The SEC's own API will hand you a number 1,000 times too large

Ask the SEC how much money Vince Holding made last year and you can get two answers, both from the SEC, that differ by a factor of a thousand.

Form 10-K                       $6,378,000
XBRL frames API             $6,378,000,000
──────────────
1,000×

We hit this while building a screen. It is worth writing down, because nothing about the wrong answer looks wrong.

What actually happened

The frames API answers a question of the form “give me this tag, in this unit, for this period, across every filer.” It returns one value per company.

To do that it has to choose. A company may report the same figure in several documents, and for Vince Holding’s fiscal 2025 there were two: the annual report, and the proxy statement sent to shareholders before the annual meeting.

The proxy is tagged in the wrong units. The 10-K is right. Frames returned the proxy.

The same thing happened the year before, in the same direction: $19,047,000 in the 10-K, $19,047,000,000 in the proxy.

This is not the SEC making an error. The endpoint selects by period and tag; it does not promise to prefer a 10-K, and the documentation does not say it will. It relayed a number a company filed. The company filed a bad number in one of its documents.

Why it survives every check you would normally run

Here is the part that matters if you build anything on this data.

Every defensive check most pipelines have passes. The value is only wrong relative to the same company’s other filings, and nothing in the response points there.

The first seven traps we have written up in this project were of a different kind: a tag missing, a period mislabelled, a company switching tags mid-series. Those make a number absent or inconsistent, and absence is loud — the missing-tag case, where 526 filers use an element most screens never query, is one of three ways reported capex understates spending. This one makes a number present and wrong, which is quiet.

What it breaks

Anything with a threshold or a ranking.

Our screen asked for companies above $1 billion of net income whose free cash flow had fallen — the screen itself is published here. Vince Holding entered that list. It is a clothing retailer whose operating cash flow for the year was $2,987,000 — a company whose entire cash generation is a third of one percent of the profit the API attributed to it.

That mismatch is what made us look. It is also not a rule you can automate, because most companies with odd cash-to-profit ratios are simply companies with odd cash-to-profit ratios.

The check that works

One extra request per company that survives your screen.

The company concept endpoint returns every filing of a given tag for one company, each labelled with its form. Ask for the same tag and compare.

https://data.sec.gov/api/xbrl/companyconcept/CIK0001579157/us-gaap/NetIncomeLoss.json

The response puts them side by side:

If the frames figure does not match an annual report figure for the same period, drop the company and say so.

Do not require an exact match

This is the part we got wrong first, and it cost us three healthy companies.

Proxy statements routinely round to the nearest million. Our first version required the frames value to equal a 10-K value exactly, and it threw out Thermo Fisher, Chipotle and Corpay — none of which had anything wrong with them.

Chipotle’s proxy figure differs from its 10-K by 0.016%. Vince Holding’s differs by 99,900%. A 1% tolerance sits comfortably between them — 64 times the rounding gap, and nowhere near the real error. We settled there.

The point generalises: you are not hunting for disagreement, you are hunting for a change in magnitude. Write the rule that way and it stays quiet on the cases that do not matter.

What this does not mean

It does not mean the data is unreliable. The SEC’s structured filings are the best public financial dataset that exists, they are free, and the overwhelming majority of what is in them is clean. In the screen that turned this up, one company out of forty-six was affected.

It means the dataset is exactly as good as what was filed into it, and that a pipeline which never looks at a second document has no way of knowing which is which. The safeguard costs one request per surviving company and a few lines of code.

What would make this wrong

• We have not audited how frames chooses. We observed the behaviour on two consecutive years for one company. The SEC does not document a form-type preference, and we are not claiming to know the selection rule.
• The proxy may be tagged correctly for a purpose we do not understand. We treat the 10-K as authoritative because it is audited; that is a judgement, not a rule the API states.
• One company is one company. We are not claiming a rate. What we can say is that a screen of 3,631 filers surfaced this once, and that once was enough to put a wrong name in a published table if we had not checked.
• The 1% tolerance is a choice, calibrated on the four examples we hit. A different dataset may need a different number, and the reasoning matters more than the value.

Check it yourself

Open the company concept link below. Search the response for 6378000 and you will find both values, each with its form field. The frames link returns the larger one for CIK 1579157.

Our implementation of the comparison is a function called verifyAgainstFilings in the screen script in this site’s repository, and the tolerance is one constant at the top of it.

If a figure here does not match a filing, tell us and it will be corrected on the article and on the corrections log, with the date.