An Efficient Loan Default Risk Assessment Framework Using TPE-Optimized CatBoost Model

Abstract

Reliable estimation of loan default risk plays a vital role in ensuring financial system resilience and enabling sound credit decision-making in today’s lending landscape. Although conventional statistical techniques offer transparency, they frequently struggle to model the intricate nonlinear patterns embedded in high-dimensional financial datasets. To address this limitation, the present study introduces a robust framework for default risk evaluation that employs a CatBoost model optimized through a Tree-Structured Parzen Estimator (TPE). This methodology combines Bayesian-driven hyperparameter tuning with gradient boosting techniques to improve both predictive accuracy and the calibration of probability estimates. The performance of the proposed model is validated using three widely recognized benchmark datasets: German Credit Risk, Polish Companies Bankruptcy, and Taiwan Credit Card Default. The proposed approach is evaluated in comparison with several conventional machine learning models, namely Logistic Regression, Random Forest, and Support Vector Machine. The assessment is carried out using a diverse set of performance indicators, including ROC–AUC, recall, F1-score, Brier score, and the Kolmogorov–Smirnov (KS) measure. Empirical analysis across multiple datasets reveals that the CatBoost model enhanced through TPE-based optimization consistently delivers superior results relative to the benchmark methods, particularly in terms of classification strength, probability estimation accuracy, and effective differentiation of risk levels. Notably, the improvement is more pronounced in datasets characterized by class imbalance, emphasizing the advantage of the adopted optimization technique. The results indicate that combining advanced boosting methods with systematic hyperparameter tuning yields a dependable and scalable framework for credit risk evaluation, offering meaningful benefits to financial institutions through more precise and data-informed lending decisions.