Comparative Analysis of Ensemble Gradient Boosting Classifiers for Credit Default Prediction using ML Optimization Framework

The growing use of ML (machine learning) methods in credit risk evaluations has exacerbated the trade-off among accuracy in prediction and model understanding, posing challenges to compliance with regulations under frameworks such as the Europe’s General data protection regulation (GDPR) and Basel III/IV guidelines. This study compares five machine learning classifiers—logistic regression, random forest, gradient boosting machine, extreme gradient boosting (XGBoost), and Light gradient boosting machine (LightGBM) by using the statistically calibrated credit risk database designed after benchmarking German Credit dataset.

The study helps in three ways. Firstly SMOTE enhanced training framework used to address the class disparity found in credit portfolios. Secondly, Shapely Additive Explanations (SHAP) used in both local as well as global levels to assess the impact of borrower attributes on estimated default probability. Third, Kolmogorov-Smirnov (KS) separation analysis and cost-sensitive threshold optimization used to match classification judgments to asymmetrical monetary loss functions.

The results showed Logistic Regression model alongside calibrated based regularization methods had the best discriminatory performance of each model tested, with an area under receiver operating characteristics curve, AUC of 0.6516 under cross-validation while retaining full interpretability and auditability. LightGBM provides in-depth insights on feature interactions, mainly combined effect of borrower age and credit exposure on default risk. Under a cost structure in which false-negative errors were penalized fivefold, the best classification threshold stabilized at 0.317, lowering projected misclassification costs by 23% compared to the traditional threshold of 0.5.

The results obtained indicate that interpreted ML techniques can aid in effective credit risk assessment while increasing transparency and enabling regulatory compliance.