This desk spent yesterday morning flagging a live vendor gap on the September Chicago PMI: one calendar had the consensus at 51.0, another at 51.2. Two tenths of a point. The release has now printed, and the print itself has two values — 58.8 and 55.8, three full points apart, from publishers reading the same MNI release on the same afternoon. The argument we were having about the forecast was fifteen times narrower than the disagreement about the answer.
The consensus was two tenths apart. The actual is three points apart.
RTTNews, quoting MNI Indicators directly, gives the September Chicago Business Barometer as 58.8, up from 47.1 in August, against an expected 51.0. Trading Economics gives 58.8 from 47.1 against a forecast of 51.2. Fisco’s New York wrap gives 58.8 against 51.0. And GuruFocus gives 55.8 — same month, same prior of 47.1, same consensus of 51.0, a number three points lower.
Three readers on 58.8 and one on 55.8 is not a coin flip, and this desk’s working figure is 58.8. But that is not the point. The point is that a system trader who wanted to know whether the print beat expectations by four and a half points or by seven and a half points had, for most of yesterday evening, no way to tell from the open web without picking a publisher. The prior was agreed by all four. The consensus was agreed to within two tenths. The one field that is supposed to be a fact is the one with the spread in it.
We have published vendor disputes about consensus figures for three weeks. A consensus is an opinion and it is allowed to differ. An actual is a measurement that already happened, and when two publishers give it three points apart, at most one of them has read the release.
Four consensus figures for ADP, and the print beat every one of them.
ADP’s September national employment report printed 90,000 at 12:15 UTC against a prior of 38,000. Four consensus figures were in circulation for it. FXStreet carried 72,000. Fisco carried 75,000. This desk recorded 73,000 at one vendor and 74,000 at another at the pre-US slot yesterday and deliberately did not publish either, on the ground that a three-thousand spread was not worth an article.
That judgement was right for the wrong reason. The spread between the four forecasts was three thousand jobs. The gap between the highest of them and the print was fifteen thousand. Whichever calendar you were reading, you were set up for a number in the low seventies and you got ninety, and the distance between the vendors was a fifth of the distance between all of them and reality.
This is the shape the file keeps finding. Arguing about which vendor has the right consensus is a way of spending attention on the smallest quantity on the page.
Four PCE fields, four misses, all in the same direction.
The August personal consumption expenditures release at 12:30 UTC undershot on every field anyone had a forecast for. Core PCE came in at plus 0.2 per cent on the month against an expected plus 0.3, and plus 3.0 per cent on the year against an expected plus 3.3. The headline deflator came in at plus 0.3 per cent on the month against plus 0.4, and plus 3.4 per cent on the year against plus 3.7. Four fields, four misses, every one of them soft.
That is a different kind of information from a single beat or miss, and it is worth separating. One field missing by three tenths is noise. Four correlated fields all missing on the same side is the forecasting community having been wrong about the same thing in the same direction, and it is why the October rate path repriced on the print rather than shrugging. If you run a rule that sizes off surprise magnitude in a single headline field, you measured three tenths yesterday. The release was bigger than that.
Eleven hours after the print, the box was still empty.
At roughly 23:20 UTC — eleven hours and five minutes after ADP printed — Investing.com’s own economic-calendar page for ADP Nonfarm Employment Change still showed the September row with a blank actual and a blank forecast, carrying only the 38,000 prior. The scheduled time on that row, 12:15, is correct. The number that arrived at it is not there.
And the same publisher is capable of the opposite. The Bank of Japan Tankan printed at 23:50 UTC and Investing.com’s calendar row for it carried an actual of 24.00 against a 25.00 forecast within three minutes, while Trading Economics still showed the June survey and described the September release as upcoming. One publisher, one evening: a number populated in minutes on one release and absent after eleven hours on another. That is not a slow feed. That is two different pipelines behind one calendar, and you cannot tell from the page which one you are looking at.
This desk published the same failure at FXStreet three days ago on the Dallas Fed survey, and it is the third instance in four days. The lesson has not changed and it is worth restating in the plainest terms: a calendar is a schedule, not a record. It will tell you reliably when something is due. It will not reliably tell you what happened, and a rule that reads the actual out of a calendar cell is a rule that will read a blank.
What this does not tell you
We have not established which of 58.8 and 55.8 is right. Three readers against one is a weight of evidence, not a verification, and we did not reach MNI’s own release page. If you need the Chicago number to be exact for a rule, go and buy it from the issuer.
We also cannot settle the release time. RTTNews puts the Chicago print at 10:15 a.m. Eastern, which is 14:15 UTC; this desk’s own week-ahead carried 13:45 UTC. That is a single source against our own note, thirty minutes apart, and we are recording it rather than publishing it as a finding.
Nothing here says anything about direction. Four soft PCE fields and two strong activity prints on the same afternoon is not a trade; it is a reason to expect the rate-path repricing to be argued about for several sessions. And the 90,000 ADP figure is one month of a series that has been revised heavily all year.
