A loan-level empirical study of what actually predicts repayment when formal credit histories are thin or absent, using U.S. securitized auto loans to identify underwriting signals that extend responsible credit in Emerging Markets, then calibrating those findings directly to African consumer-lending conditions.
This paper examines how consumer credit can be efficiently extended in environments where formal credit histories are limited or incomplete. Using loan-level data on U.S. securitized auto loans disclosed under Regulation AB II and accessed via Wharton Research Data Services, it studies repayment performance, delinquency transitions, and default outcomes in a market where lenders combine collateral values, borrower characteristics, and contract design rather than relying solely on credit scores. The analytic sample comprises 239,374 unique loans drawn from 154 trusts issued by 18 sponsors, yielding just over 4 million loan-month observations between November 2016 and December 2024. Estimating discrete-time hazard, logit, and linear-probability models with trust-clustered standard errors, the paper finds that loan-to-value ratio, vehicle age, and contract maturity retain economically and statistically significant predictive power after conditioning on credit scores, and that removing score information from the model reduces explanatory power only modestly. The results are then calibrated to African consumer-lending conditions, showing that conservative loan-to-value limits and faster amortization can materially reduce projected default. The paper positions collateral-backed consumer credit as a distinct informational architecture for extending responsible credit where histories are thin.
The estimated hazard applied to alternative covariate profiles. These are calibrations, not forecasts, the policy-relevant signal is the gradient across design choices, not the absolute level.
Monthly asset-level performance data on securitized auto loans, accessed through Wharton Research Data Services, deduplicated to one observation per loan-month across amended filings.
A complementary log-log hazard for delinquency onset, a logit for the delinquent-to-default transition, and a two-way fixed-effects linear-probability model to benchmark explained variance, all with trust-clustered standard errors.
Performance is estimated with the continuous score, with score collapsed into three coarse bins, and with score excluded entirely, isolating exactly how much predictive power sits outside formal credit histories.
Results hold after treating prepayment as a competing exit rather than censoring it away, and the stress-period interaction isolates how the same underwriting signals behave under the 2020 and 2022–2023 macro shocks.
This paper is one piece of a question I keep coming back to: what does responsible underwriting actually look like when a borrower's history is incomplete, whether that's a thin file in Nairobi or an unscored file in Ohio. The finding at the center of this work, that collateral structure and contract design carry real predictive signal independent of a credit score, doesn't respect the line most financial-inclusion research draws between emerging and developed markets. I'm extending it in two directions next: recalibrating the hazard model directly against the current subprime auto cycle to see whether the stress-period result holds up against live data rather than only the 2020 to 2023 window I originally tested, and building a comparable dataset on US thin-file and unscored borrowers to test the same framework where, so far, I've only calibrated it. The goal isn't a longer paper. It's one underwriting framework that holds up whether the borrower is in Nairobi, Nashville, or anywhere else formal credit history hasn't caught up with actual creditworthiness.