At 06:18 UTC this morning this desk published four live consensus disagreements on today’s PMI board and said, in print, that we would not publish a fifth because the figure behind it was one house’s own forecast. Three of the four have now resolved. The two widest disagreements turned out to be irrelevant, and the narrowest one — a single tenth of a point on euro area manufacturing — was the only one that changed what a surprise-scaled system would have done. The size of a vendor disagreement told you nothing about whether it mattered.
52.7, and three different verdicts on the same number
Euro area flash manufacturing printed 52.7 at 08:00 UTC. Nobody disputes that. What the print means depends entirely on which feed you keyed your consensus to, and this morning there were three candidates in circulation.
Against Forex Factory’s calendar, which carried 52.6, the print is a beat of one tenth. Against the data table on FXStreet’s own release page, which reads “Consensus: 52.7”, the print is exactly in line and the surprise is zero. Against the 53.3 we declined to publish, the print is a miss of six tenths.
One release, three feeds, and a surprise that is positive, zero or negative depending on which one your calendar happened to scrape. A system that scales position size by surprise magnitude would have sized long, stood aside, and sized short — on identical data.
The narrowest disagreement on the board — one tenth — was the decisive one. That is the finding, and it is not the one we expected.
We drew the band one tenth too narrow, and the print landed in the tenth we excluded
This morning’s article said that had 52.6 and 53.3 been published side by side as a split, the gap would have been wide enough to flip the sign of the surprise “for any print between 52.8 and 53.2”.
The print was 52.7. On our own stated band, that is outside it, and the illustration should have failed.
It did not fail. 52.7 is above 52.6, so it is a beat. 52.7 is below 53.3, so it is a miss. The sign flipped at exactly the number we excluded. The band should have started at 52.7, not 52.8; we set the lower bound one tenth too high and in doing so understated our own case by the only tenth that was going to matter.
We would rather record that as an arithmetic error in our favour than quietly widen the band after the fact. The mechanism held. The boundary we drew around it was wrong, and it was wrong in the direction that made us look more cautious than we needed to be.
The two wide splits were swamped, and that is not luck
Euro area services was the widest disagreement on the board — 51.4 against 51.7, three tenths. It printed 53.0. That is a beat of 1.6 or 1.3 depending on the feed, and no sizing rule in use anywhere treats those two differently. The composite went to 53.1 against 51.5, a beat of 1.6.
UK manufacturing was split 51.5 against 51.4 and printed 52.0 — a beat of half a point either way. Immaterial.
There is a mechanical reason the wide splits did not bite and the narrow one did, and it is worth stating because it generalises. A vendor disagreement matters when the print lands inside the gap, not when the gap is large. A three-tenths gap that the print clears by a point and a half is noise. A one-tenth gap that the print lands on top of decides whether the release was an event at all. You cannot know in advance which you are getting, which is precisely why the gap width is not the thing to screen on.
The figure we said had vanished came back
At 06:18 UTC we wrote that the euro area services consensus of 51.7, which this desk had published at 23:15 the previous night, was “not present on any reader we checked” and that we could not tell whether it had been an error, a stale render or a real revision.
It is present. FXStreet’s release page, published 08:04 UTC, gives the services consensus as 51.7 in its own prose. The figure did not vanish; it was absent from the surfaces we read at 06:10 and back on a different surface two hours later.
That correction is due and we are making it plainly. It also downgrades the observation from “a forecast disappeared overnight” to something duller and more useful: the same publisher serves different consensus figures on its calendar page and its release page, and which one you get depends on when and where you look.
The same release page demonstrates it twice over. Its prose says manufacturing “was expected to drop to 52.6”. Its data table, in the same document, says “Consensus: 52.7”. Those are both FXStreet, both about the same release, both published at 08:04.
A fifth split opened in the maturity that was agreed this morning
US flash PMIs print at 13:45 UTC, after this slot, so the US services split — 55.8 against 56.0 — is still open and we are not marking it here.
But US manufacturing, which both of our readers put at 53.6 this morning with no disagreement at all, is no longer agreed. Forex Factory still carries 53.6. A broker desk note timestamped 11:19 UTC carries 53.5. That is a one-tenth gap that opened during the session, in the one line on today’s board that had nothing wrong with it eight hours ago.
We are recording it rather than building on it, because one tenth on one new reader is not yet a claim. But the pattern this article is about says a one-tenth gap is exactly the kind you should not dismiss on width.
What this does not tell you
It does not tell you which feed is right. We have no privileged access to the survey of economists behind any of these numbers, and it is entirely possible that 52.6 and 52.7 are both faithful renderings of different polls taken at different times. The argument here is not that somebody is wrong; it is that your system is silently picking one.
It does not tell you that one tenth always matters. It mattered today because the print landed on it. On most days it will not, and reading this as “screen for narrow gaps” would be the wrong lesson — the right one is that gap width is not a filter at all.
It does not settle the US split, which resolves after this was written. And the 53.5 we cite for US manufacturing is a single reader on a broker page we have not used before; treat it as one desk’s reading, not as a second source.
We also have no explanation for why the 51.7 was absent at 06:10 and present at 08:04. We looked. We are not manufacturing a cause.
What to key your sizing to
The operational answer is unglamorous. If your position size is a function of surprise magnitude, the consensus is an input to your system and inputs get named, versioned and logged. Record which feed supplied the number, and the timestamp you read it at, in the same row as the trade. Today that one field is the difference between a long, a stand-aside and a short on euro area manufacturing.
The weaker but cheaper alternative is to stop scaling by surprise on releases where your feeds disagree at all, and take a fixed, small size instead. That costs you the good days. It also means a tenth of a point in somebody’s database cannot invert your position.
