A Machine Learning–Based Framework for Predicting Loan Default Risk

Abstract - The growth of online lending has made efficient credit risk management more important than ever. Old-fashioned, manual review processes are often slow and inconsistent, especially when dealing with thousands of loan applications. To solve this, we built an end-to-end machine learning system that predicts the likelihood of a borrower defaulting. Our approach covers every step—from cleaning the data and creating insightful features to testing different models and deploying the best one. We compared four common algorithms (Logistic Regression, Decision Tree, Random Forest, and XGBoost) on a public loan dataset. The results clearly showed that advanced "ensemble" models perform best. Specifically, XGBoost achieved 87% accuracy and an excellent ROC-AUC score of 0.91, while also making fewer costly mistakes (like incorrectly labeling a risky borrower as safe). To make this research useful in the real world, we packaged the winning model into an interactive web application using Streamlit. This work provides a clear, complete blueprint that banks and lenders can adapt to make faster, smarter, and more reliable lending decisions. Key Words: Loan Default Prediction, Machine Learning, Explainable Artificial Intelligence, Credit Risk Assessment, XGBoost, SHAP