A strong credit history can tell a lender that someone has paid their debts on time. It can’t necessarily predict whether they can keep doing so.
For example, a borrower could still make payments on time while simultaneously drawing down savings. Or behind the scenes, they could be relying on cash advances or going without a paycheck. Those signals can show up in bank transactions before they appear as missed payments on a credit report.
That is one of the problems Plaid** is trying to address with two credit risk models it introduced earlier this month: LendScore 2, an update to its original model, and LendScore Arc, which adds analysis of the timing and sequence of bank transactions to help lenders assess borrowers.
“The cash flow data also helps you spot people who may look really good on a credit file, but actually are in [financial] trouble,” Brett Manning, a credit product manager at Plaid, told This Week in Fintech in a recent interview.
Much of the argument for using bank transactions in lending has focused on expanding access for people with limited or poor credit histories. But Manning points to another use — identifying borrowers who would ordinarily qualify but may struggle to repay.
San Francisco-based Plaid was founded in 2012 and started by connecting consumer bank accounts to financial applications. It has since pivoted into lending, identity verification, credit reporting, fraud prevention and payments. It introduced its original LendScore in October 2025, combining bank transactions and balances with information from activity across Plaid’s network, such as recent connections to lending applications, to estimate repayment risk.
The latest iteration updated the core model and introduced specialized versions for auto, home and short-term lending. It also introduced LendScore Arc, which uses a different approach to analyzing transaction data.
Both LendScore 2 and Arc use cash-flow and network information to assess repayment risk. Arc adds the ability to learn patterns from the order and timing of transactions, rather than relying entirely on financial measures engineers define in advance.
Manning said Plaid expanded the loan repayment records used to train its models for the latest tools, and sought to make the dataset more representative of the broader U.S. population. The training data included more tradelines, while the models incorporated thousands of features rather than hundreds.
The company tested more than 20 variations for different borrowers and loan types before settling on the expanded lineup, he said.
Suddu Seshadri, Plaid’s head of data and AI, says LendScore 2 delivers 42% greater predictive power than traditional credit data alone. That comparison differs from its reported results for Arc, which measure improvement over Plaid’s own core model.
Previously, engineers defined the measures a model would examine, such as average balances, total deposits or spending over a particular period. That approach depends on engineers anticipating which characteristics will help predict repayment.
Arc combines established financial measures with a model that learns additional patterns from the order and timing of individual transactions. It uses a transformer, the same broad type of technology behind large language models, to analyze those relationships.
Relying on summaries can create limits, noted Seshadri. Monthly totals show how much was deposited and withdrawn but do not preserve the sequence connecting those events.
“What we are trying to do is model this as a backbone in terms of how things are happening in terms of transactions, or in terms of accounts and connections, and trying to look at what is in order, what can be predicted in a pattern, what is an anomaly, and what should we do with that anomaly or with the pattern itself,” Seshadri told This Week in Fintech.
Suddu Seshadri, Plaid’s head of data and AI
For example, whether expenses cluster immediately after payday or spread throughout the month can help predict repayment, Manning said. He emphasized that Plaid is not asking a chatbot to make lending decisions. The model produces a score lenders can use alongside credit reports and other information.
Early testing found that Arc improved predictive performance by 20% for deep-subprime borrowers and 24% for super-prime borrowers compared with Plaid’s core model, Seshadri said. Those figures measure predictive improvement, rather than equivalent changes in approvals or defaults.
The results suggest different uses at opposite ends of the credit spectrum.
For applicants with weaker credit profiles, bank transactions could show they can afford a loan. Manning pointed to borrowers who consistently add to savings or have enough income left after expenses to take on debt.
Brett Manning, Plaid’s credit product manager
For applicants with strong credit, those transactions could expose problems that have not yet affected their repayment records.
Such signals can include overdrafts, essential versus discretionary spending and active credit seeking.
“We can see that in cash flow data before it necessarily turns up in late credit card payments,” he said.
Plaid trains its models on different earnings patterns and also uses incomplete and disrupted account histories during training to help the models recognize missing information that would otherwise send incorrect signals. For example, a declining savings balance could reflect a planned expense. Irregular deposits could be normal for a freelancer. Income arriving in an unconnected account could leave a lender with an incomplete picture.
Identifying those gaps can flag where a borrower needs to provide additional information, Seshadri said.
The model’s complexity makes it hard to explain why it produced a particular score.
When engineers define a financial measure in advance, its meaning is relatively straightforward. A relationship the model learns among transactions can be harder to describe.
Plaid says it measures how individual transactions contribute to a score, groups those contributions into financial categories and supplies reasons lenders can incorporate into explanations of their credit decisions.
“The explanation shouldn’t be like, hey, AI found something,” Seshadri said.
Plaid’s reported improvements come from its own testing. The company did not provide customer results in the interviews showing how Arc has changed loan approvals or defaults in practice.
Manning argues that spotting financial trouble earlier can also protect borrowers from taking on debt they cannot manage.
“No one is served well by receiving a loan that they can’t afford to repay,” he said.
**(Disclosure: Plaid is the parent company of This Week in Fintech; it does not have any say in our editorial process)
