Affirm’s expanding customer activity raises the stakes for underwriting: its model must assess not only a borrower’s credit record, but also how that record is changing. The company has put a transformer-based model into use across its U.S. checkout, looking for patterns in the timing and sequence of credit events that conventional approaches may overlook. That design reflects a broader challenge in lending: a snapshot can show where a borrower stands, while the path that led there may tell a different story.
The model has already allowed Affirm to approve some applicants its prior system would have declined. Those added approvals produced 3.4% more completed purchases than a control group at comparable risk, according to the company. Affirm says the result also exceeded a similar expansion using its previous machine-learning models. The finding is specific to the test described by the company; it doesn’t establish that every additional approval will perform as well, or that the model will produce the same result as its use expands.
More activity makes underwriting more consequential
Affirm has extended beyond point-of-sale installment loans with the Affirm Card, its app and marketplace, and the Affirm Money Account savings product. Customers are using the service more often and for smaller purchases. In fiscal 2026, active consumers totaled 27.8 million, up 21%, while transactions per active consumer increased from 5.8 to 7.0. Those figures point to more frequent contact between customers and the company’s products, which makes the quality of each credit decision more consequential across the relationship.
That growth puts more weight on the quality of the underwriting behind the product range. A credit score condenses a complex financial history into one measure, but it can hide whether a borrower is recovering from past problems or slipping further behind. Affirm’s new model aims to capture that direction by examining when credit events occurred and how they unfolded. Similar systems typically seek to extract useful sequences from information that would be harder to represent as a single static score.
“The time dimension of multiple purchases, multiple credit events in a customer’s life wasn’t being represented with particularly high fidelity,” Affirm President Libor Michalek said in a recent podcast. His description points to the gap the company is trying to address: credit events don’t arrive as isolated data points, and their order may matter. A missed payment followed by a period of improvement can carry a different signal from a pattern that is deteriorating. The model’s task is to distinguish those patterns well enough to improve decisions, not simply to process more history.
What the test does—and doesn’t—show
The early purchase results suggest that adding this context can widen approvals while keeping risk comparable. The test gives Affirm a practical measure to track as customer activity grows. That’s a useful result, but it’s also a narrow one: completed purchases measure whether newly approved applicants went on to transact, while the company’s risk comparison addresses a separate part of the underwriting question.
This is more significant than it looks because approval growth can be valuable only if the added activity remains consistent with the risk standard used in the comparison. A model that approves more people but misreads changing credit conditions could create costs that aren’t visible in an initial purchase result. The company’s reported outcome therefore merits attention, but the available details don’t settle how the model performs over longer periods or across different customer groups.
Implications as customer use grows
Affirm’s wider product reach gives it more opportunities to observe customer activity, while also raising the importance of how that information informs credit decisions. The company is presenting the transformer model as a way to read event timing and sequence with greater precision. And yet, more data alone doesn’t guarantee better underwriting; the value depends on whether the model’s signals remain useful as customer behavior and the mix of products change.
If you’re working in this space, the result to watch is not only whether approvals rise, but whether comparable-risk performance holds as the system encounters a broader range of applicants and transactions. Affirm has reported an encouraging test outcome against its stated comparison. The next question is whether that advantage persists in routine use. That’s the harder measure.
…
The post Affirm wants its data to tell a more complete customer story appeared first on Tearsheet.
