Affirm Underwriting Reads Credit as a Sequence

Affirm underwriting now uses a transformer-based model at U.S. checkout to read the order and timing of events in a consumer’s credit history. Affirm says its first controlled deployment produced 3.4% more completed purchases while keeping risk comparable, but that performance claim remains company-reported rather than an independent audit.

Key takeaways

The model is live in the United States, not a laboratory preview.
It evaluates sequences in credit histories alongside familiar bureau measures.
Affirm says thin-file and no-FICO applicants were among those newly approved.
The practical test is whether repayment quality holds after the model scales.

Affirm underwriting changes the unit of analysis

Traditional credit models often reduce a borrower’s history to summary variables: balances, utilisation, account counts and payment history. Affirm says the new model keeps those signals but also learns how events unfold over time. A recent balance change, a repayment and a new account can carry different meaning depending on their order and distance from one another.

That is the useful mechanism behind the announcement. Transformer architectures are designed to identify relationships inside sequences. Applied to credit, they can look for temporal patterns without engineers hand-coding a separate feature for every possible path. The model is therefore not simply a chatbot attached to a loan form; it sits inside a real-time underwriting decision.

What the 3.4% result does and does not prove

Affirm’s announcement says an initial control-group comparison generated 3.4% more completed purchases. The additional approvals included people with limited credit histories and no FICO score, and the company says those loans performed better than a comparable expansion under its previous machine-learning models.

The number is commercially meaningful because a lender can increase merchant conversion only if credit losses do not erase the gain. Yet the disclosure does not provide the sample size, observation period, delinquency curve or dollar value of the added loans. Readers should treat 3.4% as an early operating result reported by Affirm, not a general claim that transformer models make lending safer.

Disclosed item	What is known
Deployment	Live at U.S. checkout
Reported lift	3.4% more completed purchases
Newly reached group	Some thin-file and no-FICO applicants
Not disclosed	Sample size, loss rate and observation window

The business consequence is a tighter feedback loop

Affirm underwrites each purchase rather than assigning one permanent decision to a customer. That means new repayment outcomes can feed later model generations, while merchants see a decision at the moment of sale. The architecture could help the lender distinguish temporary noise from a repeated pattern, but it also concentrates responsibility in monitoring, explanations and adverse-action controls.

Affirm says it built a proprietary explanation layer that keeps the system fast enough for checkout and produces explanations comparable with traditional models. That is important because credit decisions cannot become an opaque consequence of model complexity. The lender’s August 2026 annual filing separately says its risk models are designed to comply with partner-bank policies and its own underwriting standards.

For the wider fintech market, the shift resembles the infrastructure logic behind Ryft’s payments expansion and Tare’s private-credit infrastructure: the competitive advantage sits in decision plumbing, not just a consumer-facing screen. The durable advantage will depend on data quality and loss performance, not the model label.

Affirm’s transformer underwriting model is a live credit-decision system that reads the timing and order of credit events; its reported 3.4% checkout lift is promising, but the long-term credit outcome has not yet been independently demonstrated.