There's a quiet assumption baked into most consumer lending strategy in Africa: that the real constraint on responsible credit expansion is data, specifically the credit bureau file that's still missing for most borrowers. Build better credit bureaus, the thinking goes, and underwriting gets solved. I spent the better part of a year testing that assumption against four million loan-month observations from the U.S. auto lending market, and the evidence points somewhere else entirely.
Here's the test I ran. Using loan-level disclosures on 239,374 securitized auto loans, I estimated a discrete-time hazard model of delinquency onset three ways: once with the borrower's credit score included, once with score collapsed into coarse bins, and once with score removed from the model entirely. If credit history really is doing the heavy lifting in underwriting, removing it should gut the model's explanatory power. It didn't. Pseudo-R² fell from 0.087 to 0.059, a real but modest decline, while loan-to-value ratio, vehicle age, contract term, and payment burden all held their statistical significance and their economic magnitude, seemingly unbothered by the absence of a score.
That's the part of this work I'd want a skeptical underwriter to sit with the longest, because it quietly undercuts an assumption most credit teams never actually test. What it tells me is that collateral structure and contract design carry information that's largely separate from borrower history, not a weaker substitute for it, but a genuinely independent signal.
The finding that should really reframe how blended finance structurers think about thin-file markets shows up in the stress-period results. During the COVID shock and the 2022-2023 rate-tightening episode, the marginal predictive power of credit score roughly halved, while leverage and amortization structure became more predictive, not less. The signal everyone assumes is essential gets weaker exactly when the cycle turns, and the signal embedded in the deal structure gets stronger, which is close to the opposite of how most credit models are built to behave.
I calibrated this to African lending conditions, thinner files, higher payment burden, weaker recovery infrastructure, and the design margins that moved projected twelve-month default weren't "get better credit data." They were loan-to-value discipline and amortization speed. A conservative LTV limit alone brought projected default from 83.8 percent down to 49.8 percent in the thin-file scenario. That's not a marginal improvement you have to wait years for bureau infrastructure to mature into. It's a lever available today, in the underwriting room, on every deal.
I'd push this further than the paper itself does, because the paper is careful and an opinion piece doesn't have to be. The emphasis on closing Africa's credit information gap has, in some corners of the impact investing conversation, become a way of deferring the harder work of underwriting discipline. Bureau infrastructure is worth building. It's just not the precondition for responsible lending that it's often treated as. Lenders who wait for it are leaving a structurally sound underwriting toolkit on the table in the meantime, one built on assets and contracts, not files.
If you're structuring private debt into underserved markets and your risk model leans heavily on a credit history that doesn't exist yet, the real question isn't how to get better history. It's whether your contract terms are doing the work your data can't.