Predictive Analytics for Credit Risk Assessment: A Machine Learning Approach to Loan Default Prediction (International Journal of Science Strategic Management and Technology)

Summary

• The research analyzes a secondary dataset of 45,000 loan records from Kaggle to evaluate machine learning approaches for credit risk assessment.p.3
• The training set comprised 80% of data (30,039 observations) while the testing set contained 20% (7,510 observations).p.4
• The dataset exhibits class imbalance, with 29,500 non-default cases compared to approximately 8,000 default cases.p.4
• Random Forest achieved the best overall performance with 92.69% accuracy, 88.19% precision, 81.51% F1-score, and 97.04% ROC-AUC.p.5
• Decision Tree yielded the highest recall at 77.46%, compared to Random Forest's 75.77% recall.p.1
• Logistic Regression recorded an accuracy of 89.24% and a ROC-AUC of 94.95%.p.5

Key insights

1. Interest rate is identified as the single most influential variable for predicting loan default risk.p.6
2. Financial and credit characteristics show a significantly stronger relationship with default risk than borrower demographic traits such as age or employment experience.p.6
3. Prior default history is a critical predictor, as borrowers with past defaults exhibit a substantially higher likelihood of recurring default.p.5
4. Loan purpose impacts default rates, with medical and debt-consolidation loans displaying higher default risk than venture loans.p.5
5. While Random Forest provides superior overall accuracy, Decision Tree is better suited for prioritizing the detection of actual defaulters due to its higher recall.p.5
6. The unhandled class imbalance (78% non-defaulters to 22% defaulters) contributed to relatively low recall scores across all models.p.3