When the statement is a photograph
Two files can look identical on screen and be entirely different documents. One was generated by a bank and carries its characters inside it — every amount is already a number. The other is a picture of a page, and every amount has to be recognised before it can be read. Underwriting treats them as the same thing, and almost everything that goes wrong in extraction goes wrong on the second kind.
- A text layer is read; an image is guessed at
- The statement carries its own answer key
- Only material misses are allowed to call a read unreliable
Read versus recognised
A statement downloaded from a banking portal contains text. The amount in row fourteen is stored as those digits, in that order, and extracting it is a matter of finding the right row rather than deciphering anything. Mistakes are still possible — a column boundary in the wrong place, a wrapped description split across two lines — but they are structural mistakes about layout, not about what the characters are.
A scan, a print re-scanned by a broker, or a photograph taken on a phone contains none of that. It contains pixels that look like digits, and recognition turns them back into characters with a probability attached. A 5 becomes a 6. A decimal point in a poor scan disappears and $1,234.56 becomes $123,456. A faint column rule shifts and a withdrawal lands in the deposit column, which is the worst of the three because the row still looks perfectly reasonable.
Both kinds of file are accepted, and they deserve different amounts of suspicion.
The statement brought its own answer key
The instinct is to solve this with better recognition, and better recognition is worth having. It is not the defence, because it cannot tell you when it has failed — a confident misreading looks exactly like a confident correct reading.
What saves the read is that a bank statement is a redundant document. It prints its own totals for deposits and withdrawals. It usually prints how many credits, debits and cheques it contains. It prints a balance at each end of the period, and often a running balance on every row. Those figures were not put there to verify an extraction, and they do exactly that: a misread digit changes a total, a missed row changes a count, and a wrong amount breaks the chain from the balance above it to the balance below.
So a scan is not trusted because the recognition was good. It is trusted because the numbers it produced add up to the numbers the bank printed on the same page — and when they do not, the ratios are withheld rather than published over rows known to be wrong.
What is checked, and what is allowed to be alarming
Every check is surfaced. Only some of them are permitted to mark a read as unreliable, and the distinction is the point.
A parse that barely produced anything
A month of trading that yields a handful of rows did not go well, whatever the rows say. This is the check that catches a file that was never really read at all — an image-only page, a password-protected export, a document that is not a statement.
Totals missed by a material amount
The extracted deposits and withdrawals are compared with the printed ones, and a gap counts as material when it is big enough to move the read rather than merely to fail a tight equality. A sub-percent difference is noted. A missing page is a finding.
A balance chain that stops following
Where each row carries a printed balance, each one has to follow from the one above plus that row’s amount. A break says a row is missing, duplicated, or misread — and it says where.
Signs and dates that cannot be right
Amounts on the wrong side of zero, and rows dated outside the period the statement declares. Both usually trace to how the document was written rather than to what it says: an accounting negative in brackets or with a trailing minus, or a numeric date whose order was settled wrongly.
Why a small gap is not an alarm
It would be easy to fail any statement whose extracted totals do not match to the cent, and it would be the wrong trade. Statements round, they print a fee summary in one place and the fees in another, and perfectly good extractions land a few pence away from perfect.
A report that raises an alarm over that teaches its readers that alarms are noise, and the cost of that lesson is paid later, on the file where the gap was a missing page. So informational checks stay informational, and only the ones that change what a number means are allowed to say a read is unreliable.
How to give a file the best chance
Prefer the bank’s own PDF. Downloaded from the portal, it carries its text, and every recognition error above simply does not exist. It is also the version whose provenance is easiest to stand behind.
If it has to be a scan, keep the summary page. The totals block is the answer key. A scan without it can still be read, but nothing is left to check the reading against, and the report will say so rather than pretend.
Never photograph a screen. A phone picture of a monitor adds moiré, glare and perspective to a problem that already has enough of all three. If the file exists digitally, send the file.
Common questions
Can scanned bank statements be analyzed?
Yes. A scan has to be recognised rather than read, so it carries more risk of a misread digit or a misplaced column, and it is checked harder: the extracted figures are compared with the totals, counts and balances the statement itself prints.
What kinds of error does recognition actually make?
Digits confused for one another, a lost decimal point, and column boundaries drifting so an amount lands on the wrong side of the ledger. The last is the most dangerous, because the resulting row still looks entirely plausible.
How do you know a scanned statement was read correctly?
Because a bank statement is redundant by design. Its printed deposit and withdrawal totals, its row counts and its balances all have to agree with the extracted transactions, and a misread figure breaks at least one of those agreements.
What counts as a material mismatch?
A gap between extracted and printed totals large enough to change the read — the signature of a missing page — as opposed to the small differences that fail a tight equality without moving any number that matters.
Should I send a photo of a printed statement?
Only as a last resort. Where a digital PDF from the banking portal exists, it carries its own text and removes an entire class of error, and it is also the easiest version to stand behind when a funder asks where the file came from.
Related
Send the file you have
Upload a portal PDF or a scan and see what the statement can and cannot vouch for.