The transformer-based model is now live, according to a Thursday (Sept. 17) announcement shared exclusively with PYMNTS. It builds on 14 years of Affirm underwriting every purchase individually, in real time, with machine learning models developed in-house. That approach is the basis of what Affirm promises its customers. Each purchase gets its own decision, based on what a person can responsibly repay that day, with no late or hidden fees.
The early results cut against the usual trade-off in lending. In its initial deployment, Affirm used the model to approve additional eligible applications its existing system would have turned down. Measured against a control group, that produced 3.4% more completed purchases, and those loans performed better than a comparable expansion under Affirm’s previous machine learning models.
“Seeing a credit history more clearly means we can responsibly say yes to more people,” Affirm President Libor Michalek said in the announcement.
The Model Finds More Signal in the Same Data
A credit score compresses a consumer’s record into a single number. That works when there’s a long record to compress. Applicants with short histories give a score-based model less to work with, and the safe default has long been to decline them.
Affirm’s models have long used credit bureau measures such as balances, utilization, account counts and payment history. Those measures still matter, but the transformer reads the same data differently. It identifies patterns within and across a consumer’s credit accounts, including how they change over time, without a separate measure having to be designed for each pattern in advance.
“We’ve steadily accelerated the amount of data we use to train each generation of our underwriting models,” Michalek said. “What’s exciting about the transformer model architecture is that we can now find new information within the data we already have.”
We’d love to be your preferred source for news.
Please add us to your preferred sources list so our news, data and interviews show up in your feed. Thanks!
More Approvals Only Work if the ‘No’ Gets Better
Michalek was clear that the point isn’t to loosen standards.
“Underwriting is the heart of what we do,” he said. “The goal isn’t to approve every transaction, it’s to make the right decision for each one. We don’t benefit from extending credit that can’t be repaid, which means saying yes to more people only works when we get even better at saying no.”
That per-purchase discipline is what Affirm CEO Max Levchin has described as underwriting “every transaction in addition to every single person, every single time,” as he put it in a LinkedIn post earlier this year. The transformer becomes one more input into that decision.
Building a better model was only part of the work. Affirm also built a proprietary algorithm that produces the same level of explainability as its traditional machine learning models while keeping the transformer fast enough to run at checkout. Validation and ongoing monitoring check that those explanations stay accurate and reliable. Affirm has published details on how it built and tested the model in a technical blog post.
Consumers Are Using Affirm for Smaller, More Frequent Purchases
The model arrives as Affirm’s customers lean on it more often. As PYMNTS reported after Affirm’s fiscal fourth-quarter results, the company processed 53 million transactions in the quarter, up 41% year over year, while gross merchandise volume rose 36% to $14.1 billion. Average order value declined 4%.
Active consumers grew 21% to 27.8 million, and transactions per active consumer rose 20% to 7.0 over the trailing 12 months. More transactions at a lower average ticket points to consumers using Affirm as a regular way to pay for everyday purchases, on terms they know upfront, rather than only as financing for the occasional big one. A model that sees more of those consumers clearly, and says yes to the ones who can repay, widens who gets access to that option.
Add as Preferred Source
