Buy now, pay later platforms grew quickly by offering a credit product that required minimal friction at checkout. No hard credit inquiry, fast approval, small payment amounts spread over weeks. The demographic that found this most appealing was, unsurprisingly, exactly the demographic that traditional credit models can't see: young adults, recent immigrants, and people with thin or no bureau files.
That alignment between BNPL's product design and the credit-invisible population created a structural problem that has become clearer as the segment has matured. The very customers BNPL was best positioned to serve are the ones the default models were least equipped to evaluate. The result has been elevated default rates on a segment that, assessed properly, would not warrant them.
The BNPL underwriting design problem
Early BNPL platforms used one of two underwriting approaches: soft bureau pull (which gets you a thin-file non-answer for 25–30% of applicants) or rules-based transaction history for repeat customers (which has no data at all for new applicants). Neither approach is designed for thin-file populations. Both are proxies that fail silently — they don't tell you an applicant is high risk, they just don't tell you they're low risk.
The decision point in this design is the fallback. When a bureau pull comes back unscorable, what does the model do? Many BNPL platforms set a default conservative limit — approve for a low dollar amount ($100–$250) regardless of actual creditworthiness, with no consideration of the applicant's actual financial profile. This "approve thin-file at low limit" approach sounds prudent but creates a self-fulfilling problem: the applicant gets a product designed for someone with demonstrated bureau history, uses it, and the platform learns nothing useful about their creditworthiness because the limit was too small to stress any behavioral signal.
Why thin-file defaults look worse than they are
The reported default rates on thin-file BNPL customers often reflect selection effects rather than inherent credit risk. Consider what happens at a platform that handles thin-file applicants conservatively:
Creditworthy thin-file applicants — those with good financial behavior but no bureau history — are approved for the same low limit as high-risk thin-file applicants. At low dollar amounts, the behavioral differentiation between good and bad risk is compressed. A person who would reliably repay $800 installment obligations behaves differently on a $150 limit than on a $450 limit. The signal-to-noise ratio in the repayment behavior at low limits is poor.
Meanwhile, the platform's "thin-file customers" bucket is treated as a homogeneous risk category in reporting and model evaluation. Average defaults in that bucket are high. The conclusion drawn: thin-file customers are high risk. The actual explanation: the underwriting process lumped creditworthy and uncreditworthy thin-file customers together with no discriminatory signal, and then charged the creditworthy ones with the default rate of the uncreditworthy ones.
This is not a hypothetical error. It's a predictable consequence of applying undifferentiated approval logic to a heterogeneous population.
The credit stacking problem compounding BNPL defaults
One of the features that makes BNPL structurally different from credit cards is bureau invisibility — historically, most BNPL tradelines were not reported to the major credit bureaus. This creates a credit stacking problem: a consumer can hold active BNPL obligations across four or five platforms simultaneously, and no single platform can see the others when evaluating a new application.
From a credit risk perspective, an applicant with $800 in active BNPL obligations split across three platforms, plus a pending auto loan payment, is in a materially different financial position than an applicant with no current obligations. But if none of those obligations appear on the bureau pull, the credit model sees an identical input profile for both applicants. The obligation-laden applicant defaults at a higher rate. The platform attributes this to "thin-file risk." The actual cause is obligation stacking invisibility combined with inadequate cash-flow evaluation.
Bureau reporting for BNPL products has improved — the major credit bureaus have introduced short-term installment plan tradeline categories and several platforms have begun reporting. But coverage remains incomplete, and for thin-file applicants, the circularity persists: the BNPL tradeline that appears on the bureau is itself a thin tradeline that doesn't generate a meaningful score on someone with no prior bureau history.
What a cash-flow model sees differently
Consider a specific scenario: a 24-year-old applicant with a checking account at a regional bank, no credit card history, and three prior BNPL purchases all paid on time. Her bureau file is unscorable. The BNPL bureau pull returns no score. The standard model approves her at the default low limit.
Her bank transaction history tells a different story. Over 18 months, her monthly income from a salaried employer has been $3,100, deposited on the same two dates each month without exception. She pays $950 in rent via ACH, always 2–3 days before the first of the month. Her minimum balance in the 5 days before each deposit averages $280 — she doesn't cut it close. Her overdraft count in 18 months is zero. Her net cash flow trend over the lookback window is slightly positive — she's slowly building savings.
This is not the risk profile of a thin-file default candidate. This is a creditworthy borrower in the early stage of building a credit history. A cash-flow model on this transaction record would assign a materially better score than the unscorable bureau pull implies. The appropriate BNPL limit is not $150. It's $400–$600, priced for her actual risk level.
Contrast that with a different thin-file applicant at the same platform: similar age, similar income level on average, but with high income variability (gig platform), frequent overdrafts, a declining balance trend, and evidence of multiple active BNPL-sized payment outflows. Same bureau result: unscorable. Very different cash-flow risk profile. This is the applicant who generates the elevated default rate that gets attributed to the entire thin-file segment.
Model design implications for BNPL risk teams
The diagnosis above has concrete model design implications. BNPL platforms serving thin-file populations need to:
- Stop treating unscorable bureau returns as a single risk bucket. The unscorable population is as heterogeneous as the scoreable population. Averaging over it produces meaningless risk estimates.
- Build cash-flow feature pipelines specifically for the thin-file fallback path. When the bureau pull returns unscorable, trigger a bank account connection request and route to a cash-flow model, not a conservative rules-based fallback.
- Calibrate limits to the alternative data risk score, not to the bureau score or its absence. A thin-file applicant with a strong cash-flow risk profile should receive a limit appropriate to that profile, not the bureau-equivalent low limit.
- Track default rates by thin-file sub-segment, not by thin-file aggregate. Model validation on the thin-file population requires separating the cash-flow-scored sub-segment from the unscorable-and-approved-anyway sub-segment. Aggregate default rates are misleading.
The Regulation B consideration for BNPL thin-file decisions
ECOA and Regulation B apply to BNPL products that constitute credit. A policy of approving thin-file applicants at a systematically lower limit than bureau-scoreable applicants with equivalent financial profiles may constitute disparate treatment if the thin-file population is disproportionately composed of protected class members — which, given the demographic composition of credit-invisible Americans, it often is.
We're not saying any lender currently applying this policy is in violation — that requires a full adverse impact analysis that depends on the specific population, limit distribution, and feature set. What we are saying is that the regulatory exposure from undifferentiated conservative thin-file treatment is not zero, and a properly differentiated cash-flow underwriting approach reduces that exposure while simultaneously improving portfolio performance.
The platforms that get this right will approve more of the creditworthy thin-file customers at appropriate limits, produce better portfolio economics, and build a defensible compliance posture simultaneously. The ones that treat thin-file as an undifferentiated high-risk bucket will continue paying the cost of that conflation.
