Enhancing Agricultural Financing: A Data-Driven Approach to Credit Risk Prediction for Farmers in Sub-Saharan Africa

Access to financing poses a significant challenge for farmers, particularly in regions such as sub-Saharan Africa, where high perceived risks limit credit availability. This work presents a data-driven approach utilizing machine learning to enhance credit risk prediction and improve lending decisions for smallholder farmers. We constructed a dataset incorporating key demographic, financial, and agricultural characteristics of borrowers. Various machine learning models, including Logistic Regression, Random Forest, Support Vector Machines (SVM), Decision Tree, Naive Bayes, and K-Nearest Neighbors, were evaluated using three feature selection methods: All Features, Lasso selection, and Forward Selection. The findings indicate that machine learning models provide a statistically significant improvement over traditional microfinance assessment techniques, which yield only 45% accuracy. In contrast, the top-performing configuration; Decision Tree with Forward-selected features achieved a study-leading accuracy of 95%, precision of 92%, and F1-score of 94%. While Random Forest achieved high accuracy (93%) with all features, it proved highly sensitive to feature reduction. SVM and Logistic Regression emerged as the most stable models across feature sets, with the latter achieving 100% recall. The use of Lasso-selected features contributed to more stable predictions and minimized performance variability across most models. These results highlight the capacity of AI-driven models to provide high-precision risk assessments, ultimately facilitating greater financial access for farmers and supporting agricultural development.