September job openings printed 7,523 thousand against consensus figures clustered near 7.1 million — a beat of roughly 420,000, above every estimate in at least one publisher’s survey. The agency that published it describes the month as “little changed at 7.5 million.” Both sentences are true, and a system keyed to the headline verb and a system keyed to the issuer came out of 15:00 UTC with opposite signs. The reason the two readings diverge is the thing worth your attention: the same release that delivered September rewrote August in five fields, and the month-on-month change everyone quoted depends on which version of August your calendar was holding.
Two true sentences about one number
The level is as solid as a figure gets on this channel. Four publishers give it identically — 7,523 thousand, 7.52 million, 7.523 million, 7.52 million — and the issuer’s text rounds it to 7.5 million with an openings rate of 4.5 per cent. There is no measurement dispute here at all, which is unusual enough to note.
The characterisations are not identical. One publisher titles it a jump and says openings surged. One says it hit the level and topped forecasts. One says openings unexpectedly rose. One reports it as up 333 thousand and, in the same article, quotes the issuer’s own “little changed.” Not one of those four verbs is the issuer’s word, and the issuer’s word is the one attached to a statistical convention rather than to an editor’s judgment about surprise.
Both framings are defensible because they measure different distances. Against the prior month the change is +333 thousand on a base of 7,190 — under five per cent, which the agency treats as not meaningfully different. Against the forecast the miss is about 420 thousand. If your rule reads a release as a surprise, it fired hard. If your rule reads a release as a change in the level of labour demand, it did not fire at all. Neither rule is broken. They are answering different questions and tonight they disagreed, and a desk that stores only one of the two numbers will not be able to tell later which question it answered.
The release rewrote its own prior in five fields
This is the part no headline carried. In the September release the issuer revised August openings up by 111,000 to 7.2 million, August hires up by 79,000 to 5.3 million, August total separations up by 105,000 to 5.2 million, August layoffs and discharges up by 127,000 to 1.8 million — and August quits down by 22,000 to 3.0 million. Five fields, and one of them moved the other way.
Do the arithmetic the two ways and the gap is visible. August as published on 29 September was 7,079 thousand. August as the September release now states it is 7,190. So a calendar still holding the figure its own vendor published five weeks ago computes 7,523 minus 7,079, or +444 thousand month-on-month. The issuer’s arithmetic is 7,523 minus 7,190, or +333. Same print, a 111-thousand difference, and none of it is a measurement error — it is a revision arriving inside the document that makes the comparison possible.
And that stale prior was live on the morning of the release. One briefing we read carried the day’s calendar with job openings prior 7.079 and consensus 7.13, and quits prior 3.066 with consensus 3.05. Both of those priors are the pre-revision figures; the issuer has since moved them to 7,190 and 3,044. A system that scored tonight’s print against that calendar got the surprise roughly right and the change badly wrong, in the same instant, from the same row.
We have written before that you should check whether the prior was revised. We had been treating that as a thing to check before a release. Tonight makes the sharper point: on this series the revision is the release, and there is no moment at which the prior is stable enough to check.
Our consensus spread was a rounding error against the error
Yesterday morning at 11:40 UTC this desk published that three readers were 34 thousand apart on this figure, and gave that spread a clause of its own. The print missed all three by roughly 420 thousand. The dispersion we measured was about one-twelfth of the error we did not.
Widening the reader set this morning does not help. We now have 7.10 million on one card, attributed to a named feed; about 7.164 million on a broker calendar; 7.1 million unattributed on a third; 7.13 on a fourth; and one publisher declining to give a number at all and saying only that the print came in above every estimate in its survey. That is a 64-thousand spread across four figures, and it did not matter either. The publisher that refused to pick a number produced the most useful sentence of the five.
This desk keeps arguing about which vendor has the right consensus and then being wrong about the print by an order of magnitude more than the argument was worth. We are not going to keep re-deriving the lesson, so here it is as a rule: the spread between vendors’ consensus figures carries no information about the variance of the release. It measures how much the forecasters copied each other, which is a fact about forecasters. Use a house’s conditional threshold if you want something that survives the print; a point estimate, and a fight about the third digit of a point estimate, will not.
“Lowest since March 2025” is scored against a number this release moved
One publisher reports that layoffs fell to the lowest level since March 2025. Layoffs and discharges printed 1,641 thousand at a 1.0 per cent rate, which the issuer calls “changed little.” The August figure on the other side of that comparison was revised up by 127,000 to 1.8 million — in this release. So some part of the fall is a revision to the month being compared against, and we cannot tell you how much, because the claim is single-sourced and the publisher does not show its series.
We are not saying it is wrong. We are saying it is one publisher’s claim, resting on a prior that moved inside the document that produced the claim, and that is a different kind of weak than a number two vendors disagree about.
Quits are the field that actually did something: 3,365 thousand, up from a revised 3,044, a 2.1 per cent rate, and the issuer puts the rise over the year at 375,000. One publisher gives the month-on-month as +321,000. That same publisher carries two different quits consensus figures on one page — 3.05 million in its body and 3.096 million in its key takeaways — and states a beat of 315,000, which only works against the first of the two. We have published before that a consensus is not one number. Here it is not one number inside one article.
The one place the issuer names a driver, it is unglamorous: openings increased in state and local government excluding education, by 104,000.
Tonight’s cluster, and which prior your system is holding
ADP’s October report is at 13:15 UTC. The hour is three-sourced and we are reasonably confident in it; the date is less settled than it looks, because the aggregator giving us that hour labels the release “estimated” and says in its own text that ADP had not formally confirmed it at the time of writing.
Before you score it, decide which prior you hold, because one publisher’s single page for this release carries three. Its body says September was +90,000, citing the company’s own media centre. Its table says prior 44,000 and labels it July. Its sidebar for the 2 September release says +38,000 against a 48,000 consensus with the prior revised to +46,000. Its consensus is 65,000 in the body, attributed to one feed and “other market calendars,” and “not yet published” in the table two screens down. We publish no ADP consensus of our own tonight, because the only figure we found contradicts itself on the page that carries it.
ISM Services is at 15:00 UTC, and this one is clean at the issuer. ISM’s own release calendar puts the Services report on 4 November and states the standing rule as the third business day of the month at 10:00 a.m. EST. That label is worth a second: a week ago “EST” on that page was wrong, because the United States was on daylight time; since Sunday it is correct, so the page became accurate without being edited. 15:00 UTC is also the minute of the New York option cut, as it has been since the clock moved — two things in one minute, for the third session running.
For scoring it, September printed 54.9 against 55.1 expected, down from 55.4 in August, with Business Activity 56.5 from 61.7, New Orders 59.8 from 60.9, Employment 50.1 from 47.8, Supplier Deliveries 53.2 from 51.3 and Prices 74.0 from 72.6. Note which of those is in the headline: the index is the mean of Business Activity, New Orders, Employment and Supplier Deliveries, and Prices — the one at 74.0 — is not in it. A services print that moves on prices moves the sub-index and not the number your filter reads.
The rest of the day: the ECB wage tracker at 09:00, Lane at 13:20, the weekly petroleum report at 15:30, and the Banque de France governor at 18:45. One calendar also carries the Treasury’s quarterly auction plan with no hour against it at all, which is the row this desk said on Monday was somebody’s inference rather than a scheduled event. It is still somebody’s inference, and now it has a date and no time.
What this does not tell you
We did not read the issuer’s detail tables, so every per-category figure in thousands above — 7,523, 5,402, 5,319, 3,365, 1,641, 313 — is a secondary reading cross-checked between publishers, not a number we took off a BLS table ourselves. The issuer’s own text rounds to one decimal and we have reported its rounding separately from the publishers’ precision.
We also cannot tell you what the market did with it beyond a direction. One Tokyo headline says the ten-year rose on the print; nobody we could open publishes where it settled. Our session piece this morning refuses that number for the same reason.
No October consensus is given here for ISM Services, because we could not read one at a source we would stand behind — a prediction-market bucket is not a consensus. The ADP consensus is omitted for the reason given above. And none of this is a view on Friday’s payrolls: this release is a September reference month counted on the last business day of September, and the surveys that disagreed about employment yesterday both closed collection on 27 October.
One procedural note for anyone scraping. At the time we checked, more than eight hours after publication, the issuer’s summary page for this series was still serving the August release, stamped 29 September, while the September release sat at a different address stamped 3 November. We cannot tell whether that is a stale cache or the page’s normal behaviour, and we are not asserting a fault. But a scraper pointed at the page you would reach by searching for this series read a five-week-old release at a moment when the new one was public.
Related
- FX event calendar — ADP at 13:15, ISM Services and the New York cut both at 15:00, petroleum at 15:30
- Signals — scoring a release whose prior moves inside the release
- EA presets — news-window settings for two releases 105 minutes apart
