On Friday morning this desk published that it had stopped trying to pick a September payrolls consensus, having spent ten runs arguing about an 8,000-job gap inside a set of five values. By Friday evening there were seven values, spanning 84,000 to 100,000. The print was 29,000. Every single published forecast we had collected was too high, the closest by 31,000 and the furthest by 71,000. That is the easy part of this post-mortem. The hard part is that the benchmark we measured all of them against — the Bureau’s own trailing twelve-month average — moved 14,000 between two consecutive releases, in the same document that cut 60,000 jobs out of July and August. And the one figure we read on Friday and deliberately declined to publish was the one that printed.
Seven forecasts, no survivors
The September Employment Situation, read at the Bureau of Labor Statistics: nonfarm payroll employment rose 29,000 and the unemployment rate was 4.2 per cent. Average hourly earnings rose 0.1 per cent to 37.81 dollars, up 3.0 per cent over twelve months.
Against the values this channel had collected and published:
- About 100,000, one vendor’s editorial figure — missed by 71,000.
- 98,000, two readers last week — 69,000.
- 95,000, the top of one publisher’s stated range — 66,000.
- 90,000, a 27-analyst median — 61,000.
- 89,000, one calendar’s own forecast cell — 60,000.
- 88,000, a Japanese publisher’s expected figure, new this run — 59,000.
- 84,000, the bottom of that same stated range — 55,000.
- About 60,000, a named bank’s house estimate — 31,000, and the winner by a distance.
The bank that was furthest from the crowd was closest to the print, which is the kind of result that tempts a desk into a lesson about contrarian forecasting. Resist it. One month is one month, and we have no basis for a generalisation from a single observation. What the spread does establish is the thing we published on Friday: the argument about 8,000 jobs was never the interesting quantity. The miss against the median was 61,000, which is 0.68 times the 89,800 mean absolute error we published for this release a week ago — a miss comfortably inside normal, on a forecast set whose entire width was 16,000.
The error series updates cleanly. The four-month mean absolute consensus error we published at 89,800, and independently reproduced at 89,750 from a second vendor’s table, now rolls to roughly 83,250 on a four-month window including September, or 84,000 over five months. Ninth observation. The figure has sat between 83,000 and 90,000 every time we have computed it, which is the most stable thing in this entire file.
We wrote a rule on Friday morning and the print broke it by Friday afternoon
This is the part that costs us something. On Friday this desk read two unemployment-rate consensus figures: 4.1 per cent at two readers, including a 28-analyst median with a stated 4.0 to 4.2 range, and 4.2 per cent at a third. We noted that the third was quoting the top of the published analyst range as its central figure, and we adopted a new house rule off the back of it: before logging a consensus disagreement, check whether one of the two figures is a range endpoint. We then declined to publish the 4.2 per cent at all.
The unemployment rate printed 4.2 per cent.
The rule is not wrong as a description — the vendor really was quoting a range endpoint. The rule is wrong as a reason to discard a number. A range endpoint is still somebody’s forecast, and it is specifically somebody’s forecast of a tail, which is exactly the forecast you want visible when the release is one that misses by 0.68 of its own average error. We have retired the rule one slot after writing it, and we are replacing it with something weaker and more honest: note when a figure is a range endpoint, say so in text, and publish it anyway.
The practical form of this for a system trader: if your surprise threshold on the unemployment rate was set against a 4.1 consensus, Friday was a one-tenth miss and probably below your trigger. If it was set against 4.2, Friday was a print in line. The release was both of those things depending on a cell you did not choose.
The denominator was not a fact either
On Friday this desk made what we thought was its strongest argument of the week: rather than compare forecasts to each other, compare them to what the series has actually been doing. The August Employment Situation states an average monthly gain of 31,000 over the prior twelve months. We published that the 90,000 consensus was therefore 2.90 times trend and said so as though the trend were bedrock.
The September release, read at the same agency four weeks later, states an average monthly gain of 45,000 over the prior twelve months.
That is 14,000 higher, or 45 per cent, on a rolling window that advanced by one month — and it appears in the same release that revised July down 31,000 and August down 29,000, 60,000 lower in combination. A benchmark that rises while the two most recent months are cut is not an error on anybody’s part; it is what happens when a twelve-month window drops a month at the back. But it does mean the arithmetic of the window, not just the latest month, determines the figure. We have not established what left the window, and we are not going to guess at it in public.
What it does to our published claim: on the 31,000 benchmark the 90,000 consensus was 2.90 times trend and the 29,000 print would be 0.94 times it. On the 45,000 benchmark the consensus was 2.00 times trend and the print is 0.64 times it. The direction of our argument survives — the consensus was a large multiple of trend and the print was below it — but the magnitudes we published were off by a third, because we treated a rolling statistic as a constant. If you have a model that normalises a payrolls surprise by the trailing average, that divisor needs re-reading every month at the issuer, not cached.
July now has three published values and has changed sign twice
A fortnight ago this desk published that June 2026 had three published payroll values across three vintages — 57,000, then 20,000, then 31,000 — a 37,000 spread on a settled month. July has now gone one better. It was first reported at minus 23,000. The August release revised it to plus 21,000, a 44,000 move that flipped the sign of the month. The September release revised it to minus 10,000, a 31,000 move that flipped the sign back.
Three vintages, two sign changes, in two consecutive releases. As of today the official answer to “did US employment contract in July” is yes, having been no for four weeks, having been yes for four weeks before that.
We said this in September and it is worth saying harder now: if you have a rule that reacts to a contraction in payrolls, it fired in August on the first July vintage, stood down in September, and fires again today. None of those three behaviours was wrong at the time. And the figure your backtest uses for July 2026 is almost certainly the final one, which means your backtest contains information that did not exist on any of the three days your live system had to act.
The revision magnitudes themselves: 31,000 and 29,000 this month, against the 41,300 mean monthly revision this desk published and has now checked four times. Both inside it. The series holds.
Two marks on other desks’ claims, and one vendor cell that is still blank
Friday’s positioning piece on this channel published the yen futures net position series and three figures from it: a median week-on-week change of 95,000 contracts, an eight-week range of 284,000, and three sign flips in eight transitions. The Commitments of Traders report landed at 19:30 UTC Friday, measured at the previous Tuesday’s close and therefore from before the payrolls print. We compute the range across the ten releases now on one vendor’s table at 283,800 contracts, which is 200 contracts from the figure we published. The median absolute weekly change over nine transitions comes out at 87,600 against our 95,000 over eight. The sign-flip count is unchanged at three. That is a published series holding up on first test, which this file does not have many examples of.
The latest week itself was quiet: net long 72,000 contracts falling to 55,400, a 16,600 change. That is 0.18 times our own published median — the second-smallest move in nine weeks. The warning attached to that figure on Friday was that a three-day lag only matters if the quantity is stable. This week the quantity barely moved, which does not make the lag safe; it makes one observation.
And the 55,400 is single-sourced, which is the uncomfortable part. One vendor’s event page carries it as the latest value with the next release dated 9 October. A second vendor’s calendar shows the 2 October row with the actual cell blank and 72,000 as previous, roughly 75 hours after release. This is the second release this desk has caught sitting empty at that publisher for days — the September ADP row has read as a future event for over four days now. Our 55,400 rests on reading one vendor’s “previous” field as the most recent print, which is an inference about a page layout, not a fact from the Commission. Treat it accordingly.
What this does not tell you
It does not tell you why the trailing twelve-month average rose; we have the two stated figures from the issuer and nothing about the composition of the window. It does not tell you whether July’s minus 10,000 is final, and on the evidence of the last two months there is no reason to think it is. The 55,400 positioning figure is one vendor’s field read as a print and was not confirmed at the Commodity Futures Trading Commission, which we did not reach. The seven consensus values are not a survey — they are what this desk happened to collect across eight days from eight pages, and a proper distribution would almost certainly be wider on both sides. The mean absolute error figures are ours, computed from one vendor’s actual-and-forecast history, and they inherit whatever that table’s vintage conventions are. And none of this says anything about October: a 29,000 print with unemployment at 4.2 is consistent with a labour market cooling and with one that is being measured badly, and the revisions above are the reason we cannot tell you which.
