Enhancing credit risk models with machine learning: Cash advance prediction model and an early warning system model for default risk

This dissertation consists of two papers.The first paper, “Enhancing Credit Risk Assessment with Machine Learning: An Early Warning Model for Customer Default Prediction,” presents an advanced Early Warning System (EWS) for predicting retail bank customer default risk using machine learning. Utilizing a proprietary dataset drawn from 159 million observations and comprising 631 customer-specific variables, the model achieved a predictive accuracy of 86.15% during the modeling phase and maintained high performance one year post-modeling. These results demonstrate the model's robustness, reliability, and adaptability across varying economic conditions. Key findings indicate that higher overdraft usage, greater overall liabilities, recent payment delays, lower credit limits, and younger customers are associated with an increased likelihood of default. In contrast, longer credit histories and higher approved credit limits are linked to lower default risk. Additionally, customers with a higher predicted risk of default tend to acquire more new financial products and exhibit reduced transaction activity in demand deposits compared to others.The second paper, “What Drives Cash Advance Usage? Unveiling Cash Advance Usage Patterns Through Machine Learning,” investigates the determinants of cash advance usage—an underexplored but critical component of short-term credit. Utilizing a comprehensive dataset of seven million financial consumers and 180 variables, this study introduces a novel methodological framework that combines rolling-recursive sample splitting with advanced machine learning techniques, including XGBoost and SHAP analysis. This approach effectively captures the evolving nature of consumer financial behavior, enabling robust and adaptable predictive modeling. The model demonstrates exceptional performance, with an AUC of 0.90 and a recall of 0.88. Findings reveal that cash advance usage is significantly influenced by both financial factors—such as overdraft balances and credit limits—and demographic factors, including education level, gender, and age. These results highlight socio-economic vulnerabilities and the interplay between financial distress and demographic disparities in shaping consumer reliance on high-cost credit. Beyond academic contributions, this research offers actionable insights for financial institutions, providing a foundation for advanced credit risk models and targeted interventions. It also fills a vital gap in the literature and underscores the transformative potential of applying machine learning in consumer finance analytics.