credit-risk-prediction

•Built an end-to-end ML pipeline on 150,000 customer records, cleaning and engineering 13 final features while

handling a 93.3%/6.7% class imbalance.

•Trained and compared Logistic Regression, Random Forest, and XGBoost, selecting XGBoost for achieving the

highest recall (0.62) — critical for correctly identifying high-risk customers.

•Used SHAP to identify Total Times Late and Credit Utilization as the top drivers of default risk, improving model

transparency and trust.

•Deployed the solution using FastAPI and Streamlit.