NSF Events, Revenue Variability, and Other Signals Hiding in Plain Sight
Firms with volatile expense patterns exit at 21% against 3%. A working catalogue of cash-flow risk signals, what the evidence supports, and where it stops.
Alfred BEditorial Reviews
An underwriter reading bank statements is usually looking for one thing: does the income support the payment. That question is answerable from a total. Most of what a transaction history knows is in the shape of the flows rather than the sum of them.
The research is unusually clear here. FinRegLab found that standing alone, cash-flow metrics generally performed as well as traditional credit scores. The JPMorgan Chase Institute found small businesses with volatile expense patterns exited at 21% against 3% for firms with regular weekly patterns. Balance behaviour, payment rhythm and negative-balance frequency carry real predictive weight, and none of them appear on a bureau file.
This is a working catalogue of those signals, what the evidence says about each, and where the evidence runs out.
The four families
FinRegLab's July 2019 research described what participating lenders actually derived from cash-flow data: income-to-expense ratios, differences in flows of fixed and variable income, minimum balances, and the frequency of negative balance events.
Almost everything useful sits in one of those four. They are worth taking one at a time, because they fail differently.
Negative balance events
Overdraft and NSF activity is the most direct behavioural signal available in an account, and the one most likely to be dismissed as noise.
The US Consumer Financial Protection Bureau reported in December 2023 that approximately 26.5% of consumers live in households charged an overdraft or NSF fee within the prior year. Common enough to be informative rather than exceptional.
The same research found the frequent-fee group had lower average credit scores, were the most likely to hold a subprime score, and were most likely to have no available credit on a card and delinquent debt. The CFPB described a high level of financial vulnerability among many consumers incurring these fees, covering both paying bills today and weathering future shocks.
For business accounts, FinRegLab's June 2025 work associated NSF transactions and low or negative ending balances with higher default risk, using a three to six month window.
The property that makes this signal valuable is timing. An NSF appears in an account the day it happens. It appears in a credit file only if something eventually defaults.
A caution worth being specific about. No public source supports a particular threshold. FinRegLab explicitly did not evaluate participants' proprietary models and anonymized findings to protect them, so there is no published ranking of which cash-flow attributes predict best, and no verifiable "three NSFs in ninety days" rule anywhere in the literature. Anyone quoting a specific cut-off is quoting a house rule, not research.
Revenue variability
Variability is where the strongest single finding in this area sits, and it is about businesses.
The JPMorgan Chase Institute, analyzing 1.3 million small businesses and 3.1 billion transactions between October 2012 and February 2018, found that firms with volatile expense patterns exited at 21%, compared with 3% of firms with regular weekly patterns. A sevenfold difference in failure rate, keyed to the rhythm of outflows rather than their size.
The same research found regularity improves with maturity: 69% of firms had regular cash flow patterns in year one, rising to 78% of surviving firms by year four. So irregularity is both a risk marker and something that resolves for the firms that make it.
The underwriting implication is that pattern deserves its own variable. Two businesses with identical revenue and identical margins are not identical risks if one pays in a weekly rhythm and the other in unpredictable bursts.
Balance and buffer
The JPMorgan Chase Institute's September 2016 analysis of 597,000 small businesses and 470 million transactions found the median small business held a cash buffer covering 27 days of typical outflows. The 25th percentile held 13 days, the 75th held 62. Restaurants were lowest at 16 days.
Twenty-seven days is a thin margin, and it reframes what a minimum balance means. For a business at the median, a single delayed receivable is most of the buffer. Minimum balance is not a wealth measure. It is a measure of how many days of disruption the business can absorb before it cannot pay something.
Income composition
The distinction FinRegLab drew between fixed and variable income flows is the one most often collapsed in practice.
Two applicants with identical annual deposits are different risks if one receives 26 near-identical amounts and the other receives irregular sums from three sources. The total is the same. The predictability is not, and predictability is what a payment schedule depends on.
This matters more each year as fewer workers are paid in the shape verification systems expect.
The catalogue, and what each signal is good for:
| Signal | What it indicates | Evidence |
|---|---|---|
| NSF and negative balance frequency | Current financial strain; associated with lower scores and delinquent debt | CFPB, Dec 2023; FinRegLab, Jun 2025 |
| Expense pattern regularity | Business failure risk; 21% vs 3% exit rates | JPMorgan Chase Institute, 2018 |
| Days of cash buffer | Absorbable disruption; median 27 days | JPMorgan Chase Institute, Sep 2016 |
| Fixed vs variable income mix | Predictability of repayment capacity | FinRegLab, Jul 2019 |
| Income-to-expense ratio | Headroom against obligations | FinRegLab, Jul 2019 |
Does any of this beat a credit score?
Roughly matches it, and improves it in combination.
FinRegLab's July 2019 finding was that standing alone, cash-flow metrics generally performed as well as traditional credit scores, suggesting both can provide meaningful predictive power. Its September 2019 small business work found cash-flow variables predictive across a highly heterogeneous set of participants, and frequently improving prediction when combined with traditional scores.
More recently, FinRegLab reported in July 2025 that models combining cash-flow data with machine learning raised approvals by about 4%, roughly two million additional card accounts and 152,000 mortgages annually at 2023 volumes, while reducing approvals among consumers likely to struggle.
The combination is the finding. Not the replacement.
Building this into a decision
- Compute the four families before adding anything exotic. Income-to-expense, fixed versus variable mix, minimum balance, negative-balance frequency. These are the ones with published evidence behind them.
- Add pattern regularity as a distinct variable for business files, given the 21% against 3% exit finding.
- Express buffer in days rather than dollars. A median of 27 days across US small businesses is the context that makes a balance meaningful.
- Treat NSF timing as its own input. Recency and clustering carry information that a count discards.
- Set your own thresholds on your own book, and document that they are yours. There is no published number to borrow.
- Run these alongside the bureau file rather than instead of it.
What the evidence doesn't cover
Every source above is US. No Canadian regulator or research body has published equivalent work on NSF or overdraft as default predictors, and we could not find Canadian buffer or variability benchmarks.
There is also no peer-reviewed work isolating NSF event counts as a standalone predictor, and no public ranking of which cash-flow attributes matter most. That gap is deliberate on FinRegLab's part rather than an oversight, since publishing it would expose participants' models.
So the signals are established. The precise weights are yours to find.
Common questions
Are NSF events predictive of default?
The CFPB found consumers with frequent overdraft or NSF fees had lower credit scores and more delinquent debt, and FinRegLab associated NSF transactions with higher default risk in small business lending.
How much cash buffer does a typical small business hold?
The JPMorgan Chase Institute found a median of 27 days of typical outflows, with the 25th percentile at 13 days.
Does irregular cash flow predict business failure?
The JPMorgan Chase Institute found firms with volatile expense patterns exited at 21%, against 3% for firms with regular weekly patterns.
How many NSF events are too many?
No published research establishes a threshold. FinRegLab withheld attribute-level rankings to protect participants' models, so any specific cut-off is a house rule rather than a research finding.
Do cash-flow signals beat credit scores?
FinRegLab found they generally performed as well standing alone, and frequently improved prediction when combined with traditional scores.
Carousel returns categorized transaction data ready for balance, variability and NSF analysis. See the merchant intake suite
