How to Extract Transactions from a PDF Bank Statement
A PDF looks like a table but stores positioned text fragments with no rows, columns or reading order. Extracting transactions means rebuilding that structure. Here's how the process works under the hood, and what it means for the accuracy of your data.
Step 1: Get the text out
PDFs downloaded from online banking embed a text layer that can be read precisely, with exact characters and positions. Scanned statements are images — extraction must fall back to OCR, which introduces a recognition step and therefore a margin of error.
Good pipelines detect which kind of document they're holding and only OCR when necessary.
Step 2: Reconstruct the table
With text and positions in hand, the extractor infers column boundaries from alignment, groups fragments into rows, joins wrapped description lines to their parent row, and propagates dates that print once per group. Bank-specific rules make this reliable for known layouts; a generic reconstruction engine handles the rest.
Step 3: Normalize and validate
Raw strings become typed data: dates to ISO 8601, amounts to signed decimals, paired money-in/money-out columns to debit and credit fields, currency detected from symbols and context.
The strongest correctness check is the running balance: each balance should equal the previous balance plus the transaction amount. A chain that reconciles end-to-end is strong evidence that no row was missed or misread.
- ISO dates eliminate locale ambiguity
- Signed amounts allow direct summation
- Balance validation catches missing or corrupted rows
- Confidence scores mark rows needing review