A novel AUC-based feature selection method: empirical insights from machine learning in the credit scoring problem

This study introduces a novel AUC-based feature selection method designed to enhance the performance of machine learning models in credit scoring, particularly in the context of imbalanced datasets. Unlike traditional approaches that rely on statistical significance or accuracy-based metrics, our method directly incorporates Area Under the Curve (AUC) as an objective function during the feature selection process. We compare the effectiveness of Logistic Regression (LR) and Random Forest (RF) models, showing that RF, when combined with AUC-driven feature selection, significantly outperforms LR in terms of discriminatory power. Through extensive cross-validation and empirical testing using real-world credit data, the results demonstrate that the proposed method improves model generalization and stability. Our findings contribute to the literature by emphasizing the importance of optimizing for AUC during feature selection, an area previously underexplored, and by validating the suitability of tree-based models like RF for credit scoring tasks involving complex, high-dimensional, and imbalanced data.