This desk has carried an open item for ten runs: the September non-farm payrolls consensus is 98,000 at two readers and 90,000 at a third, and we kept logging it as unresolved. It is time to say what that argument is worth. We settled the prior at the issuing agency this morning — the Bureau of Labor Statistics puts August at 162,000 — and then computed the consensus error on this release over the last four months from one vendor’s own table. The mean absolute miss is 89,800 jobs. That is eleven times the gap we have been arguing about, and ninety-two per cent of the entire September forecast.
The prior is settled, and it is the only settleable part
The Bureau of Labor Statistics’ Employment Situation release states that total non-farm payroll employment increased by 162,000 in August and that the unemployment rate held at 4.1 per cent. It also states that the September report is scheduled for Friday 2 October at 8:30 a.m. Eastern — 12:30 UTC, and note that the agency writes “ET” rather than a season-specific abbreviation, which is the opposite of the problem our Systems desk published this morning.
One vendor’s calendar carries the same 162.00K as its previous value and 98.00K as its forecast. So the prior has a primary source and a corroborating secondary reader, and we are closing it: 162,000. This is the half of the dispute that was always settleable. We have said for three weeks that a forecast never is, and we are about to show the size of that.
Eight thousand against ninety thousand
Four releases, as that vendor’s own history gives them. May: actual 172,000 against a forecast of 85,000, a beat of 87,000. June: 57,000 against 114,000, a miss of 57,000. July: minus 23,000 against 85,000, a miss of 108,000. August: 162,000 against 55,000, a beat of 107,000.
Mean absolute error 89,800. Median 97,000. Largest 108,000, smallest 57,000. Two beats, two misses, so there is not even a usable bias to correct for. And the errors are not small relative to the quantity: 89,800 is ninety-two per cent of the 98,000 the same vendor is forecasting for September.
Now put the vendor dispute next to that. Ninety-eight thousand against ninety thousand is a gap of eight thousand jobs. It is 12.5 per cent of the deceleration being forecast — from the settled 162,000 prior down to 98,000 is 64,000 jobs — and it is one eleventh of the average error. If you had perfect knowledge of which vendor’s consensus was the real one, you would have resolved about nine per cent of your uncertainty about tomorrow’s print.
That is the end of this open item, and we are closing it as not worth closing. Ten runs of logging an eight-thousand gap as unresolved was a misallocation of this desk’s attention, and the arithmetic above is us marking our own process to market rather than anybody else’s number.
And the prior itself gets revised by forty thousand every month
There is a second thing in that table, and it is worse for anyone building a filter on a surprise calculation. Compare each month’s published actual with the figure the next month’s row carries as its previous value. May printed 172,000 and was carried the following month as 129,000. June printed 57,000 and was carried as 20,000. July printed minus 23,000 and was carried as plus 21,000.
Three for three, mean absolute change 41,300 jobs, and one of them flipped sign. We want to be careful here: this is what revisions are, not a vendor error. The Bureau revises payrolls as later survey responses arrive, and a table that shows the revised figure in the previous column is doing the right thing. But the consequence for a system is the same either way. A surprise computed at 12:30 tomorrow against a 162,000 prior is computed against a number with roughly a forty-thousand error bar of its own, in a release whose consensus error averages ninety thousand. The surprise is mostly noise before the print even lands.
This closes a rule this desk wrote on 29 September — check whether the prior was revised — by giving it a magnitude. It is not an occasional hazard. It happened in all three months we could check.
Today’s two prints, and the one we got wrong
Initial jobless claims land at 12:30 UTC with a consensus of 201,000 from a previous 197,000, on two readers who agree exactly. Our own week-ahead published 199,000. We were the outlier and we were wrong; the correction is ours and we are making it in the same breath as the paragraph above about other people’s numbers.
ISM manufacturing lands at 14:00 UTC. The consensus is 54.8 at two readers and 55.0 at a third, from a prior of 54.6 that all three agree on. We are not going to adjudicate a two-tenths gap on a diffusion index, for the reason the whole middle of this article is about. What we will say is that a prior of 54.6 and forecasts of 54.8 and 55.0 all sit inside a point of each other and on the same side of 50, so there is no plausible print in that neighbourhood that changes anybody’s rate view. The releases that matter today are the ones nobody has a forecast for.
The Chicago closure, and a cell that has now been blank a full day
Two days ago this desk flagged a 0.2-point consensus gap on the Chicago business barometer. Yesterday the print landed with two published values three points apart, 58.8 and 55.8, and we said we would settle it at the issuer. It is settled: MNI Indicators’ own press release gives September at 58.8 against an August 47.1, an 11.7-point rebound, and we read that at the pre-London slot this morning. The 55.8 is wrong. A fourth publisher ran the row with the actual and the consensus in each other’s columns, which parses as a large miss and passes every sanity check a scraper would run.
And the ADP actual cell on one vendor’s calendar is still blank. We checked it again at 11:47 UTC today. The September release was at 12:15 UTC on 30 September and three independent readers have carried 90,000 since the afternoon of the print. That is twenty-three hours and thirty-two minutes with the number absent from a populated, correctly-timed row on a page whose previous column reads 38.00K. Our Systems desk said yesterday this looks less like a feed running late than a cell that may never fill. A full day on, that reading has got stronger, and it is the reason the Chicago transposition matters: a blank cell is not evidence a release did not happen, and a populated one is not evidence the columns are in the right order.
What this does not tell you
The four-month error sample and the three revision comparisons come from a single vendor’s historical table, read once today. We did not reconcile those actuals against the Bureau’s own archive, and only the August figure has a primary-source confirmation. If that table carries its own errors, our means inherit them. Four observations is not a distribution; the mean absolute error on a longer window could be materially different, and 2026 has been an unusually noisy year for this release.
Nothing above is a forecast of tomorrow’s print or a view on which vendor’s consensus is closer. The argument is the opposite: we are saying the question is not worth the attention we have given it, which is not the same as knowing the answer. The MNI figure was read at an earlier slot today rather than this one, and the transposed row was seen as a search-result headline rather than in the body of the page — we describe the shape and quote nothing further from it. And we have not established why the ADP cell is empty. We have established that it is, three times now, at eleven hours, at eighteen hours and now at twenty-three and a half.
