Abstract
The majority of Indian small farmers belong to the category of thin-file customers who fail to meet the requirements for the payment history needed for credit bureau ratings and thus alternative data credit rating has been recommended as a solution. The practice itself is plagued by issues with measurements, since the accuracy is exaggerated due to resampling before the train-test splitting. This study addresses both gaps. We surveyed 265 small and marginal farmers across seven districts of North Coastal Andhra Pradesh and built a two-stage model on the 4C framework, comprising 22 ordinal items across four constructs. Oversampling was confined to the training partition of each fold. Under this protocol, CatBoost reached a Receiver Operating Characteristic – Area Under the Curve (ROC-AUC) of 0.783 for credit access and 0.819 for repayment, with Matthews correlation coefficients of 0.363 and 0.474 respectively. When the same model was trained on data oversampled before splitting, it reported 0.956 for access, an inflation of 0.173 AUC units arising from protocol alone. SHapley Additive exPlanations (SHAP) ranks were then tested across ten independent resamples: crop insurance enrolment ranked first for access in every resample, with a rank standard deviation of zero, while bank distance and farm location led the repayment stage. A four-tier score constructed from score quartiles showed a 17.2 percentage-point access gap between the top and bottom bands.
Keywords: Alternative Credit Scoring, CatBoost, Explainable AI, SHAP, Thin-file Borrowers.