How do I extract transactions from a PDF bank statement?

A PDF does not contain a table. It contains characters or pixels placed at coordinates, and the table is something the human eye assembles from the spacing. Extraction is the work of reconstructing that structure.

Two kinds of file hide behind the same extension. A statement exported by a bank normally carries a text layer, where every character is stored exactly. A statement produced by a scanner or photographed on a phone carries only images. The first can be read precisely and cheaply. The second requires OCR, which is slower and introduces character level errors.

After the characters are recovered, the parsing problems begin: finding where the transaction table starts and stops, mapping the columns, joining descriptions that wrap over three lines, deciding whether a trailing minus means a debit, and reading dates according to the country that issued the statement.

The final step is the one that separates a usable extraction from a plausible one. Statements carry their own checksum in the form of the balance chain. Running that check tells you whether anything was missed.

A UK statement uses day first dates and separate debit and credit columns. An Indian statement uses lakh grouping in the amounts. A European statement uses a comma as the decimal separator. All three break a parser that assumes one convention, and all three are read correctly by one that reads the convention off the document.

A transaction description reads over three lines: the merchant, then a reference number, then a city. Treated as three rows, the extraction produces two rows with no amount and one that is wrong.

A statement footer repeats a summary of totals. Imported as transactions, those summary lines silently double part of the month.

Extracted transactions can be sorted, categorised, totalled, compared across months and exported into accounting software. None of that is possible while the data lives inside a PDF.

Doing it properly also removes the largest source of error in manual bookkeeping, which is retyping figures, and the balance check gives you evidence that nothing was dropped.

Skipping the balance check. Without it, a missing page looks exactly like a quiet month.

Processing only the first page, or the first few, on a long statement.

Assuming one date format across statements from different countries or different banks.

Importing summary and carried forward lines as if they were transactions.

Combining overlapping statements without deduplication, which doubles every transaction in the overlap.