SAN FRANCISCO — Affirm has rolled out a new transformer-based machine learning model for real-time credit underwriting that the company said is allowing it to approve more consumers, including some with limited credit histories or no FICO scores, without taking on greater levels of risk.
Affirm said the model, which is now being used at checkout in the U.S., analyzes the order and timing of events in a consumer’s credit history to identify patterns that may not be captured by traditional credit measures.
In its initial deployment, Affirm said the model allowed it to approve additional applications that its existing underwriting system would have declined. Compared with a control group, the new approach resulted in 3.4% more completed purchases, according to the company.
Better Performance
Affirm said the additional loans also performed better than a comparable expansion in approvals using its previous generation of machine learning models.
The company, which provides buy-now, pay-later and other payment products, said it has spent 14 years underwriting individual purchases in real time using machine learning models developed internally.
“We’ve steadily accelerated the amount of data we use to train each generation of our underwriting models,” Affirm President Libor Michalek said in a statement. “What’s exciting about the transformer model architecture is that we can now find new information within the data we already have. Seeing a credit history more clearly means we can responsibly say yes to more people.”
Looking Beyond Traditional Credit Measures
Affirm said its underwriting models have traditionally incorporated credit bureau information such as outstanding balances, credit utilization, numbers of accounts and payment histories.
The new transformer-based model is designed to go further by identifying patterns within and across credit accounts and analyzing how those patterns change over time, according to Affirm.
Rather than requiring the company to develop a separate measure for each potential pattern, the model can learn relationships from the sequence and timing of credit events, Affirm said.
The architecture is similar in concept to transformer technology used in other areas of artificial intelligence, where models analyze relationships among elements in a sequence. In Affirm’s case, the sequence involves events in a consumer’s credit history rather than words or other content.
Affirm said the approach allows it to extract additional underwriting information from data it already possesses.
The company has significantly increased the amount of information used to train successive generations of its underwriting models. Affirm investor materials show the volume of data points used for model training increased from 3 billion in November 2022 to 186 billion as of May 2026.
Designed for Real-Time Decisions
Affirm said one challenge in adopting the transformer architecture was ensuring that the more sophisticated model could still make decisions quickly enough to be used during a consumer transaction.
The company said it developed a proprietary algorithm intended to maintain the same level of explainability available with its traditional machine learning models while allowing decisions to be made in real time.
Affirm said it also uses validation and ongoing monitoring to assess whether the explanations generated by the model remain accurate and reliable.
“Underwriting is the heart of what we do,” Michalek said. “The goal isn’t to approve every transaction, it’s to make the right decision for each one.”
Michalek said Affirm has no incentive to extend credit that consumers cannot repay and that expanding approvals depends on improving the company’s ability to distinguish between transactions it is willing to finance and those it is not.
Affirm said its underwriting approach evaluates each purchase individually based on what it determines a consumer can responsibly repay at the time of the transaction.
