0.9658, precision of 0.7561, recall of 0.7561, F1-score of 0.7561, ROC-AUC of 0.9663, and PR-AUC of 0.8311 on the test set. SHAP analysis showed that sales growth, operating expenses, profitability, liquidity, capital structure, and missingness patterns in selected financial ratios played key roles in risk prediction. Moreover, the “critical-risk” group included 62 companies and captured 56 of the 82 actual bankruptcies. Independent implementation of the same framework on the Taiwanese dataset also yielded an ROC-AUC of 0.9579. These findings indicate that the proposed framework can serve as a decision-support tool for prioritizing financial reviews and designing early interventions.
Keywords : bankruptcy prediction early warning system explainable machine learning xgboost shap class imbalance financial ratios