Bank statement parser

A statement from an unlisted bank usually parses fine, because the four named templates share the same underlying line parser and the generic fallback is what actually does the work for everything else. The pattern it looks for is a line that begins with a date, followed by a description, followed by one or two amounts — the second being the running balance. That layout is close to universal in retail banking, whichever institution printed the statement. Leave the chip on Auto: it falls back to Generic when no signature matches, and the preview footer tells you whether the template was auto-detected, manual, or a fallback. Where it does break down is a statement laid out as widely separated columns, or one with debit and credit in separate columns.

Ambiguous dates are resolved from evidence across the whole document, and only when there is none does the tool guess. Any date in the statement with a day above 12 proves the order for every other date, and that inference overrides the bank template when the two disagree. When a statement contains nothing but ambiguous days, the warning box says so and tells you which way they were read. Fix it with the Date format chips: pick DD/MM/YYYY or MM/DD/YYYY and the whole table re-parses instantly, so you see the corrected dates before exporting. The CSV always writes ISO YYYY-MM-DD regardless of what the statement printed, which is what makes the column sort correctly in a spreadsheet.

Rows printed as 05/31 with no year take their year from the statement itself. The parser collects every full date anywhere in the document, discards outliers more than a year away from the most common one — a card expiry or a "customer since" line must not stretch the period — and uses the end of that span. If the document contains no full date at all, it falls back to the current year and warns you loudly that it assumed one. The Statement year field overrides all of that: type a four-digit year and every yearless row uses it. Always read the amber box before importing into accounting software, because a wrong year on a December or January statement is painful to unpick later.

Sign is taken from how the amount is printed, and the common conventions are all recognised. Parentheses around a figure, a leading or trailing minus, and a DR or DB suffix all produce a negative amount, while a CR suffix produces a positive one. Currency symbols and codes attached to the figure are stripped before parsing, and both decimal conventions are handled — whichever of the comma or the period comes last is treated as the decimal separator, so 1.234,56 and 1,234.56 both read correctly. What the tool cannot infer is a statement with separate debit and credit columns and no sign at all: those amounts come through positive, so you will need to negate the debits after import. The preview colours negatives red, which makes this obvious.

A line only becomes a transaction when it has a recognisable date, a description and at least one amount. Descriptions that wrap onto following lines are stitched back together, but only across a few continuation lines, so a very long merchant description broken over many lines can be abandoned. Summary and page furniture are skipped on purpose: totals, opening and closing balances, brought and carried forward lines, statement periods, account and sort code lines, page numbers and column headers are all filtered out so they cannot masquerade as transactions. The warning box counts any dated row that had no amount beside it. If whole sections are missing, try a different template — and remember a scanned statement yields nothing at all.

Your statement is never uploaded, which is the entire reason this tool lives in a browser rather than on a server. The PDF is opened by pdf.js inside the page, the text is read and parsed there, and papaparse assembles the CSV locally before the download fires — no account number, balance, merchant name or filename is transmitted, and there is no endpoint that could receive one. Nothing is stored between sessions either: press Reset or close the tab and the parsed transactions are gone. The tool is free with no signup, and the CSV contains only the four columns and your own rows. Scanned statements are the one case it cannot handle, since there is no text to read — run OCR PDF first.