How it works

From a bank statement PDF to an underwriting picture

Six stages run on every file: read the statement, separate trade from financing, find the recurring debits, measure the burden, check the balance behaviour, check the document. Every figure stays attached to the transactions it came from.

  • Text-based or scanned PDFs from US and UK banks
  • Positions found, typed, and counted per period
  • Burden measured on true trade revenue
  • Every figure tied to its source rows
Live analysis
MetrikData dashboard with a merchant case ready for a bank statement upload

A real statement, merchant anonymised. February 2026, 74 transactions, four positions.

The six stages

What happens between the upload and the report

  1. Upload the statements

    Drop in the PDFs. Text-based or scanned, one month or six, from any major US or UK bank. Nothing is retyped and nothing is keyed in by hand.

    Every transaction, date, amount and running balance is read out of the file. Where a statement does not reconcile, the report says so instead of quietly guessing.

    Out of this stage: 74 transactions, daily balances for every day in the period, and the opening and closing figures reconciled against the statement header.
  2. Separate trade from financing

    A deposit is not automatically revenue. Advance proceeds landing mid-month are financing. Money the owner moves in from their own savings is not a sale. Refunds and returned items are reversals, not income.

    Each of those comes out before anything is divided, because the denominator decides the answer. On this file, $144,776 of deposits became $120,355 of true revenue once $1,041 of advance proceeds and $23,380 of own-account transfers were removed.

    Out of this stage: True revenue, the non-revenue credits that were removed, and payer concentration across the period.
  3. Find the recurring debits

    Funders are identified first by who they are. A known payee on a debit is the strongest single signal, and the registry behind it is kept current separately from the software.

    Payee alone is not enough, because the same lender can hold a monthly term loan on one merchant and a daily advance on another. That is where the rhythm of the dates decides: regular intervals at a stable amount read as a position, and the interval tells you which kind.

    Out of this stage: 4 positions, each with cadence, payment count, average payment and the dates behind them.
  4. Measure the burden

    Daily and weekly advances count toward stacking. Monthly fixed payments are treated as debt service and kept out of the stacking count, because they behave differently against cash flow.

    Every ratio is taken against true revenue, not gross deposits. The difference is shown below on this file.

    Out of this stage: Burden and debt-to-revenue, daily cash left after obligations, and the room under a payment ceiling you set.
  5. Check the balance behaviour

    Closing balance for every day in the period, days under your threshold, days below zero, and the weekday the debits cluster on.

    On this file the account never went negative, yet it closed at $223 on its lowest day and Wednesday carried 67% of the weekly debit load. Both of those matter more than the monthly average.

    Out of this stage: Daily balances, low and negative day counts, NSF and returned item counts, and the weekday pattern.
  6. Check the document itself

    Six deterministic checks run on the file: metadata, balance chain, timestamps, fonts, round-amount deposits and duplicate rows. The same PDF returns the same result every time it is run.

    These read the document, not the bank. They tell you whether the file is internally consistent. To confirm the account behind it, match the statement against a bank-verified source such as a direct bank connection.

    Out of this stage: 6 checks with what each one looked at, and the source PDF kept alongside the report.
Why the denominator matters

The same payments, two different answers

Both numbers below are correct arithmetic on the same file. Only one of them describes the business.

Against gross deposits

19.2%

$27,754 divided by $144,776, including the advance that funded mid-month and the owner’s own transfers in.

Against true revenue

23.1%

$27,754 divided by $120,355, after financing inflows, own-account transfers and reversals are removed.

What comes back

Five pages, every number traceable

One report: Summary, Revenue, Balance, Statement and Transactions, all read from the same file.

Summary

Burden, debt-to-revenue, net cash flow, lowest close, negative days and NSF count, with the monthly roll-up underneath.

Positions

On the summary page: each position with cadence, payment count and average, the order they were taken on, and scheduled payments against what cleared.

Revenue

Deposits against true revenue, what was removed and why, payer concentration, and the room under a payment ceiling you set.

Balance

A closing balance for every day, low and negative days, and the weekday the debits cluster on.

Statement

Six deterministic checks on the document, with the source PDF kept on the same page.

Transactions

Every row, tagged. The figures on the other pages come from these rows and stay tied to them.

Where it stops

We surface the picture. The funding decision is yours

There is no approve, no decline, and no grade standing in for your read. Nothing on the report is labelled good or bad, because that call depends on a credit box we do not know.

There are limits worth stating plainly. A bank statement shows money that moved through the account, so an obligation repaid before funds ever arrive is not visible as a position. The document checks read the file, not the bank. And a single debit from an unfamiliar payee is held for review rather than counted, because one payment is not a pattern.

Everything the report does assert is attached to the transactions that produced it, so you can disagree with any of it on the evidence.

FAQ

Common questions

How long does one statement take?

Minutes, not an afternoon. A single month is usually done before you have finished reading the submission email. Longer or scanned files take more.

Does it work on scanned statements?

Yes. Scanned pages are read as well as text-based ones. They cost a little more in credits because they take more work to read.

What if a payee is one you have never seen?

It still gets caught if it behaves like a position. Payee identity is the first signal, but a debit that repeats on a fixed interval at a stable amount is surfaced whether or not the name is familiar.

What about a payee that only debited once?

It is held for review rather than folded into burden. One payment is not yet a pattern, and guessing at it would move a ratio you rely on.

Can it see every obligation a merchant has?

No, and nothing reading a bank statement can. Repayment taken as a percentage of card settlement never reaches the account as a debit, so it does not appear as a position. Where that pattern is present it is surfaced as a note on the summary rather than counted.

Does it decide the deal?

No. There is no approve, no decline, and no grade. You get the positions, the burden, the balance behaviour, the document checks, and the rows behind all of it.

What does it cost to start?

Free credits on sign-up and no card required. If you would rather not create an account yet, send us a deal you have already funded and compare the read against what you concluded.

Run it on a file you already know

Send us a deal you have already funded. You know what is in it, so you can check the read against your own conclusion before anything else is discussed.

About the figures on this page. They come from one real Bank of America statement for February 2026, merchant anonymised: 74 transactions, $144,776 in deposits, $161,316 in debits, $120,355 of true revenue, four positions carrying $27,754 of debt service.