Drift-Resilient Early-Warning Models for SME Loan Distress: A Cross-Institution Validation Framework

Abstract

Early-warning models for small and medium-sized enterprise (SME) loan distress are usually evaluated within a single lending institution, where apparently strong discrimination may reflect local portfolio composition, underwriting policy, and data-definition choices rather than risk relationships that transfer elsewhere. This article develops a multi-institution validation framework for drift-resilient SME early-warning models, building on the multi-institution architecture that Rashid and Sattiraju propose for SME loan distress and transaction-fraud pilots across U.S. community financial institutions, and applies it to public data. The framework separates two questions that single-site evaluation conflates, temporal stability within an institution and transfer to an institution the model has never seen, and specifies harmonized outcomes, leakage-safe partitions within each site, staged cross-site transfer tests, unsupervised drift detection with D3 and OCDD, and site-level reporting. It is implemented on 375,261 SBA-guaranteed small-business loans, with eight industry-sector portfolios standing in for institutions because the public file identifies neither lender nor approval date. Transferred models ranked borrowers as well as site-specific ones (leave-one-site-out AUC 0.68 to 0.79), but calibration did not transfer: predicted charge-off rates were off by up to 8.4 percentage points, and between-site heterogeneity was high (I² = 0.96). Local recalibration on about 1,000 labelled loans cut the calibration error by roughly 72 percent, whereas 100 labels made it worse. At their default thresholds, D3 and OCDD rarely flagged the cross-site shifts, and no detector statistic tracked the performance cost of transfer. The results support site-level reporting and labelled local recalibration over label-free drift alarms.