From 26.4% to 3.0%: earnings filings almost never reach the SEC while the market is open

In 2003-2007, more than one in four S&P 500 earnings announcements were filed with EDGAR while the market was open. In 2022-2026, one in thirty-three. Here is the full series year by year, the four checks we ran on it before saying anything, and the part this measurement does not know — which is not small.

Announcements measured	64,938, from 808 companies, Apr 2003 – Sep 2026
Market open, 2003-2007	3,127 of 11,849 — 26.4%
Market open, 2022-2026	375 of 12,556 — 3.0%
Only the 516 companies present in both blocks	25.7% → 3.2%
Recomputed from the SEC's raw header	26.4% → 2.8%
Companies whose own rate falls	351 of 498

Download the full CSV · what this file is · the dates, company by company

01   What is being counted

A listed company reports results by filing an 8-K with item 2.02. Every one of those filings carries the time EDGAR accepted it. If that time falls between 09:30 and 16:00 New York, the filing landed while the market was open.

That is the whole calculation: a subtraction and a percentage over the 64,938 rows of a file we publish in full, free and without an account. The arithmetic is no achievement. What is hard is having the series with times at all, and what has to happen before publishing a number like this are the checks, which are half of this article.

One distinction worth keeping: announcing results is not filing the accounts. The report (10-Q or 10-K) comes afterwards, sometimes more than a week afterwards. This is about the announcement.

02   The shape: it falls every year, and there is no step

Two blocks twenty years apart are easy to pick to suit an argument. This is the full series, uncut:

Each bar is the share of that year's announcements filed with EDGAR between 09:30 and 16:00 New York. The dotted line is the same count redone from the raw header of the filing, with no time converted at all (section 04). 2003 starts on 25 April, the file's first announcement, and 2026 runs to 1 September: both are partial years.

The fall is continuous. The largest year-on-year drop is 3.2 points, from 2003 to 2004, followed by 2.7 (2009), 2.7 (2008) and 2.5 (2005). None of them stands out from the rest: there is no single year where the series jumps and stays down, which is what a new rule or a change of system would leave behind. It falls, and it keeps falling.

We are not saying why. We have not measured that and we do not know. A series can be described without inventing a cause for it.

03   The universe is not the same in 2003 and in 2026

Comparing 11,849 announcements with 12,556 is fine. Comparing different companies is not. The first block has 642 companies with at least one announcement; the second has 661; 516 are in both. It could be that what changed is not anyone's behaviour but the list.

So the count is repeated using only those 516: 25.7% → 3.2%. Practically the same.

And it can be tightened further, by comparing each company against itself. Of the 498 with four or more announcements in each block, 351 lower their own rate and 26 raise it; the remaining 121 do not move, and 120 of those were already at zero in both blocks. The median change is −11.1 points. This is not a composition effect: it is the same companies changing their hour.

The figure that does point at composition, stated too: the 145 companies that appear only in the second block announce with the market open 2.0% of the time. They push downwards, but not by much, and without them the fall is still there.

04   The check that makes the rest believable

There is one way this whole article could be false, and it is this: the acceptance stamp EDGAR's API returns ends in "Z" — meaning it claims to be universal time — but for part of the companies that clock is already New York time. Treating an already-New-York stamp as universal moves the filing four or five hours, which is exactly what it takes to cross the open or the close. If that convention had drifted over the years, the decline in the chart would be an artefact of the stamp and not anyone's behaviour.

The mechanism — that the "Z" is not always universal — was found by Tilman Ambach, cross-checking filings by hand. The question of whether the stamp could be trusted was put on the table by Ian Gow. Neither is our finding; our part is having run it across all 64,938.

The check was done twice, from weaker to stronger.

First: only the companies whose stamp always converts. There are 624 companies whose stamp behaves the same way in every one of their filings. Restricted to them, the count gives 25.6% → 2.9%.

Second, and this is the good one: no convention at all. The raw header of the filing — the SGML block that sits at the top of the file, ahead of the document — carries its own ACCEPTANCE-DATETIME field, and that one is always in Eastern time. Nothing to infer, nothing to convert. Re-reading those headers for 64,936 of the 64,938 announcements, the classification agrees with the published one on 99.78% of rows and the split comes out 26.4% → 2.8%. That is the dotted line in the chart, and it sits on top of the bars.

05   "After the close" is not a time

Something else visible in the time column that did not fit any of the counts above. The 27,142 rows labelled "after the close" are not a moment: they are a six-hour band, and it holds everything.

The three bands add up to 64,836 and the file holds 64,938: there are 102 announcements with no publishable time. A hundred of them come from two companies whose stamp behaves both ways at once, so neither reading can be settled; the other two are 2009 filings whose stamp lands on midnight exactly, which is to say with no real time at all. They are flagged as such in the CSV rather than assigned by guesswork.

A filing that lands at 16:05 and one that lands at 20:40 carry the same label and are not the same event for anyone measuring the price reaction. Using the label instead of the time puts things four and a half hours apart in the same box.

Both bands have also tightened against their edge. In 2003-2007 the median after-close filing was at 16:45 and the median pre-open one at 08:08; in 2022-2026 they are 16:16 and 07:02.

06   What this measurement does not know, and how much it weighs

That doubt can be given a number. If the release were fifteen minutes ahead of the filing in every case, 8,707 rows would cross the close and become intraday. That is 13.4% of the whole file, or, if you prefer, roughly a third of the 27,142 after-close rows. Either phrasing is fine; what is not fine is mixing them and saying "13% of the after-close ones", which is wrong by a factor of 2.4.

And there is something more uncomfortable to say, because it changes how the headline reads. That adjustment does not fall evenly on the two blocks. In 2003-2007 the median after-close filing was at 16:45 and only 17.3% of them landed in the first quarter of an hour; in 2022-2026 the median is 16:16 and 44.2% land there. With the flat fifteen-minute adjustment, the first block would go from 26.4% to 30.4% and the second from 3.0% to 22.7%: the gap between the two blocks would shrink from 23.4 points to 7.7.

That is not an estimate of what actually happens, and it is not an elegant way of withdrawing the article either. It is a sensitivity test with a flat assumption that is almost certainly wrong: a company filing at 16:02 is doing it against the bell on purpose, and its release does not look like it went out at 15:47. But we cannot prove that from this file, so it goes in writing. What is measured, and depends on no assumption, is where the document lands. Closing the question properly needs a source that timestamps the press release, and we do not have one.

And every snapshot is dated. The SEC does not document this field anywhere, and it also repairs the JSON backwards: filings that come out unconverted today can turn up converted weeks later. This is undocumented behaviour that is inconsistent with itself, not a breach of anything promised. The figures on this page come from the file generated on 4 September 2026 and from a re-read of the raw headers done on 6 September 2026. Another snapshot can move the last decimal.

07   How to check it without going through us

A figure that can only be checked by going through our website is not checkable. The three steps, with what each one needs:

• The chart and the year-by-year. They are in the CSV, at https://quant500.com/api/descarga/anuncios.csv: no account, no key, CC0 licence. Grouping by year and counting the rows whose session column says during is twenty lines of code. The script we used, medir_hora_del_anuncio.py, is in the project repository and prints every figure on this page, the chart included.
• The stamp check. sello_sgml.py downloads the raw ACCEPTANCE-DATETIME of each filing and convencion_por_empresa.py groups the result by company. Both run against the public CSV and against the SEC: they need nothing of ours. If you would rather not use our code, any filing's header sits at www.sec.gov/Archives/edgar/data/<cik>/<accession without dashes>/<accession>.txt, in the first few lines, and the field can be read straight off it.
• Any single row. Every CSV row carries its link to the original filing on EDGAR. Any figure here can be checked against the SEC's own paper without asking us, and if something does not add up, the error is ours.

08   Who each piece belongs to

The stamp mechanism was found by Tilman Ambach. The question of whether the stamp was reliable was put by Ian Gow. That the acceptance time is an upper bound and not the moment of the announcement was pointed out by Bill McDonald. Ours is the file, the re-read of the 64,936 filings and the counts on this page, mistakes included.

This is not investment advice. Quant500 publishes data and its historical reconstruction. No figure on this page is a suggestion to buy or sell.