Which XBRL tag is free cash flow, and why the obvious one is wrong

The two tags, and the arithmetic

Free cash flow, in the form almost every valuation uses, is cash generated by the business minus what it spent on the assets it runs on:

Both come from the cash flow statement, both are tagged in the us-gaap namespace, and you fetch each one on its own from the company concept endpoint:

That is it for the happy path. The rest of this page is the ways it goes wrong, every one of them checked against live filings rather than remembered.

Two details that are easy to miss before anything else can be right. Filter to annual forms, which means 10-K for domestic filers and 20-F or 40-F for foreign ones, because the same tag also carries quarterly facts. And check the period length: subtract start from end and keep only facts covering roughly a year, since a 52-week retailer year and a leap year are both normal and a stray nine-month stub is not.

Trap one: the tag that answers but stopped years ago

Five of the twenty companies below report capital expenditure under PaymentsToAcquireProductiveAssets. Several of those used PaymentsToAcquirePropertyPlantAndEquipment first and moved. EDGAR keeps serving the old tag, because the old filings are still real filings. If your code takes the first tag that returns HTTP 200, this is what you get:

Amazon is a twentyfold error. Nvidia is worse in proportion and obvious enough that somebody would probably catch it. AT&T is the dangerous one: $20,647m against $20,842m, a difference of one percent, from a filing eight years apart. Nothing about that number looks wrong. It would sit in a model for years.

The same thing happens on the cash flow side. NetCashProvidedByUsedInOperatingActivitiesContinuingOperations still answers for Apple with FY2016, for Microsoft with FY2017, for Intel with FY2018 and for Exxon with FY2016, long after each moved to the shorter tag. Twelve of the twenty companies have at least one abandoned tag that is still happily returning a frozen number.

Trap two: there is no single capex tag to begin with

Here is what twenty large filers actually use, read on 14 October 2026:

Operating cash flow is tidier but not tidy. Eighteen use NetCashProvidedByUsedInOperatingActivities, General Electric uses the continuing operations variant as its live tag, and Coca-Cola returns nothing at all.

Two more capex spellings exist and are worth asking for even though they are rare here: PaymentsForCapitalImprovements, which Walmart last used in FY2012, and PaymentsToAcquireOtherPropertyPlantAndEquipment, which none of these twenty use.

Trap three: when both tags are live and they disagree

General Electric files both operating cash flow tags for the same year, FY2025, with different values:

Six million dollars apart on eight and a half billion, so it changes nothing in a valuation. But it breaks the rule from trap one, because newest-wins cannot choose between two facts with the same date. You have to decide which question you are asking. Continuing operations excludes businesses being sold or wound down, which is usually what you want when you are forecasting forward. The total is what the company actually banked.

For a company in the middle of a large disposal those two numbers are not six million apart. Prefer one explicitly, record which you used, and show it next to the figure.

Trap four: a 200 that contains nothing

Ask the company concept endpoint for Coca-Cola's operating cash flow and it answers normally. Status 200, valid JSON, a units object with a USD key. The array under that key is empty. Asked four times in a row, it was empty every time.

The same tag, for the same company, fetched through the bulk company facts endpoint, holds 51 annual facts, the most recent being FY2025 at $7,408m. Capital expenditure is the same story: nothing from one endpoint, 51 annual facts and $2,112m from the other.

This one cost us a live bug, found while writing this page. Ask the lookup behind our own calculator for Coca-Cola today and it fails outright rather than returning the figures that a different endpoint would have given it. Failing is better than inventing a number, but it is not the right answer when the data exists one endpoint over, and it is being fixed.

Trap five: some companies have no free cash flow to compute

JPMorgan has never used any of the four capital expenditure tags. Not once. And its FY2025 operating cash flow is negative $147,782m, which is not distress, it is what a bank's cash flow statement looks like when lending and deposits move through it.

Subtracting a missing capex from that produces a number, and that number means nothing. Banks, insurers and most financial companies need a different approach entirely. The honest response from a tool is to say so rather than to show a confident figure built on an irrelevant one.

What a correct lookup does, in order

1. Ask every candidate tag for the field, not the first one.
2. Keep only annual forms, and only facts whose period is roughly a year long.
3. Among the tags that returned facts, use the one whose newest fact is newest.
4. When two tags are live for the same period, pick one deliberately and say which.
5. Treat an empty array and a 404 as different outcomes.
6. Anchor every balance sheet figure to the annual period end and drop anything more than about 45 days away, instead of mixing dates.
7. When a field cannot be filled, leave it blank and say so. Never carry over the value from the company you looked at before. That one has bitten us, and the output looked perfectly reasonable, which is what made it dangerous.

All seven rules are in the code behind this site's calculator and its company pages, which is why those pages print the tag and the period next to every figure. If a number is wrong there, you can see immediately whether the arithmetic or the source is at fault.

Check this yourself

Every figure above comes from one script hitting SEC EDGAR, and both the script and its output are published rather than described:

• fcf-tag-survey.json, the full result for all twenty companies, including the tags that returned nothing.
• fcf-tag-survey.py, the script that produced it. Run it and the numbers regenerate. If it disagrees with this page, the script is right and the page is stale.

EDGAR itself is free, needs no key, and asks only for a descriptive User-Agent and no more than ten requests a second.

Read from SEC EDGAR on 14 October 2026 · figures are annual filings only · corrections welcome at [email protected]