Best OCR for financial documents in 2026

The statements. 47 bank statements from 47 banks, one per bank: French banks such as BNP Paribas, Société Générale, Crédit Agricole and La Banque Postale, German and Austrian banks, and the neobanks a lender actually sees (Qonto, Revolut, Wise, bunq, Shine, PayPal, SumUp). 93 pages in all, in French, German and English. Every statement was anonymized before the test: names, addresses and account numbers replaced, the table layout left as it was.

The scans. Each page was rendered at 150 dpi in greyscale, turned by 0.8 degrees, blurred slightly, given sensor noise and saved as a JPEG at quality 35. That is what a branch copier emails, and it is what OCR exists for.

The answers. 492 transactions, each with its date and amount, checked by hand against the PDF wherever a tested system disagreed with the reference.

The score. For each transaction we look for its amount in what the engine read on that page, digit for digit: 1.234,56, 1 234,56 and 1,234.56 all count, 1234 does not, and each amount found is used once. A date counts when its day and month are there. Every engine ran on CPU, with its default models, and was scored by the same code.

The speed. The same 10 scans, on the same 8 CPU cores, one engine at a time. Hologlyph reads several pages at once; Tesseract reads one page at a time by default, and runs faster if you start several copies.

Hologlyph is the OCR engine Holofin builds for the documents it reads every day: bank statements, financial statements, ID documents and the scans that come with them.

On this test it read every amount the PDF's own text layer holds (the one it missed is missing from the text layer too), and it was the fastest of the five. A fair warning: Hologlyph is trained on financial documents, so the layouts of these 47 statements may be familiar to it.

RapidOCR packages Baidu's PP-OCRv4 models in ONNX, so it installs with pip and runs on CPU with no deep-learning framework. It reads whole lines at once, which works well on numbers. The cost is in the words: on Latin text it often drops the spaces, so RELEVE DE VOTRE COMPTE comes back as RELEVEDEVOTRECOMPTE.

It also glues neighbours together. On one Austrian statement it read the booking date and the amount as one string, 24.12.202315.000,00EUR. A parser that reads that as one number gets more than two hundred million euros.

docTR, from Mindee, reads in two stages: a detector finds the words, a recognizer reads them. It returns clean word boxes with their positions, which makes it the easiest of the five to build a table parser on. It needs PyTorch, and it read one more statement in full than RapidOCR while missing one more amount.

EasyOCR is the friendliest to install and reads more than 80 languages. On scanned statements it read most amounts but mangled the decimals: 7.781,98 became 7.78/,98, 50.000,00 became 50.000,oo/, and 25.000,00 became 25,00o 0o. A zero read as the letter o is a typo in a sentence. In an amount column it is a transaction that no longer parses. It was also the slowest of the five.

Tesseract is the default answer to "free OCR". On these scans it missed 1 amount in 9. Worse than the misses are the misreads: 472,07 came back as 47207 and 22,38 as 2236. Both look like real amounts. A missing number makes a balance fail loudly. A wrong number makes it fail quietly.

Reading is not extracting

An engine that reads every amount has still not told you which amount is the debit, which is the credit and which is the running balance. Statement layouts put dates on one line and amounts on the next, wrap descriptions over three lines and repeat the header on every page. Turning the text into transactions is a separate job, and it is where most extraction errors come from.

A wrong amount looks like a right one

Every misread above is a valid number. 47207 passes every format check a spreadsheet will run. The only thing that catches it is arithmetic: the opening balance plus the transactions must equal the closing balance. If your pipeline does not do that sum, it does not know whether its OCR worked.

OCR reads a forged statement just as well

A perfect OCR reads an edited statement as faithfully as a genuine one. Whether the PDF was modified after the bank produced it, or a page was redrawn by an image model, is a question OCR does not ask.