Machine Learning-Based Loan Risk Prediction for Credit Scoring

The rapid expansion of consumer lending, digital banking, and other technology-enabled financial services increases the need for timely and consistent credit-risk assessment. Traditional loan underwriting commonly relies on predefined rules and manual review of income, employment, collateral, guarantors, and credit-bureau information. Although such approaches remain important, they can be slow and may have limited ability to identify complex, non-linear relationships among borrower and loan characteristics. This study develops and evaluates a machine-learning-based loan risk prediction framework for binary classification of loans into lower-risk and higher-risk outcomes. The project uses an approximately one-million-record historical loan dataset obtained from Kaggle, with 57 attributes in the supplied schema. The methodology includes data cleaning, missing-value handling, categorical encoding, numerical scaling, outlier treatment, feature engineering, stratified train-test splitting, model training, hyperparameter tuning, and iterative refinement. Five supervised learning approaches were compared: Logistic Regression, Decision Tree, Random Forest, Gradient Boosting, and XGBoost. The reported test ROC-AUC scores were 0.7024, 0.5515, 0.7165, 0.7193, and 0.7266, respectively, making XGBoost the best untuned model. After hyperparameter tuning, XGBoost achieved a test ROC-AUC of 0.7273. The results indicate that XGBoost provides the strongest overall discrimination among the tested models, while threshold selection remains a critical business decision because it changes the balance between approving good customers and preventing default exposure. The proposed framework also includes model export using joblib and deployment through a REST API and web application. The study demonstrates the potential of machine learning as a decision-support tool for credit risk management, while emphasizing that local Cambodian loan data, cost-sensitive thresholding, monitoring, and explainability are required before production deployment.

14 Aug 2026 1 335