From prediction to decision support: the evolution of AI-driven credit scoring in microfinance

Abstract

The article examines the evolution of artificial intelligence-driven credit scoring in microfinance, where limited availability of credit histories, information asymmetry, borrower heterogeneity, informality of economic activity, and data instability complicate credit risk assessment. The development of machine-learning, neural-network, ensemble, and hybrid models expands the possibilities for automated prediction, while simultaneously increasing requirements for data quality, interpretability, and the practical usefulness of results. The aim of the study is to systematize and configuratively synthesize research on artificial intelligence-driven credit scoring in microfinance, identify the structural and informational conditions of its development, trace the evolution of modeling approaches and their operationalization, and outline methodological gaps and interrelated unresolved trade-offs. The methodological basis is a systematic literature review with configurative thematic synthesis. The analysis was structured around three analytical categories and five dimensions: structural conditions and data; data preparation and feature construction; modeling; explainability and responsible use of artificial intelligence; and model implementation and outcomes. The results show a predominance of hybrid approaches, limited representation of explainable artificial intelligence, and substantially less attention to data preparation and feature-space construction. In some studies, credit scoring extends beyond default prediction to differentiating lending conditions, supporting credit officers’ decisions, and automated digital lending. Methodological gaps and unresolved trade-offs remain in predictive performance, transparency, data quality and stability, and practical model adaptation. The analysis demonstrates that artificial intelligence expands the capabilities of credit scoring but does not eliminate the constraints of the microfinance environment; further research should jointly consider predictive performance, data stability, interpretability, uncertainty, and the practical usefulness of models.