Explainable Credit Default Prediction for Microfinance: A Gradient Boosting and SHAP Study on Consumer Credit Data

Credit scoring models underpin lending decisions for millions of borrowers, yet in many microfinance settings they operate as black boxes — lenders know the score but not the story behind it. This paper trains and compares three classifiers (Logistic Regression, XGBoost, and LightGBM) on the Give Me Some Credit consumer credit dataset (150,000 borrowers, 6.68% default rate) and applies SHAP (SHapley Additive exPlanations) to make the best-performing model's decisions transparent. All three classifiers perform within a narrow AUROC band of 0.862–0.868, with XGBoost leading marginally. The SHAP analysis reveals that revolving credit utilisation and maximum delinquency status together drive the bulk of predictive signal — consistent with established credit theory. Younger borrowers and those with lower incomes face systematically higher predicted default risk, while revolving utilisation shows a notable threshold effect near 0.5. We also find that all three tree models are poorly calibrated, with direct implications for threshold-based lending decisions. These results support SHAP-augmented gradient boosting as a practical, explainable credit scoring framework for microfinance lenders operating under regulatory transparency requirements.