Cash-flow underwriting means deciding on a loan by looking at the money that actually moves through a borrower's bank account: what comes in, what goes out, and how much is left. A credit score tells you how someone has handled credit in the past. Bank data tells you whether they can afford a new payment now, and it works even for people who have never borrowed.
Below: where the data comes from, the metrics that do most of the work, a worked two-applicant example, and where bank data misleads.
How it differs from score-based underwriting
A score-based decision starts from the credit bureau file. It measures repayment behaviour well, but says little about income and nothing about someone who has never borrowed.
Cash-flow underwriting starts from transactions. It measures capacity directly: how big and how regular the income is, what's already committed, and how close the account runs to zero.
| Score-based view | Cash-flow view | |
|---|---|---|
| Main question | Has this person repaid credit before? | Can this person afford this payment now? |
| Data source | Credit bureau file | Bank transactions, shared with consent |
| Covers people with no credit history? | Poorly | Yes, if they have a bank account |
| Sees income directly? | No, it's usually self-declared | Yes |
| Typical weakness | Thin files, stale information | Short history, categorisation errors |
Most lenders that use bank data don't throw the score away. They use both: the bureau file for repayment history and fraud signals, the bank data for affordability. Our credit underwriting guide covers the full decision process; this article goes deeper on the cash-flow half.
Where the data comes from
There are two main routes.
Uploaded bank statements. The applicant downloads PDFs and sends them in. The lender's software extracts the transactions. This works almost everywhere, but PDFs can be edited, and extraction from inconsistent layouts makes mistakes.
Consent-based account access. Through open banking, or a data connection where open banking isn't mandated, the applicant logs in at their own bank and agrees to share a fixed window of history, often several months to a year. The data arrives straight from the bank, so it's harder to fake and needs no retyping. Our explainer on open banking walks through the consent and token steps.
Either way, the raw transactions are then categorised: salary, benefits, transfers between the applicant's own accounts, rent, loan repayments, gambling, cash withdrawals and so on. Almost every metric below depends on that categorisation being right.
The core metrics
These show up, in some form, in nearly every cash-flow model.
- Average monthly inflow. Total genuine income divided by the number of months. Transfers from the applicant's own savings account and loan proceeds should be excluded, or income looks bigger than it is.
- Income stability. How much inflow swings from month to month. A simple measure is the coefficient of variation (CV): the standard deviation of monthly inflow divided by the mean. A CV near 0% means the same amount lands every month; 40% or more means income is lumpy.
- Recurring obligations. Payments that repeat and are hard to cut: rent, existing loan and card repayments, insurance, childcare.
- NSF and overdraft events. How often payments bounced for non-sufficient funds, or the account went below zero. Repeated events are one of the clearest signs of strain.
- Minimum and average balance. How much buffer the account keeps. An account that hovers near zero has no room for a surprise bill.
- Debt-service capacity. Usually expressed as a debt-service ratio (DSR): total debt payments, including the proposed one, divided by monthly inflow. Lenders often also look at what's left after all outgoings.
None of these has a universal pass mark. Thresholds depend on the product, the lender's risk appetite and local affordability rules.
Worked example: two applicants, one loan
Both applicants ask for the same loan: $6,000 over 24 months at an illustrative 14% a year. All figures below are illustrative, not drawn from real people or any real lender's policy.
The proposed payment. Using the standard EMI formula, the monthly payment is $288.08, with $913.86 of total interest. You can get the same figure from the Loan EMI Calculator by entering a loan amount of 6000, an interest rate of 14 and a tenure of 24 months.
The six months of bank data
Applicant A is salaried, recently moved to the country, and has about ten months of credit history: a single phone handset instalment of $60 a month. Too thin for a reliable bureau score.
Applicant B is a freelance designer with years of clean credit history and a good score, plus a car loan of $320 a month. Clients pay irregularly.
"Outflows" include everything leaving the account: rent, bills, spending and existing debt payments. "Lowest balance" is the lowest the account fell during the month.
| Month | A inflow | A outflow | A lowest balance | B inflow | B outflow | B lowest balance |
|---|---|---|---|---|---|---|
| 1 | $3,400 | $2,950 | $420 | $6,200 | $3,900 | $2,100 |
| 2 | $3,400 | $3,020 | $310 | $1,800 | $3,700 | −$350 |
| 3 | $3,400 | $2,900 | $380 | $4,900 | $4,000 | $900 |
| 4 | $3,450 | $3,080 | $290 | $2,400 | $3,600 | −$120 |
| 5 | $3,400 | $2,960 | $350 | $7,100 | $4,200 | $3,400 |
| 6 | $3,400 | $2,930 | $400 | $3,000 | $3,800 | $1,500 |
Step 1: average monthly inflow
- A: $20,450 ÷ 6 = $3,408.33
- B: $25,400 ÷ 6 = $4,233.33
On averages alone, B looks like the stronger applicant.
Step 2: income stability (coefficient of variation)
Take each month's difference from the mean, square it, average the squares (this is the variance), then take the square root to get the standard deviation. Here we use the population standard deviation over the six months.
- A: five months are $8.33 below the mean and one is $41.67 above. Variance = (5 × 69.44 + 1,736.11) ÷ 6 = $347.22. Standard deviation = $18.63. CV = 18.63 ÷ 3,408.33 = 0.5%.
- B: the squared differences sum to 23,333,333.33. Variance = 23,333,333.33 ÷ 6 = 3,888,888.89. Standard deviation = $1,972.03. CV = 1,972.03 ÷ 4,233.33 = 46.6%.
A's income is close to perfectly regular. B's lowest month ($1,800) is about a quarter of the best ($7,100).
Step 3: monthly surplus and overdraft events
Surplus is inflow minus outflow.
- A: $450, $380, $500, $370, $440, $470. Average $435.00. Lowest balance never below $290. Zero months overdrawn.
- B: $2,300, −$1,900, $900, −$1,200, $2,900, −$800. Average $366.67. Three months ran a deficit, and the account went below zero in two of them (lowest −$350).
Step 4: debt-service ratio
DSR = (existing debt payments + new payment) ÷ average monthly inflow.
- A: ($60 + $288.08) ÷ $3,408.33 = $348.08 ÷ $3,408.33 = 10.2%
- B: ($320 + $288.08) ÷ $4,233.33 = $608.08 ÷ $4,233.33 = 14.4%
Both look comfortable. But for lumpy income, the average hides the risk. A common adjustment is to stress-test against lean months. Using B's three lowest inflows ($1,800, $2,400, $3,000, averaging $2,400):
- B, stressed: $608.08 ÷ $2,400 = 25.3%
- B, worst single month: $608.08 ÷ $1,800 = 33.8%
A's three lowest months are all $3,400, so A's stressed DSR is still 10.2%.
Step 5: what's left after the new payment
- A: $435.00 − $288.08 = $146.92 a month on average, and positive in all six months (the tightest is $81.92).
- B: $366.67 − $288.08 = $78.59 a month on average, but negative in three of six months (−$2,188.08, −$1,488.08 and −$1,088.08). B would be relying on good months to cover bad ones.
Score-only vs cash-flow: the two verdicts
| Applicant A (salaried, thin file) | Applicant B (freelancer, good score) | |
|---|---|---|
| Credit file | About 10 months, one small instalment | Long, clean history |
| Likely score-only outcome | Decline or refer: not enough history to score | Approve |
| Average monthly inflow | $3,408.33 | $4,233.33 |
| Income CV | 0.5% | 46.6% |
| Months overdrawn (of 6) | 0 | 2 |
| DSR with new loan | 10.2% | 14.4% |
| Stressed DSR (3 lowest months) | 10.2% | 25.3% |
| Average surplus after new payment | $146.92 | $78.59 |
| Likely cash-flow outcome | Approve | Refer: smaller amount, longer term or proof of a cash buffer |
The two views nearly reverse each other. The score sees B's past discipline and nothing of A. The bank data shows A can carry the payment every month, while B can carry it on average but not in a lean month without dipping into savings or the overdraft.
That doesn't make B a bad borrower. It means the decision should price in the volatility: a smaller amount, a longer term, a check of B's savings, or a longer history to see whether lean months are seasonal.
Who benefits most
- Thin-file and new-to-country borrowers, who often have steady income but no bureau history.
- Self-employed and gig workers. Income is real but irregular, and payslips don't exist. Transaction history is often the best evidence available, provided the lender measures volatility rather than just the average.
- Small businesses, whose account inflows, card settlements and supplier payments show trading patterns a director's personal score can't.
The limits
- Short windows mislead. Three or six months can miss seasonal income, an annual bonus, or a once-a-year bill. Longer history helps, but asks more of the applicant.
- Categorisation errors. A transfer from your own savings counted as salary inflates income. A loan repayment labelled as "shopping" hides debt. Every metric inherits these mistakes, so good systems let a human review and correct categories.
- Multiple accounts. If income lands in one account and bills leave from another, a single-account view gets both wrong.
- Privacy and consent. Transaction data reveals a lot: health spending, religious donations, who someone sends money to. Collect only the window you need, explain how it's used, and set retention limits.
- Fairness. Some spending categories can act as proxies for protected characteristics. Test outcomes across groups, and avoid decline reasons you couldn't explain to the applicant.
- Rules vary by country. Whether lenders can use account data, how consent works, and what affordability checks are required all differ by jurisdiction and change over time. Check your financial regulator's guidance.
What to look for in bank-data underwriting software
There's no single best provider; it depends on your borrowers, countries and existing systems. These questions separate tools in practice:
- Coverage. Does it connect to the banks your applicants actually use, and handle uploaded statements where connections fail?
- Categorisation accuracy. Ask to test it on a sample of your own anonymised data, and measure how often income and debt payments are labelled correctly.
- Transparent metrics. Can you see how each figure (inflow, CV, overdraft count, DSR) was calculated, down to the transactions behind it?
- Configurable rules. Can you set your own stress tests and thresholds rather than relying on a black-box score?
- Explainable outputs. Can it produce reason codes you could give a declined applicant?
- Consent and data handling. How is consent captured and logged, how long is data kept, and where is it stored?
Next step
Before you assess any applicant's bank data, work out the payment you're testing against. Enter the amount, rate and term into the Loan EMI Calculator, then compare the payment with the applicant's average and lean-month surplus as shown above. For the wider decision process, read our credit underwriting guide, and for how the data gets shared in the first place, see what is open banking.
This article is general information, not financial advice.
