For researchers

Why filing dates matter

A company's 2019 revenue can have two values: the figure first published in 2020, and the figure after a later restatement or reclassification. A database that stores one value per company-year keeps the later figure and drops the first.

For most purposes that is fine. For anything where the question is what had been filed by a given date — an event study, a credit model, a trading rule, a policy evaluation dated to an announcement — it is not, because the later figure was not available on that date. The bias has a direction: restatements are not random, and neither is which firms survive to be restated.

Here, the first published figure stays. A later filing that restates the same period is not written over it, and is not stored at all (see the methodology), and a company that is struck off keeps everything it ever filed. Ask for a company as_of=2019-03-01 and you get the rows whose filings are dated on or before that day. Later restatements are not stored, so nothing has to be filtered out.

If your current source overwrites figures on restatement or drops companies that exit, the original figures cannot be recovered from it.

What is in the fileWhat is in the data

Counts are read from the delivery manifests, not typed into this page, so what is quoted here is what arrives.Counts are read from the data manifests, not typed into this page. Current vintage 2026-11-09.

Delivered as gzipped CSV, one file per market, UTF-8 with a header row, plus a manifest carrying the row count and a SHA-256 for each file. Columns are documented in the data dictionary. Every row states its reporting basis, so a group consolidation is never silently mixed with the parent's own accounts.

What it is not

• This is filed accounts, not a firm-level panel with employment, ownership or trade linkages. There are no director records, no shareholder chains and no group structures.
• Sole traders are excluded. A French entrepreneur individuel is a person, not a company, and republishing a person's finances engages the GDPR and INPI's own reuse licence (art. 3) in ways a company dataset has no reason to. 7,765 French and 1,057 Danish entities are removed — 0.52% and 0.19% of those markets. A study of the very smallest firms should treat them as outside the population. The legal basis and the identification rule are set out on the methodology page.
• New Zealand charities currently cover registered charities only; deregistered ones are not yet included, so that market does not yet have the struck-off retention the European registers do.
• Coverage begins where each register's structured filings begin, not where the companies do: Belgium from April 2019, France from 2017 (about 700 of 6.2 million filings are dated earlier), Denmark from June 2012 (earlier filings are scans or PDFs), Japan from 2016. A company older than its register's structured era has no rows for those years here.
• Turnover is not universal. Small-company reporting exemptions mean many filers publish a balance sheet and no revenue line; it is left empty rather than imputed. Measured fill rates per metric and per market are published on the methodology page — read them before you design around a field.
• Figures are reported as the filer published them, including their own rounding. Some Japanese filers present the five-year summary in millions of yen; that is their presentation, not our truncation.
• Names and identifiers come from the registers. Cross-border entity matching, LEI linkage and parent-subsidiary resolution are not included and would have to be done by you.

Citing it

Cite the vintage, not just the dataset — it is the vintage that makes a result reproducible, and two vintages are not the same file.

WU Capital Limited (2026). Trading Agent Data: company fundamentals from official filings [Data set]. Vintage 2026-11-09. Auckland, New Zealand. https://www.tradingagentapp.com/fundamentals

@misc{tradingagent_pit_fundamentals,
  author       = {{WU Capital Limited}},
  title        = {Trading Agent Data: company fundamentals from official filings},
  year         = {2026},
  note         = {Data set. Vintage 2026-11-09},
  howpublished = {\url{https://www.tradingagentapp.com/fundamentals}}
}

Source registers must be credited in their own words as well; the required strings ship with every API response and bulk delivery, and are listed on the methodology page. For Japan in particular, attribution is a condition of the licence rather than a courtesy.