Fraud is growing, and it is getting harder to see. The transactions that cost the most look ordinary in isolation and only reveal themselves in context. What we’ve heard is that this is a hard problem to solve from inside your own platform, and only gets more costly.
When we first launched Signal, we set out to make the tradeoff between risk reduction and a great customer experience an easier decision. Since then, Signal has grown more sophisticated and more specific: from point-in-time account checks to behavior measured over time, and from one model for everyone to models trained for the risk patterns of industries like gaming and remittances. Signal now protects $7B in transactions across millions of accounts every month.
We are now announcing the fourth generation of the Signal model with a step change in performance, driven by three new classes of model features.
In testing, the Bank Risk Score caught 126% more risky dollars than v3 at a 5% decline rate, and the Customer Risk score caught 26% more unauthorized returned dollars at a 1% decline rate. We saw these same gains for growth: approve 7% more good dollars at a 1% return rate.
Sequential foundation model
Signal 4 incorporates Plaid's sequential foundation model, pretrained across the network to read transaction streams in order. With it, Signal can learn more behind the meaning of each transaction, its timing, and its position relative to everything around it, alongside account characteristics. Rather than viewing a transaction as an isolated event, the model learns to interpret it within the broader flow of financial activity on an account.
This helps Signal recognize activity that appears out of place in its broader context, and maintain a coherent understanding of an account’s long-term financial patterns. For example, you might have two accounts with the same 60-day transaction count, inflows and outflows, and return history, but transaction volume is steady in one, while the other shows dormancy, then a sudden burst. In aggregate, these two accounts can seem identical, while a payment from the second account is much riskier to approve.
Over time, our sequential foundation model will continue to learn patterns using self-supervised learning tasks derived directly from transaction sequences. Due to the outsized impact of the sequential foundation model, its features now rank among 8 out of the top 20 features across both bank and customer risk models.
Cash flow health
A point-in-time balance check tells you what is in an account, not what is about to leave it. Signal 4 adds income and spending patterns: transaction categorization and diversity, income and expense variability, and whether this is the account holder's primary or secondary account. This deeper look into inflows and outflows helps Signal reveal a better understanding of financial activity and account health to forecast an upcoming shortfall or paycheck.
Return history detection
Not every return is flagged correctly, and some are never reported to Plaid at all. A model trained on incomplete labels learns an inaccurate definition of risk.
Signal now applies AI to raw transaction descriptions to identify and categorize prior returns with more confidence, including returns originated by businesses that never report to Plaid. So a return that nobody reported still shows up in the score.
Improve scores at no additional cost
Signal gets more accurate when it can see what happened after the decision. By feeding your decisions and return outcomes back to the network:
Return events are accounted for accurately and in real time, rather than inferred from transaction history later.
The model retrains on your traffic, so it gets smarter at ranking your returns over time.
Customers who share return data via API have seen a 2-3x increase in dollar recall.
Refreshed Signal dashboard
We’re also shipping two dashboard enhancements to make it easier to check your integration and monitor performance.
Integration Health: a new near-real-time view of Signal calls, decisions, returns, and errors, and alerts you of any integration issues.
Analytics: now displays return rates with a dollar/count toggle, top return codes, and institutions ranked by return volume.
Together they answer the two questions risk teams ask most often: is my integration actually feeding the model everything it needs, and where are my returns coming from?
Get started
Signal 4 rolls out broadly in early Q4 2026. For customers already using Signal, your scores will remain stable and your score distribution will stay consistent, so you do not need to change your threshold or rules.
Not using Signal yet? Your next linked account probably isn't new to us. Get in touch with us here.