Machine learning for retail credit risk scoring: a systematic literature review with insights for South African banks

Abstract

Credit scorecards remain central to bank lending, yet modern credit datasets increasingly demand models that capture non-linear, high-dimensional patterns. Traditional models such as logistic regression are increasingly constrained by their linear assumptions, particularly in emerging markets like South Africa where many borrowers have thin or fragmented credit histories. This study systematically reviews 32 peer-reviewed papers to examine how machine learning (ML) can be applied to retail credit scoring while maintaining the transparency requirements mandated by regulators and required by auditors. The study also conducts a performance comparison between ML approaches and barriers to adoption. Findings show that tree-based ensemble methods (Random Forest, XGBoost, LightGBM and CatBoost) consistently outperform traditional approaches in accuracy and stability while neural networks and support vector machines also perform well but raise transparency challenges. Explainable artificial intelligence (AI) techniques, especially SHapley Additive exPlanations (SHAP), emerge as practical tools to bridge predictive power with auditability. The review concludes that South African banks can adopt a staged, hybrid approach: using ML in data preparation and segmentation while retaining interpretable decision layers, thereby enhancing predictive accuracy and financial inclusion without undermining regulatory transparency.