Consensus-driven feature selection for transparent and robust loan default prediction.

Area of Science:

• Financial Risk Management
• Machine Learning Applications
• Data Science

Background:

• Accurate loan default prediction is crucial for financial stability but challenged by complex borrower data.
• Existing feature selection methods often yield redundant, unstable, and uninterpretable models.

Purpose of the Study:

• To develop a novel feature selection framework, Hybrid Rank-Aggregated Feature Selection (HRA-FS), to address limitations in loan default prediction.
• To enhance model accuracy, stability, and interpretability for equitable credit assessment.

Main Methods:

• Proposed HRA-FS framework integrating ReliefF, Recursive Feature Elimination, and ElasticNet via Borda count aggregation.
• Incorporated strategic feature categorization to mitigate domain dominance and ensure balanced risk driver representation.
• Evaluated HRA-FS on imbalanced real-world datasets using XGBoost classifier.

Main Results:

• HRA-FS consistently outperformed single feature selection methods across datasets.
• Achieved a high ROC-AUC of 0.965 for small firm loan default prediction.
• Identified concise, non-redundant, and interpretable feature sets, including house value and inventory turnover rate.

Conclusions:

• The consensus-driven, category-aware HRA-FS framework effectively resolves the accuracy, stability, and interpretability trilemma in feature selection.
• HRA-FS provides lenders with robust tools for equitable credit assessment, promoting financial inclusion.
• This approach enhances the development of more interpretable and reliable predictive models in finance.