Adaptive feature selection for credit scoring models

Abstract

Accurate credit scoring and delinquency prediction are critical for financial institutions, particularly when evaluating both banked and unbanked clients. Traditional feature selection approaches such as wrapper, filter, and embedded methods often require repeated retraining and may not efficiently explore optimal feature subsets in multi-class credit risk settings. This study proposes a reinforcement learning (RL)–based adaptive feature selection framework using tabular Q-learning to dynamically identify informative feature subsets for credit bucket classification and delinquency prediction. The framework was evaluated on two datasets, including proprietary credit datasets (24,533 records with 18–19 features). Using stratified 10-fold cross-validation, the proposed approach achieved 62.84% accuracy (F1=0.63) for 8-class credit bucket prediction, outperforming traditional wrapper-based methods by up to 4.2% while reducing computational cost compared to exhaustive subset evaluation. For delinquency prediction, the model achieved approximately 70% F1-score, demonstrating improved minority-class sensitivity. Compared to filter and embedded methods, the RL-based framework produced more compact feature subsets while maintaining competitive computational efficiency. These findings demonstrate that adaptive RL driven feature selection provides a practical and scalable mechanism for enhancing predictive performance in credit risk modelling, supporting automated and consistent decision making in financial institutions.