5 mins read
Affirm underwriting has moved to a new architecture. The company said on September 17, 2026 that a transformer-based model is now live at checkout in the United States, and that its first deployment approved enough previously declined applicants to produce 3.4% more completed purchases against a control group. Affirm also says those additional loans performed better than a comparable approval expansion run on its older machine learning models.
Fintechbits analysis
Why the Architecture Matters More Than the 3.4%
Affirm’s previous models fed on credit bureau summaries: balances, utilisation, account counts, payment history. Those figures compress a moving credit file into a snapshot. A transformer reads sequence instead, so the order and timing of events inside and across accounts becomes signal on its own, without an engineer designing a separate feature for each pattern in advance.
The company has been underwriting purchase by purchase for 14 years, which means it holds the one input this architecture needs: a long history of labelled repayment outcomes at the individual transaction level. Most lenders have outcome data at the account level and per month. Affirm has it per purchase, which is a materially richer sequence to learn from.
“We have steadily accelerated the amount of data we use to train each generation of our underwriting models,” said Libor Michalek, Affirm’s president. “What is exciting about the transformer model architecture is that we can now find new information within the data we already have.”
The Explainability Claim Inside Affirm Underwriting
Affirm also says it built a proprietary algorithm that keeps the model as explainable as traditional machine learning while staying fast enough to score in real time at checkout. That claim deserves the most scrutiny in the announcement. Explainability is not a nice extra for a United States lender, it is a legal requirement. Adverse action notices under Regulation B, administered by the Consumer Financial Protection Bureau, must give specific principal reasons for a decline.
Sequence models are awkward on exactly that point, because the thing driving the decision may be a pattern across time rather than a single variable a human can name. Affirm says validation and ongoing monitoring confirm the explanations are accurate and reliable. That is the right answer, and it is also an answer nobody outside the company can verify.
Affirm Is Not Alone in Using Sequence Models
Transformers applied to credit sequences are no longer exotic. Several large card issuers and bureau-adjacent vendors have published on similar approaches, and the wider pattern of financial institutions training domain-specific models is visible across the sector, including Ant Group open-sourcing a financial AI model earlier this month. What Affirm has that most competitors do not is purchase-level outcome data on a population that includes thin-file and no-FICO borrowers, plus a business model where it absorbs the loss when a decision goes wrong.
That last point is the real discipline. Affirm charges no late fees, so it cannot recover a bad decision through penalty income the way a card issuer can. Competitors in the same checkout aisle, including Klarna as it expands its merchant footprint, face the same structural test on credit quality rather than distribution.
What the 3.4% Figure Does and Does Not Say
On its own terms the result is good: an incremental approval expansion that performs better than the last expansion is what underwriting progress actually looks like. But 3.4% more completed purchases in an initial deployment is not 3.4% growth in gross merchandise volume, and the announcement says nothing about how long those loans have been seasoning.
Buy now, pay later credit sours quickly when it sours, though short duration at least means early performance is more informative than it would be on a mortgage book. Affirm has earned more benefit of the doubt on underwriting discipline than most of its peers, which is not the same as none. The interesting disclosure would be the loss rate on the incremental cohort, and that is not in here.
Affirm reports fiscal first-quarter results in November. Watch whether the approval expansion shows up in volume, and whether delinquency holds while it does, on the company’s investor relations site.
What to Watch Next in Affirm Underwriting
Three disclosures would settle whether Affirm underwriting has genuinely improved. The first is the loss rate on the incremental cohort, the applicants the old model would have declined, reported separately rather than blended into the book. The second is whether the 3.4% lift in completed purchases holds as the model meets a full seasonal cycle including holiday volume, when basket sizes and applicant quality both shift. The third is regulatory: if a state or federal examiner tests the adverse action reasons the new Affirm underwriting stack produces, the outcome will matter to every lender considering sequence models. Affirm reports fiscal first-quarter results in November, and the shareholder letter is usually more forthcoming on credit than the press release was here.
