The bank statement PDF to CSV converter reads a statement PDF in the browser, writes a CSV for Xero, QuickBooks or a spreadsheet, and proves the result against the statement itself: opening balance plus every row must equal the closing balance, to the cent. Paid converters charge per page for this. Ours is free and nothing is uploaded.
Until today every test it had passed used PDFs its own test code generated. That is a weak test: the same person wrote both the reader and the PDFs, so they share the same blind spots.
An outside benchmark
An independent, MIT-licensed bank statement parsing benchmark publishes synthetic statements from fictional banks, built around named parsing challenges. Five were public today: a US credit card, Singapore, the Netherlands (in Dutch), Hong Kong (English and Chinese) and Canada (in French). The ground truth is kept on the benchmark's server, but each file states how many transactions it holds, and every statement prints its own balances. That is enough to score a reader that proves its output.
Before: 1 of 5. The Singapore statement converted and reconciled. Three produced no rows at all. Hong Kong produced rows, then reported a broken balance halfway through.
What broke
None of the faults was specific to one bank. Each was a general gap:
- A table further down relabelled every row. The reader took its column layout from the last heading it saw. The card statement has an interest charges box below the transactions, with a "Balance" column right where the amounts were, so every amount was read as a balance and dropped. Each row now uses the heading above it.
- The euro sign. Every Dutch amount ended in byte 0x80. In the encoding most PDFs use for Western text (WinAnsi, which is Windows-1252) that byte is €. The reader decoded it as Latin-1, where it is an invisible control character, so "114,83 €" was not a number.
- French amounts. "10 598,70 $" uses a space as the thousands separator and puts the currency after the number. Headings like "Chèques et débits" and balance labels like "Solde d'ouverture" carry accents the reader never folded.
- Two accounts in one PDF. The Hong Kong statement lists a HKD current account and then a USD savings account. The USD balance starting at 150,000 looked like a missing row worth about 131,600. Now a line that names an account, followed by a balance that starts over, begins a new account. Each account is proved on its own, by an unbroken running balance and its printed deposit and withdrawal totals. You can download one account for Xero or QuickBooks, or all of them with an Account column.
- Text that is not a row. A footer printed a point or two below a transaction was merged into it. A watermark, "Totals Year-to-Date" and a download date were appended to descriptions.
After: 5 of 5, each with the right number of rows and proved by its own balances. The card statement is now recognised as a card without being told, from the phrases only card statements print ("minimum payment due", "payment due date", "credit limit"). Not one of the 99 descriptions carries a footer or watermark.
The part that did not carry over
Five files that were used to find the faults are not a test of the fixes. So before calling it done, the reader ran on two more statement PDFs from other open-source parsers, each with its own ground-truth CSV.
It failed both, and the version from this morning failed them the same way. They use two layouts the benchmark does not have:
- the date printed only on the first row of each day, with later rows blank;
- amounts printed with one decimal ("1935.3").
With those two added, the first file matches all 100 rows on date and signed amount (80 before). The second, a test file built to be awkward, matches 56 of 62 (15 before). Its six misses come from a balance printed without decimals and a description printed above its amount. The tool does not hide them: the balance check fails and names the line where it first breaks.
What this does not show
None of these statements came from a real customer of a real bank. They are synthetic or test fixtures, and the last two files are now development data too. A real statement is still the next measurement, and the balance check exists for exactly that reason: if a row is misread, the totals stop adding up and the page says by how much.
The method, file hashes and every number are in the converter's eval README.