Credit Risk Default Prediction

My Contribution

Helped build the reproducible Python pipeline, engineer applicant-level features, evaluate threshold sensitivity and lift, interpret SHAP results, audit subgroup performance, and translate the analysis into a practical recommendation.

Feature Pipeline

Transformed millions of historical records into one row per applicant. Train-only preprocessing and feature selection reduced 4,609 engineered affordability, repayment, credit-burden, recency, missingness, and cross-table features to 2,701.

Modeling

Used a stratified 60/20/20 train, validation, and frozen-test split. Tuned a heavily regularized XGBoost model with Optuna and three-fold cross-validation, then calibrated the risk scores with Platt scaling.

Model Performance

Achieved 0.8002 ROC-AUC, a +0.0305 improvement over the application-only baseline. The highest-risk 20% of applicants contained 59.5% of repayment-difficulty cases; the highest-risk 10% contained 39.3%, a 3.93x lift over random selection.

Threshold Finding

At a 0.50 probability cutoff, the model missed 94.8% of repayment-difficulty cases. That result showed why threshold selection should reflect loss severity, review capacity, and business costs rather than rely on a conventional cutoff.

Responsible Use

SHAP and subgroup audits surfaced governance concerns, including gender among influential features and meaningful performance gaps across applicant groups. We recommended reviewing the highest-risk 10% to 20%, supported by human judgment, sensitive-feature testing, fairness monitoring, adverse-action explanations, and applicant recourse.

Focus areas

• Machine Learning
• Credit Risk
• XGBoost
• Python