Generalizable Machine Learning for Corporate Bankruptcy Prediction: An XGBoost-Based Framework

Accurate prediction of corporate bankruptcy is a fundamental and challenging problem due to the rarity of default events and the heterogeneity of financial data across firms, sectors, and countries. This paper proposes a machine learning framework based on the eXtreme Gradient Boosting (XGB) algorithm for early detection of financial distress in highly imbalanced settings. The inclusion of a scaling XGB parameter mitigates class imbalance, while maintaining interpretability through feature importance analysis. The approach is evaluated across three benchmark datasets (Taiwan, USA, and the Poland Bankruptcy datasets), which represent distinct economic contexts and temporal structures. The extensive experimental results demonstrate that XGB consistently outperforms traditional classifiers, such as Logistic Regression, Decision Trees, and Random Forests, achieving superior generalization across datasets and prediction horizons. Feature importance analysis further highlights profitability, leverage, and liquidity as the most stable and discriminative financial indicators of corporate distress. Overall, the proposed framework offers a robust and generalizable solution for bankruptcy prediction, supporting applications in risk assessment, credit scoring, and financial supervision.