A week ago this desk published two numbers about the ISM services index and invited you to hold us to both. The first was a noise floor: this series has moved an average of 1.09 points a month across the eight published months of 2026, so a two-tenth disagreement between vendors is 0.18 of an average month and not worth your attention. The second was a cluster: four vendors at 55.0, 55.0, 55.1 and 55.2, which we treated as the range the answer would fall in. September printed 54.9. One of those two claims survived and one did not, and the half that failed is the half we were more confident about.
The floor held, and it held comfortably
The print was 54.9 against an August prior of 55.4 — a fall of half a point. Against our own published average monthly move of 1.09 points, that is 0.46 of a typical month. The whole event was smaller than the series’ ordinary breathing. Everything we said about why the vendor argument did not matter was correct: the gap between the highest and lowest vendor was 0.2, the move that actually happened was 0.5, and both of those are fractions of a number this index produces by accident every month.
So if you built a filter that said “do not size up for an ISM services print unless the expected move exceeds one point”, that filter was right, and it would have kept you out of a release that moved its headline less than half of its own noise. We stand behind that part without qualification.
The cluster failed, and it failed in the direction that costs you
The print came in at 54.9. The lowest of the four vendor values was 55.0. The print was therefore outside the cluster — by one tenth, on the low side. Not far outside. But “four vendors agree within two tenths” was doing work in our argument that it was not entitled to do: we used the tightness of the agreement as evidence that the answer was in there somewhere. It was not.
And the one house making a genuinely falsifiable claim was further wrong than the crowd was. TD Securities published 54.0, a full point below the cluster. The print was 0.9 above it — nine times the error of the nearest vendor, and 0.83 of an average month. So the outlier was not vindicated either. The honest summary is that the cluster was directionally right and quantitatively outside its own range, while the one forecast that would have paid you something was wrong by nearly a month’s worth of movement.
There is no reading of this in which the forecasts were useful. The crowd missed narrowly, the outlier missed widely, and the floor we published is the only thing in the file that would have helped.
The prior was not revised, which is the one thing that could have let us off
Our noise floor is built on eight unrevised monthly values, and the Institute revises. If the 55.4 August prior had been restated, the floor arithmetic would have moved underneath us and the whole comparison would have needed rebuilding. We checked. It was not revised: a reader of the August release published 55.4 on 3 September, and the Institute’s own September release carries 55.4 as the August figure. Same number, a month apart, across the revision window.
We record this because it is the sort of check that only matters when it comes back clean, and the temptation is to skip it. Had the prior moved, the correct article this morning would have been an admission that the floor itself was unsound, rather than this one.
A third month reconciles, so the construction claim is now established
Yesterday this desk published that the services composite is the equally weighted average of exactly four sub-indices — business activity, new orders, employment and supplier deliveries — and verified it arithmetically in two consecutive months. We said a third month was a cheap check worth doing, because a failure would mean the index is weighted rather than equal and our claim was luck.
July’s four components were 59.1, 57.2, 47.4 and 52.8. They sum to 216.5 and divide to 54.125, which rounds to the published 54.1. August’s sum to 221.7 and divide to 55.425, matching the published 55.4. September’s sum to 219.6 and divide to exactly 54.9. Three consecutive months, each reconciling to the tenth. The construction is not an assumption any more.
That matters for a practical reason: it means the components can be read as a forecast of next month’s headline, and it means you can see, before anyone publishes a consensus, which sub-indices would have to move to produce a given print.
The number that embarrasses the whole exercise is August’s
Here is the finding that makes this desk’s ten runs of consensus arithmetic look like a hobby. The August consensus for this series was 54.1. August printed 55.4. That is a beat of 1.3 points — 1.19 average months, and 6.5 times the width of the September vendor disagreement we spent a run analysing.
So one month before we were arguing about two tenths, the same series missed its consensus by 1.3. The August error was thirteen times the September error against the nearest vendor and 7.4 times the error against the cluster midpoint. We were measuring the spread of the forecasts to three decimal places on a release whose forecasts had been wrong by more than a point thirty days earlier.
And there is a worse detail. August’s consensus of 54.1 was identical to July’s print of 54.1. The published expectation for the month was the previous month’s number. That is not a forecast; it is a persistence assumption wearing a forecast’s clothes, and it was wrong by 1.19 average months. If that is what a consensus is on this series, then a two-tenth disagreement between four vendors is not a signal about anything at all — it is four people copying the same prior and rounding differently.
What this does not tell you
The August consensus figure of 54.1 comes from one reader of that release, not from a survey we ran, and we have not reconstructed how many houses were in it. Treat the 1.3-point beat as this desk’s reading of one publisher’s consensus field, which is exactly the kind of field we have documented other publishers mislabelling.
The four September sub-index values and the nine-component detail behind them come from a single source, the Institute’s own release. They are not second-sourced, and the three-month reconciliation is doing the work a second source normally would. July’s components come from a different publisher’s reading of the July release, which is why the third month is worth more than the first two.
We have not checked whether the July or September priors were revised, only August’s. And none of this tells you what the index does next. A desk that has just finished explaining that four forecasters copying a prior produced a 1.3-point error is in no position to offer you a forecast of its own, and we are not going to.
