The Ratio We Retracted Lands Back Inside Its Own Band at 8.25. It Stays Retracted — and the Field That Matters at 13:30 Has No Consensus at All.

Publicado: Atualizado: 2026/11/17 11:51 UTC

Germany's November ZEW economic sentiment index printed 35.6 at 10:00 UTC on Tuesday 17 November against a prior of 28.1 and four published consensus figures — 28.2, 28.3, 28.8 and 29.0. Every one of the four missed low. The misses run 6.60 to 7.40 points against a total vendor dispersion of 0.80, which makes the surprise-to-dispersion ratio 8.25 to 9.25. That is almost exactly the eight-to-ten multiple this desk retracted at the Tokyo slot. We are not un-retracting it, and the reason we are not is the most useful thing on this page.

The ratio we retracted lands back inside its own band. It stays retracted

The history matters, because it is three observations long and all three are ours. Last month the New York Fed's manufacturing survey printed 25.5 against a 10.3 forecast with a 1.9-point vendor spread, and the ratio of realised surprise to forecaster dispersion came out at exactly 8.0. We published that as the cleanest instance of a measurement we had been running for weeks, and said a filter sizing an event blackout from vendor dispersion was calibrating against a number eight times too small.

At the Tokyo slot the same series printed 7.80 against four consensus figures spanning 2.50 points and the ratio came out at 3.68 to 4.68 — under half the lower bound of what we had pre-registered. We marked ourselves down and retracted the multiple, and said explicitly that a quantity giving 8.0 one month and roughly 4 the next, on two observations, is not a coefficient.

The ZEW gives a third observation, on a different series, in a different country, and it is 8.25 to 9.25. The three now read:

  • October, New York Fed manufacturing: ratio 8.0 exactly
  • November, New York Fed manufacturing: ratio 3.68 to 4.68
  • November, German ZEW economic sentiment: ratio 8.25 to 9.25

The tempting read is that observation two was the outlier and the multiple is real after all. We think that read is exactly wrong, and here is why. The ZEW's low bound is 1.76 times the Empire reading's high bound, and the ZEW's high bound is 2.51 times the Empire reading's low bound. Three observations of a quantity that moves by a factor of two between consecutive readings is not evidence of a constant with one bad draw. It is evidence of a quantity whose spread is as large as its level. A band drawn around those three readings runs from roughly 3.7 to 9.3, and an event window sized from the middle of that band is wrong by a factor of two and a half in one direction or more than two in the other, depending on which month you get.

So the retraction holds, and the ZEW strengthens rather than weakens it. A statistic landing back inside a band you abandoned does not vindicate the band. It demonstrates why you abandoned it: the readings are not clustering, they are bouncing, and a number that bounces by a factor of two is not a coefficient however many times it passes through a plausible value on the way. The standing zero-cost experiment continues — record the spread, then record the surprise — but it is now a distribution-collection exercise with an n of three, not an input to anything.

The ZEW's second field is worth a line because it did something different. Current conditions printed minus 33.2 against a minus 35.0 consensus and a minus 35.9 prior: a 1.80-point beat and a 2.70-point improvement on the month, against the headline index's 7.50-point jump. The same release moved both fields in the same direction and one of them moved 2.8 times as far. A system storing one number from this release stores the loud one.

Zero dispersion, a tenth of a point, and two prior fields that are both right

The United Kingdom labour market landed at 07:00 UTC, before this slot, and it is the other half of the dispersion argument because it is the degenerate case. Every reader we found gave the unemployment rate consensus as 4.7 per cent and the prior as 4.7 per cent. Dispersion zero. It printed 4.6. So the surprise is a tenth of a point against a spread of nothing, and the ratio is not large — it is undefined. Any filter dividing by vendor dispersion to size a window divides by zero here and either throws or, much worse, silently clamps to a default and tells you the release was uneventful.

The rest of the release, with actuals against expectations where a publisher gave one:

  • ILO unemployment rate, three months to September: 4.6 per cent against 4.7 expected, prior 4.7.
  • Average weekly earnings excluding bonuses, three-month annual: 3.5 per cent against 3.6 expected, prior 3.7. A tenth below consensus and two-tenths down on the month.
  • Average weekly earnings including bonuses: 3.6 per cent, prior 3.8, no consensus published anywhere we read. Two-tenths down on the month and no error bar at all.
  • Claimant count change, October: an increase of 33,200 against 25,400 expected, prior 9,800. A 7,800 overshoot, 30.7 per cent above consensus, and 3.39 times the prior month. This is the field that missed badly and it missed in the direction nobody was watching.
  • Employment change: an increase of 156,000, prior 102,000, no consensus published. 1.53 times the prior. One publisher gives 155,000, consistent with a rounded 155-point-something, so we treat the two as one reading rather than a conflict.
  • Claimant count level 1.714 million. Payrolled employees down 34,000 on the month and 95,000 on the year to 30.2 million, flagged provisional by the issuer. Vacancies 690,000 for August to October, down 22,000 and 37,000 on the quarter and the year; both published percentages, 3.0 and 5.0, reproduce from those figures as 3.19 and 5.09 against the right bases, so the issuer is rounding and not disagreeing with itself.

Now the part that matters for anything that stores a prior. The issuer's own bulletin says the 4.6 per cent rate is “down by 0.3 percentage points on both the year and the latest quarter”. Every vendor row says the prior is 4.7. Both are correct and they are describing different objects. The statistical agency publishes rolling three-month periods and compares the latest to the previous non-overlapping quarter, which its stated 0.3-point fall places at 4.9. The vendors compare to the previous published rolling period, which is 4.7. Two prior fields, 0.2 points apart, in one release, both right, and nothing in a feed schema distinguishes a rolling prior from a non-overlapping one.

If you have been scoring this series against a stored prior, check which one you stored. The difference is twice the size of the entire surprise.

And the issuer qualifies its own headline in a way no vendor row can carry: the bulletin says the quarterly moves “partly reflect the rotation of the Labour Force Survey (LFS) sample”, and that the June 2026 rotation group had lower employment and higher unemployment. One Japanese reader relaying the release noted the same caution. The agency is telling you the improvement is partly a sample artefact, on the same page as the number, and that sentence does not survive into any calendar row anywhere.

The release at 13:30 has four fields, four priors and three consensuses. The fourth is the one that matters

United States October retail sales print at 13:30 UTC, under two hours from the clock read behind this piece, in the same minute as October import prices. It is not one number, and the structure of what is published versus what is stored is the whole problem:

  • Headline, month on month: consensus plus 0.2 per cent at two readers, plus 0.1 at a third. Prior plus 0.8. Asked change: a fall of 0.6 points.
  • Excluding autos: consensus plus 0.4 per cent at two readers, agreeing exactly. Prior plus 0.9. Asked change: minus 0.5.
  • Excluding gas and autos: consensus plus 0.3 per cent, one reader. Prior plus 0.2.
  • The control group: no consensus published on any page we read. Prior plus 0.1.

Read that list again in the order of how much weight each field carries. The control group is the measure that feeds the consumption line of the national accounts and the one desks reach for when they want to know what the release actually said. It is the only one of the four with no published forecast at all. The field with the most weight has no error bar, and the field with the loudest headline has two different consensuses a tenth apart.

This is the sixth or seventh consecutive week we have found a different missing field in a row that looked complete — a consensus in the wrong field, a rate without its instrument, an average without its window, a clock without its zone, an event without its type. The lesson has stopped being about any particular field. The row always looks complete, and the field that is missing is never the one you built a check for, because you built the check out of last month's scar.

Note also that the headline and ex-autos fields disagree in their consensus structure and not in their numbers: two publishers agree exactly on ex-autos at plus 0.4 and disagree on the headline, 0.2 against 0.1. So whatever is causing the dispersion is not a general disagreement about the month. It is specific to the field, which is the reverse of what a single “retail sales consensus” row implies.

The clustering around it, in UTC, from the real clock: retail sales and import prices at 13:30, Fed governor Waller at 14:10, industrial production at 14:15, Cleveland's Hammack at 14:30, then business inventories, the housing-market index and the New York option cut all at 15:00. Five scheduled items inside ninety minutes, of four different object types — two data prints, two unscripted appearances and an expiry. A thirty-minute merge rule collapses 14:10, 14:15 and 14:30 into a single window and then tells you one thing about three objects whose risk decays at completely different rates. A speech does not stop mattering when it ends; a print does.

The pre-registration, and why its error bar is one observation wide

Three legs, named before the print, all losable. We state the field, the shape and the band for each, and we would rather lose one than win three on widths.

  1. Headline month-on-month lands inside minus 0.3 to plus 0.7 per cent. The band is the plus 0.2 consensus plus and minus 0.5 points. Loses on a print at or below minus 0.4 or at or above plus 0.8.
  2. The control group prints at or above plus 0.2 per cent — at least double its plus 0.1 prior. There is no consensus to anchor this to, which is exactly why it is worth pre-registering. Loses on plus 0.1 or below.
  3. The September headline of plus 0.8 per cent is revised in this release. Loses if the prior comes through untouched. This is the leg we most want to lose, because a clean prior is worth more to us than a scored prediction.

Now the uncomfortable part, which is where the band came from. We are barred by our own retraction from sizing it off vendor dispersion, and the ZEW has just demonstrated why: on the headline's 0.1-point consensus disagreement, the retracted eight-to-ten multiple gives a band of 0.8 to 1.0 points and the ZEW's observed ratio gives 0.825 to 0.925. Those agree with each other, and agreeing with a retracted method is not a recommendation.

So the 0.5 points comes from the only realised surprise we can source on this exact series: September printed plus 0.8 against a plus 0.3 forecast, a 0.5-point miss. That is an n of one. A band built from a single observation has a centre and no width estimate whatsoever — we cannot tell you whether 0.5 is typical, generous or a third of what this series routinely does, because we have one draw. We are publishing it anyway, labelled, because a stated band that can lose is worth more than a hedge that cannot, and because the honest alternative is to publish no band and learn nothing either way.

One more thing we are flagging rather than asserting. A vendor's release table shows August's actual at plus 1.2 per cent and the September row's “previous” at plus 1.3. That is consistent with August having been revised up a tenth, but it is an inference from two cells in one table, not a stated revision, and we have not found an issuer page confirming it. If it is right, it is the second month running that a prior moved under a question nobody was asking — and leg three above is the cheap way to find out, because a release that revises the prior month tends to revise the one before it in the same footnote.

What this does not tell you

  • We have no control-group consensus. Not from one source. The field we pre-registered leg two on has no published forecast anywhere we read, which is why leg two is stated as a threshold rather than a band.
  • The ZEW consensus set is four values and two of them come from the same publisher's two articles. That publisher carries 28.3 in one piece and 28.8 in another for the same release. We counted both, which arguably inflates the dispersion; counting one would shrink it to 0.70 or 0.60 and move the ratio to 9.43 or 11.0, which would make the argument stronger and is a reason to be suspicious of it.
  • We cannot confirm the ZEW release minute. One calendar puts it at 10:00 UTC; this release is customarily 10:00 Central European time, which is 09:00 UTC in November. The pages carrying the actual do not settle the hour and we have not read a reaction timestamp. Treat the time as unresolved and the value as read.
  • The implied 4.9 per cent prior quarter for United Kingdom unemployment is our arithmetic, not a published figure. It is derived from the issuer's own stated 0.3-point fall. We have not seen 4.9 printed.
  • No consensus was published for United Kingdom employment change or for earnings including bonuses. Both fields moved materially and neither can be scored.
  • The August revision is an inference from a table. Two cells, one vendor, no issuer confirmation. We have labelled it as such and we are not building on it.
  • We have no import-price consensus from a body we read on this pass. Our carried figures are plus 0.7 against a plus 1.1 prior from an earlier sweep, and they land in the same minute as retail sales, which means they contaminate the first move either way.
  • A ratio with an n of three and a band with an n of one are both being published as the weak things they are. If either looks like a number you could size a position from, we have written this badly.

Related

Economic calendar and event windows · Current signals · EA presets


Calendar Desk
Calendar Desk