Abstract Machine learning (ML) models have become a central component of modern loan approval systems as financial institutions seek to improve prediction accuracy, automate decision-making, and reduce operational costs compared to traditional statistical approaches. By leveraging large and complex datasets, ML techniques can capture nonlinear relationships among borrower attributes that conventional models often fail to identify. However, the increased adoption of these models has raised significant ethical and regulatory concerns, particularly regarding algorithmic bias, fairness, transparency, and accountability in credit decisions. This paper provides a comparative analysis of three widely used ML models—logistic regression, random forests, and gradient boosting—with a focus on their predictive performance, susceptibility to bias, and implications for regulatory compliance. Using evidence from publicly available loan datasets and existing empirical studies, the analysis demonstrates that more complex ensemble-based models generally outperform simpler models in terms of accuracy and risk prediction. Nevertheless, these performance gains are frequently accompanied by reduced interpretability and a higher likelihood of disparate outcomes across protected demographic groups. The findings emphasize that accuracy alone is insufficient for responsible credit decision-making. To ensure ethical integrity and compliance with financial regulations, the integration of fairness-aware learning methods and explainable artificial intelligence (XAI) techniques is essential when deploying ML-based loan approval systems.
Manjiri Abhay Bilurkar (Sat,) studied this question.