Most models in financial services learn one task. A foundation model learns the domain. Train one on enough financial activity and it builds a general understanding of how financial behavior evolves over time, which can then be adapted to underwriting, payment risk, or cash flow intelligence with a fraction of the labeled data a purpose-built model would need.
In May we introduced two of them. The transaction foundation model learns what an individual financial event means. The sequential foundation model learns how those events relate to one another over time.
The case for building this way was always that a model reading a broad view of the sequence of someone's financial behavior would outperform one reading a summary of it. We tested that hypothesis. In ACH payment risk, the model caught more return dollars at the same action rate. In underwriting, it produced less default risk at a fixed approval rate. In cash advance, lower losses. Same integration, same data, read more carefully.
A score that is an input to a lending decision has to come with reasons a person can understand. A score supporting a payment has to arrive before the payment does. So we built both: a framework that links the model to the reason-code categories credit and cash advance teams already use, and the infrastructure to run these models in real time, everywhere we need them.
Giving a reason, not just a score
Financial decisions require more than a score. The analyst reviewing a borderline credit application needs to know what pushed it there. The risk team about to move an approval threshold needs to know what the score is responding to. And when a lender declines an application, they need to tell the applicant why, with clear reasons such as a high expense-to-income ratio.
When a score is built from features we define ourselves, those reasons come easily, because the features are already written in human language. A sequential model works differently. Its output rests on the shape of account history rather than on a list of features, which makes for a better decision and a harder one to put into words.
To trace each score through our model, we used integrated gradients. We started from a reference applicant, a composite at the center of the population, and moved step by step toward the real applicant, measuring how sensitive the score was to each input along the way. The result is a contribution for every transaction in the history, and together these account for the gap between the real applicant's score and the reference score. We total those by category, rank whatever pushed the score toward decline, and surface the largest as candidate reason codes — each one traceable to the specific transactions behind it — for lenders to incorporate into their own adverse action notices.
That capability is what makes LendScore Arc possible: our credit risk score built on the sequential foundation model, and it arrives with reason codes built in.
Fast enough for a live payment
Plaid Signal estimates the likelihood an ACH payment will return, and it has to make that prediction before the payment is initiated. Our first foundation-model-powered version took roughly 136 milliseconds to score a batch of 300 transactions in testing. Accurate, but too slow to run at Signal's volume for every payment.
Profiling showed where the bottleneck was. Around 80% of the time was spent not in the sequence model but in the transaction encoder, since a 300-transaction history means 300 passes through that encoder before the sequence model does any work at all.
So we made the transaction encoder smaller. Using knowledge distillation, a machine learning technique in which a larger model teaches a smaller one, we trained a half-size encoder on large-scale transaction data to reproduce the original encoder's outputs. It processes a transaction in half the time and agrees with the original almost perfectly.
The second change was in how the model runs on the chip. Compiling it specifically for the GPU it serves on, and doing the math in a lower-precision format chosen for the task, doubled throughput again with no detectable change in output.
In testing, 136 milliseconds became 38. Just as importantly, the slowest 1% of requests now land within half a millisecond of the median. Predictability matters as much as speed when a model sits inside a live payment.
What this changes across our products
Three products are running on the sequential foundation model today and require no integration changes:
LendScore Arc is our first transformer-based credit risk model. It uses the sequential foundation model to capture signals from the order, timing, and interactions between transactions, then combines them with LendScore's established underwriting features. Every score comes with ranked reason codes for lenders' own adverse action notices. In early testing, Arc increases approvals by 20% for subprime borrowers.
Signal scores ACH transactions for return risk using an account’s transaction history in order directly within the payment flow. The sequential foundation model helps Signal catch 26% more risky dollars at a 1% decline rate.
Cash Advance Index reads repayment behavior as it develops across pay cycles instead of scoring a snapshot of an account, resulting in 10% fewer dollars lost.
What's next
Our foundation models still learn from transactions alone. The next versions add account balances, connection history, and financial product usage, which gives every downstream product a more complete read on the financial activity it already scores.
The grammar is learned and the model is fluent. The next chapter expands its vocabulary.
Plaid is investing in intelligent finance at the layer where it compounds, reaching every product built on top of it.
If you're building AI-powered financial experiences and want to go deeper on how these models power use cases like cash flow underwriting or ACH risk, we'd love to connect.