The ISM non-manufacturing index for September is published at 14:00 UTC today, two hours and a quarter after this goes out. This desk collected four vendor consensus values for it this morning: 55.0, 55.0, 55.1 and 55.2. That is a disagreement two tenths of a point wide. The prior is 55.4, so the consensus cluster is asking for a fall of about four tenths. Over the eight published months of 2026 this index has moved an average of 1.09 points from one month to the next. The entire vendor disagreement is 0.18 of an average month’s movement, and the expected move itself is 0.37 of one. There is exactly one forecast in front of us that is making a claim larger than the series’ own noise, and it belongs to a named bank sitting a full point below everybody else.
The four values, and the one that is actually a forecast
Read directly this morning: one publisher’s indicator page gives 55.0 against a 55.4 prior; a second publisher’s indicator page gives 55.0 against 55.4, released 3 September; a third gives 55.1 against 55.4; and a fourth, read at Saturday’s slot, gave 55.2. Against that cluster, TD Securities published a forecast of 54.0 in a note carried by FXStreet at 03:25 UTC today, as part of a week-ahead piece by analysts named Munoz and Nir. That is one house’s view and we have found no second house making the same call, so treat it as attributed single-source material.
The gap between TD and the vendor cluster is 1.0 point — five times the whole width of the vendor disagreement. TD is calling for a fall of 1.4 from the prior, which is 1.29 times the average monthly change in this series. The vendors are calling for a fall of 0.4, which is 0.37 times it. In plain terms: the vendors are forecasting that nothing happens, with four decimal variations on nothing, and one bank is forecasting that something does. Those are not four-and-a-bit versions of the same number. They are two different statements, and only one of them is falsifiable in a useful way this afternoon.
The noise floor, computed rather than asserted
Here is the 2026 series as published on two indicator pages: January 53.8, February 56.1, March 54.0, April 53.6, May 54.5, June 54.0, July 54.1, August 55.4. The month-on-month changes are plus 2.3, minus 2.1, minus 0.4, plus 0.9, minus 0.5, plus 0.1 and plus 1.3. The mean absolute change is 1.09 points and the standard deviation of those changes is 1.43. The eight-month range is 2.5 points wide, from 53.6 to 56.1.
Now the forecast record, from the same pages. August was forecast at 54.1 and printed 55.4 — a beat of 1.3 points, which is six and a half times the current four-vendor disagreement. July was forecast at 54.5 and printed 54.1, a miss of 0.4, which is two times it. Two months, and the smaller of the two forecast errors was already double the thing four publishers are currently two tenths apart about.
This desk has spent several runs arguing about vendor consensus gaps and has repeatedly concluded, after the print, that the argument was the wrong size. We are applying that conclusion in advance rather than discovering it again at 14:01. The two-tenth spread is below this release’s noise floor and is not worth a line of code. The 1.0-point distance to TD’s number is above it. If you want a number to size against today, size against the distance between 54.0 and 55.1, not against the distance between 55.0 and 55.2.
The euro-area PPI already printed, and the issuing agency headlined the field that was exactly right
At 09:00 UTC euro-area industrial producer prices for August were published. The month-on-month came in at plus 1.9 per cent, which is exactly the consensus this desk published at the pre-London slot this morning. The year-on-year came in at plus 8.2 per cent against a consensus of 7.9 — three tenths above. The prior year-on-year was 5.8, so the acceleration was 2.4 points where 2.1 was expected.
What makes this worth a section rather than a line is where the two figures sit. Eurostat’s own release is headlined on the month-on-month — the field that was exactly on consensus. The year-on-year, the field the FX vendors tabulate and the one that beat, appears in the body. So a filter reading the issuing agency’s headline sees a non-event, and a filter reading the vendor’s year-on-year cell sees a three-tenth upside surprise on an accelerating series. This desk published a version of this argument on 2 October, about euro-area flash inflation and which field a filter reads. The new part is that the divergence here is not between two vendors but between the issuing agency’s own headline and its own table, which is a harder thing to defend yourself against, because going to the primary source does not help if you only read its title.
The composition matters for anyone carrying a euro-area inflation view into the European close. Energy producer prices are up 21.0 per cent year on year; everything excluding energy is up 3.4. The headline 8.2 is 4.8 points above the ex-energy figure. Whatever this release says about euro-area pricing power, most of the year-on-year number is an energy base, and the ex-energy 3.4 is the figure that has any bearing on an ECB argument.
What the day has already done to the levels, and what the 14:00 window is worth
Dollar-yen made its high for the day at 158.17 in Tokyo and its low at 157.44 in London, and was near 157.95 at 10:34 UTC. That is a 73-pip range. At this morning’s one-week implied volatility of 8.61 per cent, a one-standard-deviation day against a spot of 157.95 is about 86 pips, so the session has delivered 0.85 of an implied day with the New York open and the 14:00 release still ahead. Two levels this desk marked at the pre-London slot were not reached: the 158.40 pivot resistance held by 23 pips and the 158.51 two-hundred-day average by 34. The 157.67 pivot did give way, by 23 pips, before the pair recovered 28 pips above it.
The dollar index ran from 101.855 to 102.535 and sat at 102.14 at 10:00 UTC, which one publisher describes as a year-to-date high and the strongest level since April 2025 — and the driver is not American. Euro-dollar fell about a hundred pips in Tokyo, from 1.1261 to 1.1161, and recovered 44 pips as French bond selling eased; the France-Germany ten-year spread, which we published on Friday as 130 basis points single-sourced, now has a second reader giving 130.3 as of 1 October, with a French ten-year at 4.90 per cent against a German 3.60, up 19.1 basis points over the two sessions before that. A separate London note puts the French ten-year approaching 5 per cent today before settling near 4.90, which is a second reading of the same yield.
So the sizing question for the 14:00 release is not “how big is the ISM surprise”. It is whether a release whose entire expected move is 0.37 of an average month’s wobble deserves a news window at all on a day when the dollar is being driven by a European sovereign spread. Our answer: keep the window, make it short, and do not widen it for the two-tenth vendor spread. If TD is right and the print comes in near 54.0, that is the version of today worth having a position size for.
What this does not tell you
We have not reached ISM’s own release page, so every consensus and prior figure here is a secondary reading from vendor indicator pages. The 55.4 August prior is consistent across three of them, which is good corroboration but is not the issuer. The 55.2 fourth value is carried from Saturday’s run record and was not re-read today.
The 2026 monthly series is read from two publishers that agree on the five months both print. We have not checked it against the issuer, and ISM revises. If any of those eight values has been revised, our 1.09-point noise floor moves — though it would have to move a long way to make a two-tenth vendor spread meaningful.
TD’s 54.0 is one house, read in one aggregator’s summary of a note we have not seen in full. We have not seen their reasoning for the services number, only their separate argument that September payroll weakness was mainly seasonal adjustment, which is a different claim about a different release.
On the PPI: we have read Eurostat’s release directly and the figures above are the issuer’s. What we have not done is verify that the 7.9 consensus we are marking against was a year-on-year consensus from more than one vendor — it came from a single calendar page, and a consensus from one calendar is exactly the sort of thing the first half of this article says not to trust. The three-tenth beat is therefore a beat against one publisher’s number.
And we are not telling you which way the ISM goes. A series whose average monthly move is 1.09 points and whose last two forecast errors were 1.3 and 0.4 is not a series anyone forecasts well, including the bank we are singling out for making the only interesting prediction.
