The Data Already Exists, So Why Isn’t Credit Reaching the People Who Need It?

Bank transaction data already shows who can repay, but it only changes a lending decision once it is permissioned, interpreted, explained and delivered into the lender’s workflow.

In short

The data to lend to more creditworthy borrowers already exists. What is missing is the work between the data and the decision.

• A decline for a thin-file applicant can reflect missing history rather than evidence that they would not repay.
• The path from data to decision breaks at five points: permission, interpretation, model fit, explainability and workflow delivery.
• Raw transactions only become a credit signal once they are categorized accurately, turned into stable features and calibrated to the lender’s own population and product.
• Lenders do not need to replace their stack. A second look at borderline and thin-file applicants is a practical starting point, and policy and final decisions stay with the lender.
• Wider access only counts if risk stays within appetite, so results should be measured on the lender’s own repayment outcomes.

What does it mean to say the data already exists?

Most adults leave a detailed record of their financial lives in their bank accounts. Deposits show when and how income arrives. Balances show how much cushion a person keeps. Recurring debits show what they have already committed to pay, and when.

With a customer’s permission, a lender can see this record through open banking and account aggregation connections. It is the same information a careful loan officer would ask about in conversation: is income steady, what is already spoken for, and what happens in the days before a payment is due.

The signal is there. Yet a large group of borrowers remain stuck at the edge of the credit system. Some have little or no bureau history. Others have a file that does not reflect their current circumstances. For many of them, a decline reflects missing information, not evidence of repayment risk.

So the real question is why access to data has not become access to credit. The answer sits in the steps between the two.

Where does the path from data to credit decision break?

Getting access to data is the first of five steps, and most lenders stall at one of the other four.

1. Permission. A lender needs clear consent for what it collects, how it uses the data and who it can share it with. Teams that cannot answer those questions find their use of data blocked by legal, privacy or partner reviews, however strong the signal is.
2. Interpretation. Raw transactions are messy. Merchant names are inconsistent, transfers between a customer’s own accounts can look like income, and gig or shift pay does not follow a salary pattern. Even a category like grocery spend can mean very different things across households, as we explain in why grocery spend is not a single signal in cash flow underwriting. Without accurate categorization and careful feature design, the data produces noise.
3. Model fit. A model trained mostly on borrowers who already had bureau files can perform less well on thin-file and non-standard applicants. It needs to be calibrated to the lender’s own population and product.
4. Explainability. Credit teams, compliance teams and borrowers need to see why a score came out as it did. A score nobody can explain tends to be sidelined in review, however accurate it is. Explainability has to survive governance, which means reasons that map to the levers a lender’s policy actually uses.
5. Workflow delivery. A signal that lives in a separate dashboard does not change decisions. It has to arrive in the decision engine, loan origination system or review queue the team already works in, through an API or a batch file.

Each of these is a solvable engineering and governance problem. But when any one of them is missing, the data stays unused.

Why do thin-file and non-standard borrowers fall into the gap?

Traditional credit scores summarize past credit behavior. They work best for people with long, conventional credit histories. They say little about someone who is new to credit, has recently changed jobs, or earns through shift work, gig platforms or self-employment.

For these applicants a bureau file offers little to go on. A lender relying on that file alone faces two poor choices. It can decline a borrower who might have repaid, or approve with very little evidence and absorb the uncertainty.

Bank transaction data adds what the bureau file lacks: how much comes in, how steadily, how much is already committed and how much cushion remains. That is a view of capacity to repay, not only a history of repaying. We cover the practical side in how to lend to thin-file customers using cash flow data.

Irregular income is not the same as unstable income. A gig worker who earns steady monthly totals through many small deposits looks very different from someone whose deposits are shrinking. A well-built model tells the two apart.

What does a lender need in place for the data to change decisions?

There is one requirement for each break point:

• A permission model with documented consent, defined permitted uses and a record of both.
• Reliable categorization and stable features that describe behavior, such as income regularity, balance buffers and obligation timing, rather than raw spending totals.
• A model calibrated to the lender’s population and product, and validated on the lender’s own outcomes wherever they exist.
• Outputs that explain themselves, with score drivers, SHAP contributions and plain-English explanations an analyst can read and map to policy.
• Delivery into the existing stack by API or batch, so the output arrives where decisions are made.

This is also where many internal builds slow down. Each requirement is a specialty of its own, and the work does not end at launch. Categorization needs maintenance, models drift and policies change.

How can a lender start without replacing its stack?

The least disruptive starting point is a second look. The lender keeps its current policy, bureau scores and decision engine, and adds a cash flow assessment for applicants who would otherwise be declined or who sit just below a cutoff. We describe the approach in why second-look underwriting is a practical starting point for cash flow underwriting.

1. Define the population. For example, applicants declined for thin or insufficient files, or those in a score band just under the cutoff.
2. Collect permissioned account data from those applicants and score them alongside the existing decision.
3. Run in shadow first. Record what the new signal would have recommended without changing any outcome, as described in how to introduce a new risk signal without breaking governance.
4. Compare against repayment outcomes on your own book, then decide where, if anywhere, to activate the signal within policy.

The lender controls cutoffs, policy and final decisions throughout. The analytics layer provides insight. People and policy still make the call.

Does wider access mean taking on more risk?

Not by default. The aim is to separate risk more precisely, not to relax standards. Extending access is justified only where the evidence of capacity supports it, and the test is the lender’s own repayment performance. Results vary by lender, product and population, so any expected uplift should be measured rather than assumed.

Fair lending needs attention too. Using more data does not make a model fair on its own. Lenders should test for proxy features, review outcomes across groups as their regulators expect, and keep explanations clear enough to stand up to review. Our article on whether cash flow underwriting can create more equitable lending goes deeper on this.

How can lenders add a second look to their current process?

The Borderline Applicant Underwriting use case shows how a lender can take a second look at applicants who sit near its cutoffs or have thin files. For consumer lending, Cashflow Score provides a standardized 1 to 100 score based only on account and transaction data, with score drivers and plain-English explanations. It complements the bureau score and existing rules rather than replacing them.

If you want to see what a second look would find in your own declined and borderline applicants, book a call with the Carrington Labs team.