Affirm Bets on Where Borrowers Are Headed

Watch more: Monday Conversation With Affirm’s Libor Michalek

Picture someone who had a bad year. A few bills went unpaid, and their credit score took the hit. Then things turned around. Every payment since has been on time. The score hasn’t caught up.

Now picture someone whose trouble is just starting. A missed payment here, another there, spread across a few accounts.

On paper, those two people can look the same. Same number of missed payments. Same answer at checkout, and it’s usually no.

Last week, Affirm announced that a new underwriting model built to tell those two apart is live at U.S. checkouts. In a conversation with PYMNTS CEO Karen Webster, Affirm President Libor Michalek went past the announcement and into the parts that don’t fit in one: what the old models were missing, who stands to gain the most and what it sets up for Affirm next.

What they were missing, he said, was time.

“The time dimension of multiple purchases, multiple credit events in a customer’s life wasn’t being represented with particularly high fidelity,” Michalek told Webster.

Put simply, the older models were good at counting what happened. They had a harder time with when it happened and what came after. The new one draws on Affirm’s 14 years of lending history and uses transformer technology, the same kind behind large language models, to follow the story in order.

Five of the 10 Deserved a Yes

Michalek gave Webster an easy way to think about it. Take 10 applicants an older model would turn down. It declines all of them because it can’t tell which ones will get in over their heads. A better model might spot five who can repay. The other five still get a no.

The bar didn’t move. Affirm can now see who clears it.

For those five people, that’s credit they would have been denied, which is a familiar story for a lot of subprime and near-prime consumers. For Affirm, it’s five customers it was sending away.

Thin Files Have the Most to Gain

Webster asked the natural follow-up. What happens when there’s barely any story to read?

Michalek said even a short history has an order to it. Credit bureau records, Affirm’s own data and cash flow information that users choose to share all add to the picture.

“Where before two very different users could look the same because of that sparseness of data, we’re really able to tease them apart,” he said.

This is where Affirm saw its biggest jump. Among consumers with something on their credit report but not enough for a FICO score, the new model improved Affirm’s ability to rank who’s likely to miss a first payment by about 2.1 times as much as the next version of its older model did. One caution on that number. It measures how well the model sorts risk, so it shouldn’t be read as twice as many approvals.

Michalek also said Affirm will keep reporting repayment behavior to the credit bureaus. Affirm can make some decisions without a FICO score, but the rest of a customer’s financial life still leans on one. It does raise a question worth watching. As more lenders learn to read the behavior behind the score, how much will the score itself matter?

A Good Prediction Still Had to Wait a Year

Michalek told Webster that Affirm had transformer models in testing more than a year ago. The predictions were good. That wasn’t enough to put them in front of customers.

Someone who gets declined needs to know why. The same application has to get the same answer every time. And all of it has to happen in under a second, while a shopper waits at checkout. Affirm kept working until the model could do all three.

What launched is a pairing. The transformer reads the payment history, and its read feeds into a more traditional machine-learning model alongside established credit, merchant and user information. Affirm said the pair beat the transformer on its own. Its first job was a second look at applicants the old system had declined. Next, it takes on full underwriting for new users, approvals and declines alike.

A Spending Limit Changes the Conversation

All of this matters more because of how people now use Affirm. Its latest earnings showed more purchases at smaller amounts, which is what an everyday way to pay looks like.

Pay later has become working capital for the household. Consumers use it to match what they owe to when they get paid, across groceries, clothes and car repairs.

Everyday use brings its own challenge. One small purchase is easy to handle. Several of them, with payments landing around the same time, can be a different story. The model has to see how a new payment fits alongside what a customer already owes.

The Affirm Card shows how that reaches the customer. Anyone with a credit card knows their credit limit, and it answers one question: How much room do I have? The Affirm Card presents a spending limit, lays out payment plans and lets users weigh options such as the fastest payoff or the lowest interest. The questions change. What am I buying? What will it cost to pay over time? Which option works for me?

The customer chooses within guardrails. The underwriting decides where the guardrails sit.

And every purchase and repayment teaches Affirm a little more about that customer. Spending power can grow when the behavior supports it and hold steady when it doesn’t. That’s how a pay-later option works its way to the top of the wallet.

Asked how he’ll know a year from now whether it worked, Michalek pointed to approvals, delinquencies, customer growth and usage. The evidence, he said, will show up in quarterly results.

The first test may come sooner than that. The holiday season is weeks away, which means more purchases in a shorter window and more payments coming due together in the new year. A model built to know where a borrower is headed will gets its chance to show it can earn the customer’s next purchase, and the one after that.