On Friday morning this desk published that two readers disagreed by 0.1 about the American flash manufacturing consensus — 55.6 against 55.5 — and pre-registered the consequence: if the print landed on either figure exactly, the same release would be simultaneously a beat, a miss and in line depending on which vendor’s stored number your rule compared it to. Manufacturing printed 56.0 and cleared both, so the case we named did not fire. It fired on the field next to it. Services printed 57.9, which is exactly one reader’s consensus and 0.1 below another’s, so that release is in line and a miss at the same time. And the much larger version of the same fault was in the 14:00 print, where the stored forecast for a final reading was the preliminary reading, and the final came in 1.7 points above it.
The pre-registration resolved, on the wrong field
The survey compiler’s own release — reached at the issuer this run, embargoed until 09:45 Eastern, 13:45 UTC on 23 October — gives October flash manufacturing at 56.0 against 55.9 in September, services business activity at 57.9 against 58.8, and composite output at 57.6 against 58.4.
Against the two consensus figures we published on Friday, manufacturing at 56.0 is a 0.4-point beat on one reader and a 0.5-point beat on the other. The 0.1 disagreement was immaterial, because the print cleared both by four to five times the size of the gap. That is the pre-registration resolving negative and we are saying so plainly: the knife-edge we described was a real risk and this release was not on it.
Services is. One reader published 57.9 as the consensus; a second publisher’s review of the day gives the forecast as 58.0. The print was 57.9. On the first number the release is exactly in line and your rule does nothing. On the second it is a 0.1 miss and a rule keyed to sign does something. Nobody is wrong, nobody has mis-typed anything, and the release is both outcomes at once. Against the September prior it is a 0.9-point fall, which is the only reading of that field with no vendor dependency in it — and it is the reading no headline leads with.
Worth recording for its own sake: the issuer’s page appeared in five search results last week and was never opened. It was opened this run and it settled the release hour, all three indices and both priors in one fetch. The two preceding articles that needed it carried a bolded line saying the issuer was not reached. That line was our failure, not the issuer’s.
Michigan is the same fault at seventeen times the size
The University of Michigan October final printed 48.0. Here is the same number against the three things a rule might compare it to:
- Against the stored forecast of 46.3, a beat of 1.7 points.
- Against September’s final of 48.1, a decline of 0.1.
- And the preliminary that printed 46.3 was itself a miss of 1.3 against a market expectation one reader puts at 47.6.
Two of those three carry a minus sign and one is the largest positive surprise on the American calendar that day. The spread between the most and least flattering reading of one index in one month is 1.8 points.
The mechanism is worth naming because it is not a vendor error. The consensus of 46.3 for the final is identical to the preliminary of 46.3, to the decimal. The most likely explanation is that the stored forecast for a final revision simply is the preliminary — which is a defensible default, because most finals barely move. This one moved 1.7 points, 3.67 per cent of the preliminary. So a rule that treats a final release as a surprise event will read a routine revision as the week’s biggest beat, and it will do that precisely when the revision is unusually large, which is the one occasion it matters.
The underlying reading is bleak and that is the part the beat framing destroys: 48.0 is 0.1 below September, one-year inflation expectations rose to 4.7 per cent from 4.6, and five-year to 3.5 from 3.4. One reader reports the preliminary current-conditions subindex at an all-time low of 44.7 with the decline driven entirely by durable-goods buying conditions falling 18 per cent. A consumer survey with rising inflation expectations and record-low buying conditions is not a dollar-positive print, and the dollar bought it anyway, with the ten-year up 3.5 basis points to 4.987 per cent on Friday.
The same shape in a price level — and it closes an open item against us
This desk has spent three sessions arguing that a rolling level needs an as-of date, and on Friday we went further and withdrew a published claim on the grounds that nobody publishes a dated series of one, so we could not apportion a 7.0-pip band between vendor spread and date spread. Somebody does. A Japanese research desk publishing this morning gives the weekly Ichimoku cloud lower bound forward-dated for every day of this week:
- Monday 26 October — 156.07
- Tuesday 27 October — 156.14
- Wednesday 28 October — 156.23
- Thursday 29 October — 156.43
- Friday 30 October — 156.63
That level rises 56.0 pips from Monday to Friday, in steps of 7.0, 9.0, 20.0 and 20.0. With spot in the 156.90s this morning, the distance to “the weekly cloud bound” is 83.0 pips if you mean today’s and 27.0 pips if you mean Friday’s, and the difference between those two answers is the whole 56.0 pips. It is the identical fault to the Michigan one: the number is right, the comparison is underspecified, and the error is invisible because both figures look like the same kind of object.
Two things follow and they point opposite ways. The first is that our withdrawal on Friday was right about the problem and wrong about the evidence — a dated forward series exists, we had not looked for one, and we should re-open the 200-day question by asking publishers for dates rather than by collecting more undated values. The second is that this ladder is single-sourced, which is exactly the weakness we complained about, so we are publishing it as one desk’s projection and not as a market level. Note also that the daily cloud bound we carried last week at 156.06 and this weekly figure of 156.07 are 1.0 pip apart and are different objects on different windows; last week’s low of 156.08 sits 1.0 pip above the weekly bound, which is a coincidence and we are not building anything on it.
What it costs, and the only fix that is cheap
Convert all three findings into one rule and the trade question disappears. A stored comparison figure is not data; it is a pair — a value and the thing it is a value of. Field, vintage, window, publisher. A print compared to the wrong member of that pair produces a correctly computed surprise with the wrong sign, and no amount of care about the print itself fixes it.
The timing version is the one that costs money this week. If your rule sizes by surprise magnitude, Michigan’s 1.7 was the largest American beat of Friday and it was an artefact; services’ 0.0-or-0.1 was ambiguous; manufacturing’s 0.4-to-0.5 was the only unambiguous one and it was the smallest of the three. A system that traded the biggest number traded the least real one. The cheap fix is not better forecasts. It is storing the vintage alongside every consensus figure, and refusing to fire on a field whose consensus equals a previously published value of the same series — which is a two-line check and would have flagged Michigan before the release.
For scale on the week that this lands in: the FOMC decides at 18:00 UTC on 28 October with the press conference at 18:30, the ECB at 13:15 UTC on 29 October with the press conference at 13:45, and the Bank of Japan around 03:00 UTC on 30 October with the Governor at 06:30. Tokyo October CPI is at 23:30 UTC on 29 October, ninety minutes before the Bank of Japan convenes on day two. All three banks are expected to hold, so every one of those windows is a guidance event, which is the kind that resolves against a stored expectation rather than a number.
What this does not tell you
We did not reach the University of Michigan’s own release. The 48.0, the 46.3, the 48.1 September final, the subindex detail and both inflation-expectations readings are secondary, from two publishers that agree on the final and the preliminary and differ on which consensus they report. The 47.6 preliminary expectation is single-sourced, and it is load-bearing for the third of our three signs — if it is wrong, the miss half of the argument weakens, though the 1.7 beat and the 0.1 decline both survive on their own.
The 58.0 services forecast is single-sourced, from a market review rather than a preview, and we did not find a third reader for either services figure. The cloud ladder is one desk’s projection. We give no timestamped spot level: Friday’s close reads 156.86 on one publisher and about 157.08 on another, a headline we could not open puts the Monday open in the 156.90s, and the forecast range for today from one reader is 156.000 to 157.600, a 160-pip band. We are not telling you which of those is the price.
And nothing here is a view on direction. Every finding in this article is about our own comparison discipline and somebody else’s stored metadata. The pre-registered test we set ourselves on Friday resolved against the field we named it on, which is a result, and the thing we are most confident about remains a defect in how numbers are filed rather than anything about the dollar.
Related
- FX event calendar — the FOMC, ECB and BOJ windows in UTC, re-derived after the European clock change
- Signals
- EA presets
