Synthetic ERP data, generated from accounting rules
Synthetic ERP data is enterprise resource planning data - purchase orders, invoices, payments, journal entries, bank statements - that never belonged to a real company. There are two ways to make it, and they are not interchangeable.
Most synthetic data tools start from a real database: they learn its statistical patterns and emit a privacy-safe copy. That works when you have a source system and permission to touch it. It fails when you have neither, and it inherits whatever the source contained.
synthetic cfo takes the other road: a cold start engine that generates complete company worlds from accounting rules alone - IFRS and GAAP treatments, real ERP table architecture, and double-entry logic. No seed dataset exists at any stage, so there is nothing to anonymise, nothing to re-identify, and nothing to clear with a privacy team.
Seven business cycles on three platforms (SAP ECC, SAP S/4HANA, Oracle Cloud): procure-to-pay, order-to-cash, cash and treasury, record-to-report, payroll, inventory and fixed assets. Every foreign key resolves, every debit matches a credit, the trial balance nets to zero, and the financial statements articulate from the operational subledgers with a tie-out workpaper. The world carries two accounting layers and says so: the record-to-report overlay (manual journals, trial balance, close) is not re-consolidated with the subledgers, so the trial balance ties to the general ledger journal layer and never to the payables or receivables documents. Documents carry a real approval lifecycle - release events, invoice holds - and every bank statement line mirrors an operational payment, receipt, payroll run or tax remittance, with the per-bank reconciliation proven to the cent at generation time.
Multi-year fraud arcs run up to five fiscal years of the same company, and a two-ERP intercompany group runs as an arc too, so multi-entity and multi-year come in one archive with an elimination schedule that carries innocent open items beside the planted fraud.
More than 25 fraud schemes are planted into the data - ghost vendors, duplicate invoices, kiting, lapping, channel stuffing, self-approval above authority - and every instance is recorded in a ground-truth answer key at the moment it is planted. Innocent look-alikes of every fraud condition exist beside the real thing, so a detector or a model can be wrong, which is what makes a score mean something. The same seed and configuration regenerate the identical dataset byte for byte, with a certificate of file hashes to prove it.
A complete sample package is free to download with no account, and the free developer plan generates on all three platforms. Programmatic access ships on every plan: an API key plus the Python client (one pip command on the evaluate page) generates, polls and downloads from a few lines of Python, or from one terminal command with no code at all.