Digital Footprints as Institutional Hard Constraints: A Multi-Source Data Fusion System for the Agricultural Credit Risk Early Warning

Agricultural credit rationing remains a persistent systemic friction driven by information opacity and limited collateral. This study develops a credit risk early-warning system by fusing multi-source institutional digital footprints (tax compliance signals, judicial enforcement records, and credit history indicators) for 1021 agricultural enterprises in China. Methodologically, we propose a Default Event Isolation protocol to enforce strict ex ante validity by discarding observations at and after the event month, and implement a two-step feature optimization pipeline that reduces 138 predictors to a parsimonious set of 50 features. Empirically, the optimized LightGBM (version 4.6.0) model achieves an AUC = 0.9345 (95% bootstrap CI: 0.8745–0.9563) and PR-AUC = 0.4421, representing a 47× lift over the random baseline under extreme class imbalance (0.94% event rate), and captures 87.4% of early-warning events by monitoring only the top 10% highest-risk firms. The interpretability analysis consistently highlights judicial boundary constraints and tax stability signals as dominant predictors, forming a “judicial baseline + tax stability” dual-core structure. A strict credit-only robustness check using bank-recorded NPL labels maintains strong predictive performance (AUC = 0.9089, 95% bootstrap CI: 0.8255–0.9591), mitigating concerns that the model’s signal is driven by label overlap. These findings suggest that integrating institutional records into automated screening pipelines can enable the earlier and more targeted identification of distressed borrowers in rural lending, even when traditional financial statements are unavailable.