A Comparative Analysis of Machine Learning Models for Predicting Loan Defaults under Imbalanced Data Conditions

Accurate loan default prediction is crucial for credit risk management. This study used Kaggle’s Loan Default Prediction dataset, applying preprocessing (cleaning, encoding, scaling) and testing six models: Logistic Regression, Decision Trees, Random Forest, eXtreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), and a neural network. Performance was evaluated using precision, recall, F1-score, and Receiver Operating Characteristic (AUC-ROC), with emphasis on handling class imbalance. Gradient boosting (especially LightGBM, AUC ~0.76) outperformed linear and tree-based models. Adjusting XGBoost’s decision threshold improved the minority class F1-score from 0.17 to 0.36 without losing Area Under the Curve (AUC). Ensemble methods like Voting Classifiers balanced recall and precision effectively. Key takeaways include model selection, threshold tuning, and ensemble strategies for imbalanced data. Future work could explore richer features, imbalance-aware loss functions, and explainability tools for transparency. This study provides a reproducible benchmark for default prediction, laying the groundwork for more robust, interpretable, and fair credit scoring systems in real-world financial applications.