Indonesian bank statement formats differ by bank, sometimes by product, sometimes by branch. Matching that starts before normalisation will blame amounts that only moved column. Land date, signed amount, reference, and narrative first. Then match.
What usually shifts
Value date versus transaction date. Separate credit columns versus one signed amount. Descriptions that truncate a VA number. A running balance inserted among movements. Excel with a three-row header. A PDF that looks like a table and is actually an image.
Rural and regional banks that ingest many correspondent statements feel this daily. OJK Regulation 34/2025 on rural-bank IT takes effect on 18 December 2026 and requires a core application plus backups. A statement parser that lives only in a staff macro will not survive a restore question.
A written column contract, not a memory
For each source, write: bank, product, filename, header rows, date column, amount column, credit sign, encoding, timezone. Version that contract. When a bank changes the export, you know what broke. Do not rely on “this month it is column C.”
Per-order virtual accounts, often settling D+0 to D+1, only help if the VA is not chopped in the narrative. If the statement keeps 20 characters, store raw narrative in bronze and parse in silver. Do not overwrite raw.
PDF, CSV, and MT940
PDFs need extraction. Extraction quality sets match quality. Test on pages with stamps and footers. Internet-banking CSV often uses semicolon separators and comma decimals. MT940 and CAMT.053 from foreign correspondents add another layer. PBI 10/2025, in force 31 March 2026, does not standardise bank files. It standardises PSP duties. Statements stay wild.
QRIS volume in the first half of 2026, 12.55 billion transactions per Kompas on 18 September 2026, does not arrive as one format. It arrives as acquirer settlement, bank movement, and merchant report, and those three will not agree on description.
Reconciliation and month-end close notes for Indonesian finance teams.
Every Monday, five minutes.
Calendars and cutoffs
National holidays delay QRIS settlement. Idul Adha 2026 merchant notices shifted payouts by several days. A parser that treats every file as D+0 will flag false breaks. Put the bank operating calendar into the date rule.
OJK’s IASC, as reported by Kompas on 8 September 2026, blocked hundreds of thousands of accounts. Missing or held lines must show up as absence, not as a parse failure.
Do not tidy narrative until it disappears. Keep the original. Tidy a copy for matching. That is bronze versus silver in simpler words.
Frequently asked questions
Why do Indonesian bank statement formats always differ?
Each bank builds its own export. There is no mandatory CSV column standard for retail statements.
What should we normalise first?
Date, signed amount, reference, raw narrative. Running balance later as a check, not as the match key.
Can we use PDFs?
Yes after tested extraction. Do not assume one sample page represents every month.
Should we change banks for a cleaner file?
Rarely worth it. Normalisation is cheaper than moving operations.
What about many branches?
One contract per source, not one macro for all. “Almost the same” branches are a bug source.
What evidence do auditors want?
The raw file, parser version, and normalised output that can be recomputed from raw.
Test Threadion on one month of your own bank data
Threadion
