Affirm's Model Approved Files With No FICO Score: What Sequence Data Reads

On September seventeenth, 2026, Affirm said in an investor release that it had put a new underwriting model into production at its U.S. checkouts. The model is transformer-based, it makes decisions in real time, and it was trained on fourteen years of the company's own transaction-level lending data.

The detail that matters here is what it approved. Affirm says the model extended credit to applicants its previous system had declined, including thin-file applicants and applicants with no FICO score at all, and that those approvals produced 3.4% more completed purchases than a control group at comparable risk.

The idea underneath it is worth understanding properly, because it has been implicit in credit modelling for years. A credit score compresses a file into summary statistics. A sequence model reads the order and the timing that the compression throws away. Two files can produce identical summary statistics and be moving in opposite directions.

Start with what a score is, mechanically.

A credit score is a number computed from a file as it stands at the moment of the pull. The inputs are summary statistics: how many accounts you have, how much of your available limit is reported as used, how many payments were late and how recently, how long the file has existed, how often it has been pulled.

Every one of those is a compression. A 30% utilization figure is a ratio computed from the most recently reported statement balance against the limit. The statement balances before that one are on your report, but they are not in the ratio. The ratio is one number from one month.

Scores are not entirely blind to time. Recency and severity of a delinquency sit inside payment history, and length of credit history is a factor category of its own, so a thirty-day late from last month is not treated the same as one from two years ago.

How finely any of that is read is not something I can tell you. Neither FICO nor VantageScore publishes its internals, so nobody outside those companies can say how timing is weighted inside the model, and I am not going to invent a mechanism to fill the gap. What can be said does not require the internals: the output is a single number, and a single number cannot carry an ordering. Whatever a score does with timing, compressing a file to one value discards sequence the file itself still holds. The statement balances and the payment dates are still sitting on the report.

The path is what gets dropped. Whether a balance arrived at 30% on the way up or on the way down is not in the ratio, and a model built on the ratio cannot recover it.

What follows is an illustration, not data. I constructed these two files to isolate the point. No real applicant is described.

Two files, A and B. Both have three revolving accounts. Both have $4,000 in total limits: two cards at $1,500 and one at $1,000. Both report a total balance of $1,200 this month, and $1,200 divided by $4,000 is 30% utilization. Both have exactly one thirty-day late in the past twenty-four months and two hard inquiries in the past twelve. The oldest account on each is twenty-six months old, the average eighteen.

On the summary statistics a score reads, these files are close to indistinguishable.

Now the sequence.

File A carried $3,100 twelve months ago, which against the same $4,000 in limits was 78% utilization. Every statement since has reported a lower balance than the one before it, down to $1,200 now. That is $1,900 cleared across twelve statements, roughly $158 a month. The late payment was twenty-one months ago, near the start of the file, and nothing has been late since. Both inquiries are eleven months old.

File B carried $160 twelve months ago, which against $4,000 is 4% utilization. Every statement since has reported a higher balance, arriving at $1,200 now. That is $1,040 added across twelve statements, roughly $87 a month, while payments moved from clearing the statement in full to paying close to the minimum. The late payment was last month. Both inquiries landed within the last sixty days.

A score is not completely fooled by this. B's late payment is recent and so are its inquiries, and recency is part of what the payment-history and inquiry categories account for, so B would very likely land below A. What the utilization ratio cannot carry is the balance path. Twelve consecutive declines and twelve consecutive increases arrive at the identical 30% and enter the ratio as the identical 30%. Give both files a clean twenty-four months and no recent inquiries, and they become hard to separate on summary statistics while remaining obviously different to anyone reading the order.

Affirm's description of the model is that it identifies patterns within and across accounts, including how those patterns change over time, without a hand-designed feature for each pattern.

That second half is the engineering point. Conventional credit modelling works by feature engineering: a person decides in advance that utilization is worth measuring, writes the definition, and the model learns a weight for it. If balance trajectory over twelve months is worth measuring, somebody has to define it first, over what window and measured how, before any model can use it. Every pattern the model can see is a pattern somebody anticipated.

A sequence model is handed the events in order, with their timing, and works out for itself which orderings carry information. The transformer architecture is the family of model used for language, and it transfers because a credit file, read as a stream of events, has a similar shape to a sentence: the units are ordered, and the meaning of any one depends on what surrounds it. A $400 balance means one thing after six months of $2,000 balances and something else after six months of $50 balances.

Fourteen years of transaction-level data is what makes that learnable. You need a very large number of sequences before the orderings that predict repayment separate from the ones that do not.

The number Affirm published is that approvals from the new model produced 3.4% more completed purchases than a control group, at comparable risk.

Read what that is, and what it is not. It is a relative increase in completed purchases. It is not an approval rate, and it is not a jump of 3.4 percentage points in one. If a thousand checkouts completed under the control system, roughly 1,034 completed under the new one, thirty-four more per thousand. That is a real commercial result and a modest-sounding one, which is about what a genuine underwriting improvement looks like rather than a loosening of standards.

"At comparable risk" is the load-bearing phrase. Approving more people is trivially easy if you will accept worse losses, so the claim only means anything if the risk held. Affirm also says these loans performed better than a similar expansion had under its previous models, which is the comparison that carries the weight: the same widening of the gate, done two ways.

What the figure does not tell you is how many of the additional approvals were thin-file or no-score applicants, or what any individual's odds are. Affirm has not published a breakdown, and I would not infer one.

This is the part most likely to be misread, so I will be blunt about it. The model governs Affirm's own approval decisions at Affirm's own checkouts. That is the entire scope.

It does not change your FICO score. It does not change your VantageScore. It does not change what the nationwide bureaus hold on you or how they hold it. A lender using a different model to read your file does not alter the file.

It does not mean other lenders will approve you. A mortgage underwriter, an auto lender and a card issuer each run their own models with their own risk appetites, and an approval at one checkout carries no weight at any of them.

It also does not mean Affirm has stopped reading credit reports. Sequence data is a claim about how a file is read, not whether it is read.

Cash-flow underwriting and sequence underwriting get discussed as one thing. They are not, and the difference is worth holding onto.

Sequence underwriting does something else. It takes the data that is already on the file and stops discarding its structure. No new source is added. The compression is removed.

The honest answer has not changed, and I am not going to pretend otherwise because one lender shipped a better model.

Build a conventional file. Mortgages, auto loans and credit cards still price off scores. An underwriting model at a point-of-sale lender does not get you a rate sheet at a bank, and the conventional file is still the instrument nearly every consequential credit decision in this country runs on.

So: one starter account reporting to all three bureaus, opened and held. Statement balances kept low relative to limits, because the ratio reads the statement. Every payment on time, because a late payment is the most expensive single thing you can put on a short file. And time, which is the one input you cannot buy.

What has changed is the shape of the interim. There are more lenders able to underwrite a file a score cannot read, and there is now a reason to care about the trajectory of your balances rather than only their level.

A transformer-based model went live at Affirm's U.S. checkouts on September seventeenth, trained on fourteen years of transaction-level lending data. Affirm reports it approved people its previous system declined, including applicants with no FICO score, producing 3.4% more completed purchases at comparable risk.

The idea underneath it is the durable part. A score compresses a file into summary statistics, and compression is lossy. Two files with the same account counts, the same 30% utilization and the same single late payment can be moving in opposite directions, and the direction is not in the ratio. A model that reads the sequence gets the direction back.

None of that touches your credit report or your scores, and none of it predicts what any other lender will do. If you have a thin file, the advice is what it was the week before: build a conventional one, keep it clean, let it age. The interim is slightly less closed than it was. That is the whole of what changed.

1. What did Affirm announce on September 17, 2026?

Affirm said in an investor release that a transformer-based, real-time underwriting model was in production at its U.S. checkouts. It was trained on fourteen years of the company's transaction-level lending data, and Affirm says it approved applicants the previous system had declined, including thin-file and no-FICO applicants.

2. Does this change my FICO score?

No. This is one lender's internal underwriting model. It does not alter your credit report, your FICO score, your VantageScore, or what the nationwide bureaus hold about you. A different way of reading a file does not change the file.

3. What does 3.4% more completed purchases mean?

It is a relative increase against a control group, not an approval rate. If a thousand checkouts completed under the control system, roughly 1,034 completed under the new one. Affirm states the comparison held at comparable risk and that those loans performed better than a similar expansion under its previous models.

4. Why can a sequence model see something a credit score cannot?

A score reads summary statistics computed from the file at one moment. Utilization, for instance, is the most recent statement balance divided by the limit, so the months of balances behind it are not in the number. A sequence model reads the events in order with their timing, so it can distinguish a balance arriving at 30% on the way down from one arriving there on the way up.

5. Does an Affirm approval mean other lenders will approve me?

No. Every lender runs its own models on its own data with its own risk appetite. An approval at a point-of-sale checkout is a statement about one company and one transaction, and it carries no weight with a mortgage underwriter, an auto lender or a card issuer.

6. Does Affirm ignore credit reports now?

No. Reading sequence data describes how a file is read, not whether it is read. Nothing in the announcement says credit reports have been dropped from the decision.

7. If I have no credit score, what should I actually do?

Build a conventional file, because mortgages, autos and cards still price off scores. One account reporting to all three bureaus, held and aged, with low reported statement balances and no late payments. Newer models widen who can get financed in the meantime, but they do not replace the file.

8. What if an automated model declines me?

You are entitled to the specific principal reasons for the decision, the same as with any other credit denial. An automated decision does not reduce that obligation, and "the model declined you" is not a specific principal reason.