Do Agricultural Subsidy Profiles Signal Bank-Authorized Lending? A Two-Stage Interpretable Machine Learning Analysis of Credit Access and Allocation

Information asymmetry constrains agricultural credit, but repayment and default labels are often unavailable for empirical analysis. Using administrative records for 2970 agricultural enterprises and cooperatives in Beijing linked to bank-authorized credit, this study examines whether multi-source subsidy profiles are associated with observed lending decisions. Two-stage gradient-boosted-tree models compare eight raw administrative features, an externally supplied composite score, and their combination; Shapley Additive exPlanations (SHAP) provides conditional model attributions. No positive authorized credit was observed for economic or share cooperatives. Among the remaining entity types, the combined model predicted accurate positive authorized credit, with an area under the receiver operating characteristic curve (ROC-AUC) of 0.97. Raw administrative features alone had little out-of-fold value for conditional authorized-credit amounts, whereas including the supplied score improved prediction. The framework supports transparent analysis of observed lending where repayment and application process data are unavailable.