Affirm rolls out a transformer-based underwriting model, says it lifted checkout approvals 3.4%
The new model reads the sequence and timing of a consumer’s credit history rather than static summary scores, letting Affirm approve thin-file and no-FICO applicants earlier systems would have declined.
Affirm has deployed a new transformer-based machine learning model for real-time underwriting at U.S. checkouts, the buy-now-pay-later lender said, in what it described as a structural shift in how it reads a consumer's credit history rather than an incremental tuning of its existing risk models.
Transformer architectures are the same general family of model behind large language models, built to weigh the order and relationship between items in a sequence rather than just their aggregate. Applied to underwriting, Affirm said that means the model considers the sequence and timing of events in an applicant's credit history — how and when past credit behavior occurred — instead of relying solely on static summary measures like a single credit score snapshot.
In its initial deployment, Affirm said the model increased completed purchases by 3.4% compared with a control group, primarily by approving eligible applicants that its previous systems would have declined, including consumers with thin credit files or no FICO score at all. The company said it paired the transformer model with a proprietary explainability layer built alongside it, intended to preserve real-time decision speed and the ability to state a reason for a given approval or denial, which is a standard requirement under U.S. fair-lending and adverse-action disclosure rules.
Affirm did not disclose the size of the initial deployment population, the underlying approval or default rates in absolute terms, or an independent audit of the model's fair-lending performance; the 3.4% figure and the description of the model's mechanics come from the company's own announcement. Automated underwriting models that use non-traditional or behavioral signals have drawn continued scrutiny from U.S. regulators over whether they can inadvertently produce disparate outcomes across protected classes even without directly considering prohibited factors, a question that applies to any lender adopting more complex, less transparent scoring architecture regardless of the explainability tooling built around it.