Predicting Client Default Payments Using Machine Learning in Production Environment

Abstract

This project investigates the application of machine learning techniques to predict client default payments in a credit card setting. Using a dataset of 30,000 Taiwanese clients, the study addresses the challenges of class imbalance, predictive accuracy, and fairness in credit risk assessment. An XGBoost model was developed and enhanced through feature engineering, resampling techniques (SMOTE/ADASYN), and class weighting to improve recall for defaulters while maintaining overall accuracy. Interpretability was achieved using SHAP values, providing transparency into model decisions. To mitigate demographic disparities, particularly across education levels, a fairness-constrained Random Forest was integrated into a two-stage cascade framework, reducing false positives while preserving high recall. The final cascade model achieved 84% accuracy, with 93% recall for non-defaulters and 53% recall for defaulters, significantly outperforming baseline benchmarks. Fairness audits revealed that education-based disparities could be reduced with minimal performance trade-offs, while age-based fairness was largely maintained. The project demonstrates a practical, interpretable, and ethically aware pipeline for credit default prediction, with deployment considerations and directions for future research in cost-sensitive learning, advanced fairness constraints, and real-time monitoring