Bank Statement Reconciliation

When a bank gives you both a CSV and a PDF for the same statement period, never trust either in isolation. Verify they agree on three things before using them downstream.

When to use

• New monthly bank statement files land in data/imports/<bank>/
• About to import bank data into Zoho / Xero / a local ledger
• Diffing annual vs monthly exports for the same bank
• Wio Business specifically — see references/wio-business.md

The 3-point verification

For each (CSV, PDF) pair:

1. Row count match. len(csv_rows) == count(date-prefixed lines in PDF). Mismatch ≠ automatic fail (PDF text-extraction can double-count wrapped descriptions), but flag it for manual eyeball.
2. Closing balance match. Extract from PDF header (regex: CLOSING BALANCE.*?([\d,]+\.\d{2})\s*AED). Must equal the last Balance column value in the CSV to the cent.
3. Chain integrity. This month's opening balance = previous month's closing balance. Computed: csv_rows[0].Balance - csv_rows[0].Amount. Off-by-cent = real gap.

If 2 and 3 pass but 1 doesn't, the data is usable. If 2 or 3 fails, stop and investigate.

Standard recipe (Python)

import csv, re, subprocess

with open(csv_path) as f:
    rows = list(csv.reader(f))
header, data = rows[0], rows[1:]
amt = header.index("Amount"); bal = header.index("Balance"); dt = header.index("Date")
opening = float(data[0][bal]) - float(data[0][amt])
closing = float(data[-1][bal])
inflow  = sum(float(r[amt]) for r in data if float(r[amt])>0)
outflow = sum(float(r[amt]) for r in data if float(r[amt])<0)

txt = subprocess.check_output(["pdftotext","-layout","-enc","UTF-8",pdf_path,"-"]).decode()
txn_lines = [ln for ln in txt.split("\n") if re.match(r"^\s*\d{2}/\d{2}/\d{4}", ln)]
pdf_close = re.search(r"CLOSING BALANCE.*?([\d,]+\.\d{2})\s*AED", txt, re.S).group(1)

assert pdf_close == f"{closing:,.2f}", f"CLOSING MISMATCH: PDF={pdf_close} CSV={closing:,.2f}"

Pitfalls

• Annual / lumped exports lie. Many banks' "full year" CSV silently drops rows that exist in the monthly exports. Always prefer monthly. If you only have a lumped file, treat it as untrusted until proven otherwise.
• PDF date-regex false positives. Multi-line transaction descriptions can produce 2nd/3rd lines that start with a date-shaped string. Off-by-1-or-2 row counts are usually this, not real gaps.
• CSV header names vary. Wio uses Balance for the running balance column; some banks use Running Balance or Closing Balance. Don't hardcode — use header.index(...) with a fallback.
• File-naming from bank apps is ugly. Mobile-app exports often land as statement(Mon DD, YYYY - Mon DD, YYYY.csv with a typo'd year (Wio's puts "2025" on a 2024 export). Rename to YYYY-MM.csv / YYYY-MM.pdf immediately, before analysis.

Bank-specific notes

• Wio Business — see references/wio-business.md for filename pattern, account filter, and the documented annual-CSV row-drop bug.