How to verify a financial number against its filing

Almost every screener, terminal and finance site you have used is built on the same raw material: XBRL, the machine-readable tagging that public companies attach to their filings. It is what makes it possible to compare five thousand companies at once.

It is also wrong often enough to change a conclusion.

Start with an identifiable claim

Write down the company, fiscal period, accounting concept, and units. Record the exact filing URL and the statement or footnote where the figure appears. A company filing-history page alone cannot establish an individual number.

For a borrowing figure, check whether the question is current debt, long-term debt, total borrowings, or debt including leases. These are different measures. Compare like periods and like units before calculating net debt or leverage.

The original version used a Kraft Heinz historical comparison without enough information to reproduce it. That example has been removed; see the dated correction below. The separate Vince Holding investigation supplies specific SEC endpoints and values for a different, documented discrepancy.

Why this happens

Companies do not all tag the same concept the same way. There are several legitimate tags that can carry borrowings, and a filer may populate one, several, or a company-specific extension tag instead. A consumer of the data that reads only one field gets whatever happened to be there. The same split runs through capital expenditure, where Amazon has not used the standard tag since 2017 and seven of sixteen companies on a recent screen report it under a different one.

The failure mode is quiet, which is what makes it dangerous:

• It does not throw an error. A missing tag returns nothing, and nothing frequently becomes zero somewhere downstream.
• The result can stay plausible. A numeric value can sort, chart, and pass a range check while describing the wrong concept.
• It compounds. Net cash, leverage, coverage and any valuation built on them inherit the error silently.

A low-debt screen can include the wrong company if it mistakes a partial debt measure for total borrowings.

How it gets caught here

Three rules to apply before publishing a material figure:

1. Standardised data locates, it does not confirm. XBRL is used to find where a number should be. The value that gets published is read out of the filing text and reconciled against the tag.

2. Absence is not zero. If a value cannot be confirmed, it is marked unconfirmed and the dependent metrics are not computed. An unconfirmed borrowings figure means no net-cash figure — not a net-cash figure that assumes zero debt.

3. Provide the inputs and method. Link the source, identify its location, and show the calculation. For a screen, disclose the population and exclusions; a changing API endpoint does not preserve a historical dataset.

What this costs

It is slower. A pipeline that trusts the standardised feed can cover the entire market in an afternoon. Reconciling against source documents cannot.

That trade is the reason this site exists. Broad, fast and occasionally wrong by four orders of magnitude is already available everywhere, for free. Narrow and checked is not.

Rule 1 is not hypothetical. In August 2026 it caught a company whose net income the SEC’s own frames endpoint reported as $6.378 billion when its annual report said $6.378 million — a thousandfold error that returns HTTP 200 and passes every schema check.

Nor is rule 2 an abstraction about missing data. Companies themselves withdraw numbers they have already published: in six weeks that summer, 17 filers told investors their own accounts should no longer be relied on, in one case because the auditor said it had never issued the opinion the annual report carried.

What would make this wrong

The reconciliation depends on the footnote being read correctly, and that step can involve both automated extraction and editorial review. It is not immune to error — it is only better than the alternative. When it fails, the correction appears on the article and on the corrections log, with the date.