Affirm Transformer Underwriting Model Raises Approvals, but Explainability Sets the Limit

Affirm deployed a new transformer underwriting model after an experiment produced 3.4% more completed purchases among eligible new users. The system approved applications that its existing models would have declined, including some consumers without FICO scores. The central tension is immediate: Affirm wants broader approvals without accepting greater credit losses.

The model is not a general-purpose chatbot attached to a lending workflow. It analyzes the order and timing of events across a consumer’s credit history. Affirm then combines that learned representation with its established underwriting features and decision system.

That architecture challenges the feature engineering used by many credit models. However, better prediction alone cannot settle the case. Affirm must also explain individual denials, control checkout latency, prevent unfair outcomes, and show that additional loans perform through complete repayment cycles.

The Affirm Transformer Underwriting Model Is Already Making Decisions

Affirm has moved its transformer system beyond research and into live U.S. checkout decisions.

The company announced the rollout on September 17, 2026. Its transformer underwriting system is initially focused on consumers without completed repayment histories on Affirm.

That group presents a difficult underwriting problem. Affirm has little or no internal repayment history for a new applicant. It must draw more information from credit-bureau records, the requested purchase, and the merchant involved.

Traditional underwriting models often reduce a credit report to engineered variables. Those variables can include balances, utilization, account counts, inquiries, and late-payment totals. Each measure captures useful information, but it compresses a changing history into a fixed summary.

Affirm’s transformer instead processes structured records from the credit report. A transformer is a neural-network architecture that can learn relationships among elements in an ordered sequence. Here, those elements represent credit accounts, inquiries, and monthly payment histories rather than words.

The difference matters when two consumers share identical summary statistics. Each person might have two late account-months, for example. Yet recent late payments spread across several accounts can imply a different risk pattern from older problems concentrated in one account.

A conventional model can capture that distinction if engineers design the right variables. They might calculate delinquency recency, concentration, or changes in utilization. However, every new variable requires a hypothesis, historical reconstruction, testing, and production maintenance.

The transformer can learn additional interactions without requiring engineers to define every pattern beforehand. It creates embeddings, which are compact numerical representations of the underlying credit records. Those embeddings preserve signals that simple counts or averages can lose.

Affirm did not replace its entire underwriting stack with one neural network. Its hybrid architecture uses two stages.

First, the transformer learns representations from detailed credit-report information. Second, an accelerated failure time XGBoost model combines those representations with established credit, merchant, and consumer features. XGBoost is a tree-based machine-learning method widely used with structured business data.

That distinction prevents an easy but inaccurate headline. Affirm has not shown that a transformer alone beats its mature underwriting system. The company says its transformer-only candidate never performed well enough to justify production.

The reported improvement came from combining learned representations with existing models. In other words, the new system extends Affirm’s established approach instead of discarding it.

Its first production role was also deliberately limited. The transformer gave a second look to eligible applications that the existing system would decline. It could produce extra approvals, but it could not initially overturn an existing approval with a rejection.

Affirm says the model now handles full underwriting for new users, including approvals and declines. The company also plans to extend the next model iteration across new and returning users.

That expansion changes the stakes. A model that only rescues selected applications has a narrower effect than one controlling both sides of a credit decision. Full deployment makes accuracy, stability, and explanations more consequential for every affected applicant.

Why Sequence Data Changes the Credit Decision

The important advance is not the transformer label; it is Affirm’s attempt to preserve time and relationships inside a credit report.

Credit reports contain nested histories. Each account has an opening date, balance, limit, status, and payment record. Those accounts also interact with inquiries, utilization changes, and other borrowing activity.

Summary features can hide the order of those events. A total late-payment count does not reveal whether problems are recent or receding. Aggregate utilization can also miss a sudden balance increase on one account.

Affirm converts tradelines and inquiries into tokens. A tradeline is the record for an individual credit account. Each token combines numerical fields, categorical fields, and an encoded payment sequence.

The model then uses attention, a mechanism that weighs relationships among different records. This allows an account’s history to influence how the system interprets another account or recent inquiry. Temporal encoding also helps the model distinguish older information from recent changes.

Affirm trains these representations against actual repayment outcomes. The objective is not to reproduce a credit score or generate a textual explanation. It is to identify patterns associated with repayment risk.

The transformer’s output becomes an input to the second-stage XGBoost model. That model also receives established underwriting variables, including transaction and merchant information. The final score enters Affirm’s broader decision system.

This setup reflects the nature of checkout lending. A consumer is not requesting a permanent credit line in isolation. Affirm evaluates a specific person, purchase, merchant, amount, and moment.

A travel purchase illustrates why transaction context can matter. The time between booking and departure can affect risk and repayment behavior. A model based only on a static consumer score would miss that purchase-level distinction.

The approach also targets consumers with thin credit files. A thin-file consumer has some credit-report information but lacks enough history to generate a standard FICO score. Missing that summary score does not mean the underlying records contain no useful signal.

Affirm reports that the hybrid model produced 2.1 times the improvement of its next XGBoost candidate among thin-file consumers. The comparison concerns improvement in risk ranking against the same production baseline.

Across the broader evaluated population, the company reports 1.8 times the gain achieved by the next XGBoost iteration. Its technical analysis expresses those gains through first-payment delinquency ranking.

First-payment delinquency measures whether an initial scheduled payment becomes overdue by a defined threshold. Affirm evaluated a version using a 60-day threshold, called FPDQ60. Better ranking means the model more effectively separates higher-risk loans from lower-risk loans.

Affirm also reported relative improvements in area under the receiver operating characteristic curve, or ROC-AUC. That measure evaluates ranking quality across classification thresholds. The hybrid model improved 3.83% against production, compared with 2.16% for the next XGBoost candidate.

Those figures require careful interpretation. A relative AUC gain is not a 3.83% reduction in defaults. It is also not a 3.83% increase in approval rates.

The most concrete commercial result came from the online experiment. Affirm says incremental approvals lifted end-to-end conversion by 1.2 percentage points, equal to a 3.4% relative increase.

That outcome connects model quality directly to merchant performance. More approved shoppers can mean more completed purchases. Yet the value survives only if repayment performance remains controlled.

The Real Contest Is Better Risk Ranking Versus Looser Credit

Affirm must prove that its additional approvals come from sharper risk selection, not a temporary relaxation of lending standards.

Every credit model faces a tradeoff between approvals and losses. Lowering a decision threshold normally approves more applicants, but it also admits more risk. A stronger model should improve that frontier.

Affirm says the incrementally approved loans performed better than a comparable expansion under its previous machine-learning models. That is the key claim behind the launch.

The company did not merely report that conversion increased. It says the extra loans produced through the transformer delivered better repayment performance than an equivalent approval expansion under earlier XGBoost models.

That comparison is more useful than a raw approval-rate increase. Almost any lender can approve additional borrowers by accepting more expected loss. The harder task is finding applicants whose risk was previously misunderstood.

Thin-file consumers provide the clearest test. Some lack a FICO score because their history is limited, not because their behavior necessarily signals high risk. A system that extracts more information from their underlying accounts can identify borrowers obscured by conventional summaries.

However, the published evidence still comes from Affirm. The company provided relative model comparisons, an online experiment, and initial loan-performance observations. It did not publish loan-level data that outside researchers could reproduce.

The phrase “performed better” also leaves important details undisclosed. Affirm did not provide the absolute delinquency rate for the incremental group. It did not publish the experiment’s sample size or demographic distribution.

Those omissions do not invalidate the result. They limit what outside observers can conclude from it.

Affirm’s scale makes even a modest conversion improvement commercially meaningful. Its 2026 operating scale included 27.8 million active consumers as of June 30, 2026. That figure rose 21% from the prior year.

The company processed $50.2 billion in gross merchandise volume during its 2026 fiscal year, a 37% annual increase. Transactions per active consumer reached 7.0, up from 5.8 one year earlier.

Affirm also reported that consumer transactions increased 45% during the fiscal year. Greater activity supplies more training examples and repayment outcomes. It also magnifies the consequences of systematic model errors.

The resulting feedback loop is central to Affirm’s competitive argument. More transactions produce more performance data. Better data supports improved underwriting, which can increase approvals and merchant conversion.

That loop places pressure on other checkout lenders, including Klarna, Afterpay, PayPal, Sezzle, and Zip. They do not need to adopt Affirm’s exact architecture. They do need to keep approval quality, conversion, and funding economics competitive.

The contest is therefore not “transformers versus no AI.” Credit providers have used statistical models and machine learning for years. The practical contest concerns who can safely identify repayable transactions that rival systems decline.

Traditional banks face a related pressure. Their models often evaluate consumers at the account level, while Affirm underwrites individual purchases. Transaction-level decisions allow Affirm to vary terms or decline one purchase without making a permanent judgment about every future request.

This flexibility can improve precision. It can also make decisions harder for consumers to anticipate. A shopper approved yesterday can still be declined for another transaction today.

The model’s success will depend on whether improved ranking remains stable across merchants, products, economic conditions, and consumer groups. A checkout experiment is an important signal, but it is not the final answer.

Explainability Is the Constraint Affirm Cannot Optimize Away

A credit model must do more than predict accurately because U.S. lenders must provide specific reasons for adverse decisions.

Transformers create an explainability challenge. Their predictions emerge from interactions among many records and learned representations. Those representations do not automatically translate into a reason a consumer can understand.

Affirm says it developed a proprietary attribution algorithm for the hybrid model. Attribution methods estimate which inputs most influenced a particular output. The company says its method provides the same rigor as its established XGBoost explanations.

Affirm also says it validates and continuously monitors those explanations. However, the company has not released the attribution algorithm or an independent assessment of its accuracy.

That verification gap matters because a plausible explanation is not necessarily a faithful one. A post-hoc method can produce a readable reason while failing to identify what actually drove the model.

The Consumer Financial Protection Bureau has made the standard clear. Its adverse action rules require creditors to provide specific principal reasons when taking adverse action.

Complexity does not create an exemption. A lender cannot defend an unclear notice by arguing that its model is too opaque to interpret.

Later CFPB guidance emphasized that generic categories may also be inadequate. If behavioral data drives a decision, a broad phrase such as “purchasing history” might not identify the relevant behavior.

Affirm’s challenge is especially demanding because its transformer captures interactions across time and accounts. The actual decision might depend on recency, account distribution, repayment patterns, and transaction context together.

Reducing that interaction to one or several accurate reasons requires care. The explanation must reflect the real model, remain understandable, and work within operational notice systems.

Fair-lending risk presents a second test. Credit records can reflect historical inequality, uneven access, reporting errors, and different borrowing patterns. A model can learn those patterns even when protected characteristics are excluded.

A stronger risk predictor is not automatically a fairer predictor. Affirm must test outcomes across relevant groups and investigate disparities that cannot be justified by legitimate credit risk.

Thin-file approvals make this question particularly important. Better use of limited records can widen access for consumers poorly served by conventional scores. It can also create new disadvantages if learned patterns act as proxies for protected traits.

The company’s public materials do not provide demographic approval or error rates. They also do not disclose how the model behaves when bureau records contain missing, stale, or disputed information.

Data consistency poses another risk. Historical training examples must reflect only information available when each original decision occurred. Otherwise, future information can leak into training and create misleading evaluation results.

Affirm says it reconstructs inputs at the relevant historical point and validates production artifacts against development behavior. It also standardizes categorical values, missing fields, padding, and temporal calculations.

Those controls are essential because minor processing differences can alter a score. They become more difficult as a system combines transformer inference, data preparation, and a downstream tree model.

The proprietary nature of the system leaves regulators, borrowers, and outside researchers dependent on Affirm’s disclosures. Independent validation would strengthen the company’s case, particularly for explanations and group-level outcomes.

Until then, the reported performance should be treated as promising company evidence. It is not proof that every decision is more accurate, fair, or understandable.

Real-Time Loan Decisions Leave Little Room for Model Complexity

Affirm’s production achievement is partly an infrastructure story because the model must finish its work within a sub-second checkout budget.

Credit decisions at checkout compete with a shopper’s patience. A model that improves risk prediction but adds visible delay can reduce conversion before it approves anyone.

Affirm says the complete computation path must remain within a sub-second latency budget. That path includes preparing structured inputs, running transformer inference, generating embeddings, and evaluating the XGBoost model.

The company kept the transformer’s parameter count relatively small. It also optimized both CPU and GPU serving paths. This is a different design target from large language models that generate text over several seconds.

The model’s size has not been disclosed. Nor has Affirm published median or tail latency after deployment. Tail latency measures slow requests that can damage the experience even when average performance appears acceptable.

Reliability matters just as much as speed. Credit reports vary in length and structure. Some contain numerous accounts, inquiries, and long payment histories, while others contain sparse or missing fields.

Affirm normalizes numerical values, assigns consistent categorical encodings, and masks padded positions. Padding lets records of different lengths fit into standardized computation batches without treating empty positions as information.

The training system uses distributed GPUs, mixed-precision computation, and streamed data. Mixed precision lowers memory requirements by using reduced numerical precision where appropriate. Gradient accumulation helps simulate larger training batches without storing every example simultaneously.

Those techniques make experimentation more efficient, but production creates a separate problem. The exported model must reproduce the behavior tested during development. Differences in preprocessing or numerical execution can change risk scores.

Affirm says it validates model outputs during the export process. The serving system then connects online data processing, transformer inference, and the existing underwriting stack.

This hybrid design reduces some implementation risk because the company retains established features and decision infrastructure. It also adds dependencies. A failure in either stage can affect the final risk score.

Model monitoring must therefore cover more than overall delinquency. Engineers need to watch input distributions, missing values, embedding behavior, explanation stability, latency, and downstream approval patterns.

Economic change adds another layer. A model trained on historical repayment patterns can weaken when interest rates, employment conditions, consumer behavior, or merchant mix shifts.

Sequence modeling does not remove that problem. It can potentially detect richer patterns, but those patterns still come from historical data. Relationships learned during one credit environment may not remain stable in another.

Affirm’s growing mix complicates evaluation further. Interest-bearing installment loans accounted for 70% of fiscal 2026 gross merchandise volume. Pay-in-X represented 16%, while zero-interest monthly installments represented 14%.

Each product can have different durations, consumer profiles, merchant economics, and repayment behavior. A model that performs well for one segment may require recalibration elsewhere.

The company plans to learn shared representations from bureau records and internal Affirm activity. Such representations could support several credit tasks without rebuilding features separately for each use case.

That strategy could accelerate future model development. It could also expand the impact of a flawed representation across several decisions. Shared foundations demand stronger validation because errors no longer remain isolated within one model.

The infrastructure achievement is therefore real but incomplete. Affirm has shown that transformer-derived features can fit inside live checkout underwriting. Long-term operations will determine whether the approach remains fast, stable, and auditable at scale.

Three Signals Will Decide Whether the Model Delivers

The next evidence should come from repayment performance, full-decision deployment, and transparent validation rather than another model benchmark.

The first signal is seasoned credit performance from the incremental approvals. “Seasoned” loans have existed long enough for meaningful repayment outcomes to emerge. Early performance can look favorable before later installments become due.

Observers should watch Affirm’s delinquency rates, net charge-offs, and provision trends over the next several reporting periods. They should also compare those measures with approval growth and product mix.

The company’s fiscal 2026 provision for credit losses increased 29%, while average loans held for investment increased 27%. Those figures cover the broader portfolio and predate a full evaluation of this rollout.

A modest movement in aggregate losses would not establish that the model failed. Loan growth, economic conditions, and mix can affect those numbers. The stronger evidence would connect model-driven cohorts with comparable control groups over complete repayment windows.

Stable cohort performance would support Affirm’s central claim. Rising losses concentrated among incremental approvals would weaken it, even if checkout conversion remained higher.

The second signal is performance after the model gains authority over both approvals and declines. The initial second-look experiment only tested whether it could rescue selected applications.

Full underwriting is more difficult. The model can now reject applicants whom the previous system might have approved. That makes false negatives, explanations, and consumer experience more visible.

Affirm also plans to deploy the next iteration for returning users. Those consumers bring internal repayment histories that were largely absent from the initial target population.

Success among returning users would show that the architecture generalizes beyond thin-file and first-time applicants. Weak results could indicate that learned bureau representations add less value when Affirm already has detailed internal data.

The third signal is evidence about explanations and fairness. Affirm says its proprietary method matches the rigor of existing XGBoost explanations. Independent testing, regulatory review, or detailed methodology would make that claim more credible.

The issue extends beyond compliance paperwork. Consumers need accurate reasons to correct errors or improve their credit position. Merchants and funding partners also need confidence that expanded approvals do not introduce hidden legal risk.

Broader industry context raises the stakes. The CFPB’s BNPL market data shows that six surveyed lenders originated 335.8 million loans during 2023. Their combined dollar volume reached $45.2 billion.

That market has continued moving beyond short pay-in-four products into longer installment credit and card-like services. As loan structures expand, underwriting decisions carry greater financial consequences.

Affirm’s launch is therefore more than an attempt to attach a fashionable architecture to lending. It demonstrates a practical route for using sequence models inside a regulated, latency-sensitive decision system.

The architecture also carries a useful lesson for enterprise AI teams. The transformer succeeded as one component inside a mature stack. Affirm retained structured features, XGBoost, controlled experiments, monitoring, and operational constraints.

What should readers ask next? Look for seasoned repayment data, results from full underwriting, and evidence that explanations reflect the model’s actual decisions. If those signals hold, the Affirm transformer underwriting model will represent a meaningful advance in credit selection. If they do not, the conversion lift will have measured only the beginning of the risk cycle.