EDGAR publishes every value a company has ever reported for a figure, each stamped with the filing that reported it. Across 6,267 companies, 397,407 period-figures were reported in more than one filing. 37,472 of them came back with a different number, at 4,175 companies — 66.6% of the population. Nothing here diffs a snapshot: the register ships the revision history with the figure.
A figure reported twice with the same number is an ordinary comparative and means nothing. A figure reported twice with two different numbers is the thing this page is about. Of those, 22,407 are large enough to measure as a percentage — the rest start from a base too near zero for a percentage to say anything. And 72.6% of the measurable ones move the figure by under five per cent — too small to make news, too large to join on.
of measured corrections move the figure by less than a tenth of a percent. Small enough to ignore in a report, large enough to break a join, a reconciliation or a hash.
move it by more than a quarter. Those are the ones that change what the number means, and they are the minority that gets written about.
of repeated total-assets figures change, against 8.6% of repeated net-income figures. The headline number is the steadier of the two.
The interesting question for anyone building on this data is not whether a figure changes but when. If corrections arrived within the quarter, a pipeline could wait them out. They do not.
of corrections land at least a year after the figure was first filed. Only 2.2% arrive inside four months.
days between a figure being filed and the same figure being filed differently — the outer edge of what "late-arriving" means here.
is the period year that gets corrected most often, with 3,083 revised figures. Periods from 1999 onward are still being restated.
There is a filing type for correcting a previous one:
the amendment, marked with an /A. If corrections lived there, a
pipeline could watch for them.
They do not. 8.4% of corrections arrive in
an amended filing. The rest arrive in the next routine
10-K or quarterly report, as a comparative column, with
nothing marking it as changed. You find out by comparing, or you do not find
out.
That is the whole practical lesson of this page. A filing is not a fact; it is a fact as of a date. Storing one without the date of the filing that asserted it throws away the only thing that would let you notice the difference.
Every figure above survives four filters. Each one was added because the unfiltered version produced a striking, quotable and false result — and each rejected row is counted and published rather than quietly dropped.
facts were excluded for not coming from a periodic report. A proxy statement tags the same element with pay-versus-performance figures: one company's 10-K reported −$19.2m for a year and its proxy tagged −$35.0bn for the same year.
revisions are a clean power of ten with the sign unchanged — the same digits filed at the wrong scale. One company filed $22,632,000 for a quarter three times, then $22,632,000,000.
revisions have a value of exactly zero on one side. One bank filed $1.818tn of total assets for 2008 three times and then 0. It did not restate its balance sheet to nothing.
revisions carry a value more than 20× the company's own median for that figure. Not every mis-tag is a clean power of ten, and the company's own history is the only scale reference this data provides.
What is left after those four is still not certified. A revision of ten times or more is almost never a restatement, and 74 of them remain — some flagged, some not, none of them checkable without reading the filings. So this page does not claim to have separated the real from the mis-tagged at the extreme. It publishes a bounded band instead, from 5% to 100%, where the revision is large enough to matter and small enough that a mis-tagged magnitude is implausible — and shows the tail separately, labelled as what it mostly is.
5,437 revisions sit inside that band. They are ordinary corporate history: a spin-off restating what the remaining company ever owned, an accounting standard adopted retrospectively, an acquisition consolidated back through prior periods.
The largest by percentage is SLM Corp, whose
Assets for 2013-12-31 moved
-93.3%. None of these is a scandal. They are what a
correctly functioning register looks like — which is the point: a pipeline
that cannot absorb these is not being defeated by fraud, it is being defeated by
normal accounting.
The values are not the only thing that changes underneath a query. The element a company uses to mean "revenue" changed across the whole market when the revenue standard did, and a query pinned to one element simply stops returning that company.
SalesRevenueNet peaked at 2,309 companies and
is down to 0.
RevenueFromContract did not exist
before 2018; it reached 2,799 companies in its first year and
2,967 by 2024. Meanwhile the generic
Revenues element carries 2,503.
A revenue series built on any single one of these has a hole in it. That is not a data-quality problem in the register — the register is recording accurately that companies changed how they describe themselves.
Four limits, and the first is the one that would change most readings.
EDGAR is a federal system but the filings in it are written by registrants, and the element labels come from FASB. Neither is published here. The extraction takes values, periods, accession numbers and form types — facts — and never the explanatory notes. This page knows that a number changed, by how much and how late, and never why.
Only total assets and net income are followed. They are the two most universally reported, which makes the population comparable, but a company can restate revenue, inventory or a segment without either of these moving at all.
The 6,267 companies are those reporting total assets in one recent frame. That skews toward companies large enough to file XBRL at all, and every ranked alternative needed a market-value source this site cannot use.
1,660 periods carry more than one reported value but end where they began. They are excluded, because first-versus-last would publish each as a revision of zero — and counted, because they are real events.
Every figure on this page traces to one of these, through a CSV in
data/edgar-restatements/. Each is checked against its source query on every
build.
Every value each of the 6,267 companies has ever reported for total assets and net income, each stamped with the accession number, form type and filing date of the filing that reported it — Keyless: no account, no token, no form. SEC requires a descriptive User-Agent and answers 403 without one — a header, not a credential, and a project URL satisfies it, so no personal contact address is sent to a third party. This endpoint is why the page needs no snapshot diffing: the revision history of every figure ships with the figure
The population — every company reporting total assets at CY2024Q4 — and the company count for each revenue element in each year from 2012 to 2024 — One request returns every company reporting one element for one period, so thirteen years of industry-wide re-tagging costs sixty-five requests rather than eighty thousand. The population is defined by an element and a frame rather than by size or index membership, because every ranked universe available would have needed a market-cap source this site cannot use
The access terms this pipeline is built to, quoted: “Anyone can access and download this information for free” and “Current max request rate: 10 requests/second” — The crawl runs at roughly four requests a second against that ceiling, on four workers with a per-worker throttle. The rate is a published limit on somebody else's free service, and the whole run is cached, so waiting costs one run and hammering costs them every run
The element names used as keys throughout this page — Assets, NetIncomeLoss, and the five revenue elements whose rise and fall the last chart traces — The element names are identifiers and are published as such. The human-readable labels and documentation strings the API returns beside each fact are the taxonomy's own text, not a government work, and are never published here; a check asserts no mart can carry a label or description column
Would have explained why any given figure was restated — the explanatory notes a company files alongside the numbers — EDGAR is a federal system but the filings in it are written by registrants, so the same reasoning applies as to sponsor text on the clinical-trials page. Reported values, periods, accession numbers and form types are facts and are published; the prose is not requested and not stored. The cost is real: this page can tell you that a number changed, by how much and how late, and never why
Would have provided the same facts as quarterly bulk archives, one ZIP per quarter from 2009 onward — Same licence, same data, rejected on method rather than terms. Reconstructing a revision history from them means downloading and diffing roughly seventy quarterly archives at 60–80 MB each; the companyconcept endpoint publishes the history directly. The bulk route would have moved several gigabytes to answer a question the API answers in a few hundred megabytes of cache
Would have let the population be the largest listed companies rather than everyone reporting one element — EDGAR publishes tickers and CIKs but not market value, and every market-data source that does is licensed or keyed. So the population is defined by a taxonomy element and a reporting frame instead — which is reproducible from this repository alone, and which the page states plainly skews toward companies large enough to file XBRL at all
I wrote a companion post — “Companies Quietly Change Numbers They Have Already Published” — in simple, everyday words.