Loan Default Risk Prediction with Fairness & Explainability

Problem Statement & Abstract

Banks and fintech lenders use machine learning to decide who gets credit, and regulators expect those decisions to be explainable and fair. This project builds a credit-risk model that meets both expectations.

Using a public loan dataset, the project engineers features such as debt-to-income ratio, credit utilisation and payment history, and handles the strong class imbalance between repaid and defaulted loans. Logistic Regression, Random Forest, XGBoost and LightGBM-style boosting are compared using ROC-AUC, precision-recall and a business cost function that weighs a missed default far more than a rejected good customer.

Each application receives a risk score, a decision band (approve, review, reject) and reasons from SHAP, phrased as adverse-action notes such as "high credit utilisation". A fairness report compares approval rates and error rates across groups and shows the effect of removing sensitive features. A Streamlit app lets a loan officer score new applications and run "what-if" changes.