Cash-Flow Data Won't Replace the Credit Score. It Completes It.
Combined bureau and cash-flow models beat either alone. Cash-flow data on its own doesn't reliably beat the score. What the research actually found.
Alfred BEditorial Reviews
The most common claim about cash-flow underwriting is that transaction data beats the bureau file. It is a good story, it flatters everyone building in the space, and the best available evidence does not support it.
What the research shows is more useful than the headline. Models combining bureau data with cash-flow data outperform either source alone, consistently. Cash-flow data on its own sometimes beats the traditional score and sometimes loses to it, depending on the lender. The reliable gain comes from using both.
That is a less exciting sentence and a much better basis for a lending strategy.
What the evidence actually says
The clearest public dataset on this comes from Charles River Associates, which analyzed cash-flow underwriting for FinRegLab in July 2019 across multiple lenders. Measuring predictive power by area under the curve, the results by participant looked like this.
Predictive power by data source, from the Charles River Associates analysis for FinRegLab (July 2019). Higher AUC means better discrimination between defaulting and non-defaulting borrowers:
| Participant | Traditional score | Cash-flow only | Combined |
|---|---|---|---|
| 2 | .640 | .652 | .660 |
| 4 | .559 | .592 | .620 |
| 5 | .573 | .572 | .659 |
| 6 | .720 | .688 | .758 |
Read the rows individually. For participant 6, the traditional score outperformed cash-flow data alone by a clear margin. For participant 5, the two were nearly identical on their own. In every single case, the combination beat both.
FinRegLab's own July 2019 conclusion put the mechanism plainly: cash-flow metrics frequently improved the ability to predict credit risk among borrowers that traditional systems scored as presenting similar risk of default. That is the actual value. Not replacing the score, but separating people the score cannot tell apart.
More recent work points the same direction. FinRegLab reported in July 2025 that machine learning models combining bureau and cash-flow data increased approvals by roughly 4%, which at 2023 volumes it estimated as about two million additional card accounts and 152,000 additional mortgages annually.
Why the two sources disagree usefully
A credit file is a record of how someone has handled formal credit obligations. A transaction history is a record of money actually moving. They answer different questions, and the disagreement between them is where the information is.
Two applicants can carry the same score, the same utilization and the same file thickness while looking nothing alike in their accounts. One holds a stable balance and pays rent on the first. The other runs to zero every month and has three overdraft events in the last quarter. The bureau cannot see the difference. Transactions can.
That is also why cash-flow data does not simply beat the score. Someone with a thin file and clean transactions may be a good risk the bureau underrates. Someone with volatile income and a long, flawless repayment record may be a good risk the transactions underrate. Neither source is complete.
Where cash-flow data carries the most weight
There is one population where the argument changes shape, because for them the bureau file barely exists.
Statistics Canada reported in September 2023, with a correction issued in February 2024, that about 1.1 million Canadian economic families, or 7.2%, are credit invisible. Among immigrant families in Canada less than two years, that figure rises to 14.8%, against 7.5% for Canadian-born families.
For those households there is no combined model to build, because there is nothing to combine. Transaction data is not an enhancement to their file. It is the only file.
One number worth being careful with
A widely repeated figure holds that cash-flow underwriting increased approvals by 27%. It comes from a 2019 assessment of a lender whose model used education and employment variables alongside cash-flow data, which means the result cannot be attributed to cash-flow data alone.
It circulates as a cash-flow statistic anyway. The Charles River AUC table above is the cleaner number to cite. Less dramatic, and it survives scrutiny.
What this means for how you build
If cash-flow data replaced the score, the right move would be a parallel underwriting stack. If it completes the score, the right move is an intake layer that reliably produces both, and a model that can use them together.
That is a less romantic project and a more tractable one. It also changes what you should be asking of a data provider. The interesting question is not whether their signal beats the bureau. It is whether their signal arrives complete, current and structured enough to sit alongside the bureau in the same decision.
The lenders getting the most out of transaction data are not the ones who replaced anything. They are the ones who stopped treating the two sources as competitors.
Common questions
Is cash-flow data better than a credit score?
Not reliably on its own. In the Charles River Associates analysis for FinRegLab, cash-flow-only models beat traditional scores for some lenders and lost to them for others. Combined models beat both in every case measured.
What does cash-flow underwriting add?
It separates borrowers that traditional scoring treats as equivalent risks, according to FinRegLab's July 2019 findings.
How many Canadians are credit invisible?
About 1.1 million economic families, or 7.2%, according to Statistics Canada, rising to 14.8% among immigrant families in Canada less than two years.
Does cash-flow data increase approvals?
FinRegLab reported in July 2025 that combined bureau and cash-flow machine learning models increased approvals by roughly 4%.
Where does the 27% approval figure come from?
A 2019 assessment of a model that used education and employment variables alongside cash-flow data, so it should not be cited as a cash-flow-only result.
Carousel's intake flows return structured transaction data ready to sit alongside bureau data in the same decision. See how verification fits your flow
