Trading Agent Data · Guide
What is point-in-time fundamentals data?
Point-in-time fundamentals data is company financial data in which every figure carries the date it was filed. Because of that date, a query for any past day returns only the figures that were on file on that day, and leaves out everything filed or restated later.
Last updated . Written by WU Capital Limited, which publishes Trading Agent Data.
Every figure has two dates
The first is the period end: the last day of the financial year or quarter the figure describes. The second is the filing date: the day the company lodged the accounts with a register or regulator, or the day the register published them. Nobody outside the company could read the figure in an official filing before the second date.
An example from our data. LEGO A/S reported revenue of DKK 83.53 billion for the year ended 31 December 2025. The Danish Business Authority published that annual report on 19 May 2026, 139 days after the year ended. A dataset keyed to the period end shows the figure against 31 December 2025. A point-in-time dataset shows it from 19 May 2026.
Source: Erhvervsstyrelsen (Danish Business Authority), CVR 54562519, read from our API on 19 November 2026. A company may announce results earlier in a press release; the date here is the register's publication of the annual report.
How long is the gap between year end and filing?
For private companies it is measured in months. These are the median days from the financial year end to the filing, from our Filing Observatory of November 2026:
| Country | Financial year | Median days to filing | Legal deadline |
|---|---|---|---|
| Denmark | 2025 | 156 | 6 months (5 for years ending before 31 Dec 2021) |
| Estonia | 2025 | 176 | 6 months |
| Belgium | 2025 | 204 | 7 months |
So for about half of these companies, a figure dated to the year end was not on the register for five to seven months afterwards. Listed companies are faster: Apple's annual report for the year ended 26 September 2026 was filed with the SEC on 3 November 2026, 38 days later.
Why it matters: look-ahead bias and restatements
Look-ahead bias is the error of testing a rule or a model with information that was not available on the date being tested. A credit model trained on year-end figures as if they were known at the year end, or a backtest that trades on them, has seen the future by several months. The result looks better than the model is.
Restatements cause a second, quieter error. A company's revenue for a year can have two values: the one first filed, and the one after a later correction or reclassification. A database that stores one value per company and year keeps the later one. For a question about what had been filed by a given day, the first value is the right one.
The same applies to an AI agent. Asked what was known about a supplier on the day a contract was signed, it can only answer correctly from figures that carry their filing date.
Period-end data and point-in-time data compared
| Question | Keyed to the period end | Point-in-time (keyed to the filing date) |
|---|---|---|
| When does a company's annual revenue appear? | On the last day of the financial year, before anyone had filed it. | On the day the filing was made or published by the register. |
| What happens when a figure is restated? | The stored value is usually replaced, so the original is gone. | The figure as first filed stays, with its own date. |
| What does a query for a past day return? | Today's version of every period that had ended by then. | Only the filings dated on or before that day. |
| What is it suited to? | A current profile of a company. | Backtests, model training, event studies and reviews of past decisions. |
How a point-in-time query works
You give a day, and the answer contains only filings dated on or before it. In our API the parameter is as_of. This call runs on the public demo key and asks for Apple's latest revenue as it stood on 2 November 2026:
curl -H "Authorization: Bearer demo" \
"https://www.tradingagentapp.com/api/v1/fundamentals?ticker=AAPL&metric=revenue&view=latest&as_of=2026-11-02"The answer is the quarter ended 27 June 2026: US$109.417 billion, from the 10-Q filed on 31 July 2026. Apple's financial year had ended five weeks earlier, but its annual report was not yet filed. Change the day to 2026-11-03 and the answer becomes the full year: US$478.834 billion, from the 10-K filed that day.
Source: SEC EDGAR, read from our API on 19 November 2026. Every row in a response carries end (the period end), filed (the filing date) and the register it came from.
The filing date is not recorded the same way everywhere
Registers date filings differently. Some record the day a filing was submitted and some the day it was published. A few publish no day-level date, and there a stated proxy is used in its place. A point-in-time dataset is only as good as its account of these dates, so ours are listed market by market in the methodology, with the known issues.
Questions
What is look-ahead bias?
Look-ahead bias is the error of testing a rule or a model with information that was not available on the date being tested. With company accounts it usually comes from treating a figure as known on the last day of the financial year, when it was filed weeks or months later.
Is point-in-time data the same as as-reported data?
No. As-reported data keeps the figures as the company first filed them, without later restatements. Point-in-time data also keeps the date each figure was filed, so it can be queried by date. A dataset can be as-reported without being point-in-time if the filing dates are missing.
Does point-in-time matter for private companies?
Yes, and usually more than for listed ones. Listed companies usually file within weeks of the period end. Private companies file with their national register months later, and the legal deadline differs by country, so the gap between the period end and the filing date is longer and less even.
Which date counts as the filing date?
The date the official register or regulator gives the filing. Registers record this differently: some give the day a filing was submitted, some the day it was published, and a few give no day-level date, in which case a stated proxy is used. Our methodology lists the date used for each market.
Can an AI agent use point-in-time data?
Yes. An agent that is asked what was known about a company on a past day needs figures that carry their filing date and source, so it can answer from what was on file and cite it. Our API and MCP server return the filing date and the register with every figure.
Where to go next
- API reference: the
as_ofparameter, Python package and MCP server. - Methodology: what the filing date means in each of the 13 markets.
- Filing Observatory: monthly figures on when companies file.
- Filing-date test: what AI models do when the filing date is missing.
- Coverage check: see which companies on your list we hold.
- AI agents and platforms: using the data from an agent.