Examining the Research Gap in Explainable AI for Credit Scoring in Nigerian Microfinance Institutions

Machine learning credit scoring models are increasingly used by Nigerian microfinance institutions to automate loan eligibility decisions, but their black-box nature creates transparency and fairness concerns that explainable AI techniques could address, a research gap that has received no Nigeria-specific empirical attention. This study examined this research gap through systematic review and original empirical evaluation. A systematic review of explainable AI credit scoring publications from 2020 to 2024 identified 48 papers, of which only one included African and no studies covered Nigerian microfinance contexts. Original experiments trained Random Forest, XGBoost, and neural network credit scoring models on a de-identified dataset of 12,400 loan applications from a North Central Nigerian microfinance institution. SHAP, LIME, and Integrated Gradients were applied as post-hoc explanation methods for each model. Explanation fidelity was measured by the accuracy of local surrogate models generated by LIME and SHAP. SHAP explanations for XGBoost showed the highest fidelity (mean explanation accuracy 94.7 percent) and the greatest consistency across repeated explanations. LIME explanations were less stable, varying by 12.4 percent across ten repeated local samples for the same instance. A fairness audit using SHAP feature importance revealed that mobile phone ownership and market association membership were the two features with the highest credit score impact, suggesting potential proxies for informal sector occupation type rather than genuine creditworthiness discriminators. The study recommends MFIs adopt XGBoost with SHAP explanations as a minimum transparency standard for algorithmic credit decisions.

Keywords: explainable AI, credit scoring, microfinance, SHAP, Nigeria