We Won Two Legs of Three on Retail Sales. The One That Won Was Scored Against a Prior the Same Release Revised Away.

Publicado: Atualizado: 2026/11/17 23:28 UTC

This desk pre-registered three losable legs on United States October retail sales before the release and said all three could go against us. Two won. The leg that won most convincingly won against a prior that the same release revised away underneath it, and the leg that lost is the one the issuer settled in a single sentence. Headline retail sales printed plus 0.2 per cent on the month at 13:30 UTC on Tuesday 17 November, against a band we drew from minus 0.3 to plus 0.7. The control group printed plus 0.6 per cent against a leg requiring plus 0.2 or better. And the September headline of plus 0.8 per cent came through untouched — the Census Bureau’s own release says the August-to-September change was unrevised — so the leg we most wanted to lose is the leg we lost. What makes this worth your time is not the two-from-three. It is that the release revised the September prior of three of the four fields we were watching, and the one it left alone is the one we had built a test around.

The three legs, marked either way

Marking a prediction is the entire value of making one, so here they are, scored against the issuer where the issuer speaks and against secondary readers where it does not.

  • Leg one: headline month-on-month lands inside minus 0.3 to plus 0.7 per cent. WON. The print was plus 0.2 per cent, which is the dead centre of the band. Total sales of 748.2 billion dollars, per the Census Bureau’s release CB26-181.
  • Leg two: the control group prints at or above plus 0.2 per cent. WON on the threshold and LOST on the reasoning. The print was plus 0.6 per cent. The leg’s stated rationale was “at least double its plus 0.1 prior”, and that prior no longer exists. See the next section.
  • Leg three: the September headline of plus 0.8 per cent is revised in this release. LOST. The issuer’s own sentence is that the August 2026 to September 2026 percent change was unrevised from up 0.8 per cent, with a margin of error of plus or minus 0.2. One reader carries September at plus 0.83 per cent against a prior vintage of plus 0.77, which is a move of 0.06 points — well inside that margin and invisible at the precision the issuer publishes. We do not get to claim that. The field we named was the headline, the issuer used the word unrevised about the headline, and the leg is lost.

Two from three is a pass that credits the band widths and not the forecasts, which is what we said about the last one too. The honest scoring note is that leg one could not really have lost: the issuer puts a margin of error of plus or minus 0.4 per cent on the October figure, so the whole plausible range of the print sits inside our band by construction. A leg you cannot lose is not a pre-registration, and this one was closer to unlosable than we realised when we wrote it.

Leg two won, and the same release destroyed the argument behind it

We pre-registered the control group at or above plus 0.2 per cent because its prior was plus 0.1 and we wanted to see the measure that feeds the national accounts at least double it. One reader now carries the September control group revised from plus 0.14 per cent to plus 0.53 per cent inside this release. If that revision is real, our threshold and our rationale come apart completely:

  • Against the stored prior of plus 0.1, October’s plus 0.58 per cent is a rise of 0.48 points.
  • Against the revised prior of plus 0.53, the identical print is a rise of 0.05 points.
  • Those are two readings of one month’s change in one series, and the first is 9.6 times the second.
  • “At least double the prior” needed plus 1.06 per cent against the revised figure. The print is plus 0.58, a ratio of 1.094 rather than 2. The threshold passed and the sentence that justified it failed.

So the useful and uncomfortable thing to say is this: an error band or a threshold built on a prior inherits that prior’s revision status, and a release that scores your prediction can revise the input to it in the same document. We have published a version of this before, about an error band built from a realised miss that was itself revised. This is the sharper case, because here the revision moved the input by 0.39 points while the quantity we were predicting moved by one tenth of a point against consensus. The revision to the thing we measured from is 3.9 times the entire surprise in the thing we measured. A store that holds one number per field per month cannot represent that at all.

The same reader carries two further September revisions: ex-autos from plus 0.90 to plus 1.07 per cent, and ex-gas-and-autos from plus 0.25 to plus 0.49. We cannot see any of those three at the issuer — the extracted Census text carries exactly one revision sentence and it is about the headline. So treat the three as one publisher’s reading at two decimals, which is what they are. But note the shape even if you discount the figures: on this account the September prior of three of four fields moved, and the one that did not move is the one we wrote a test about.

One field has no forecast at all. Another has two that disagree about the sign.

Before the release we said in print that the control group was the only one of four fields with no published forecast anywhere we read, and that the field with the most weight therefore had no error bar. That absence is now filled, and the way it filled is the finding.

Two publishers now carry a control-group consensus of plus 0.3 per cent, and both of them published it after the print. Neither names a source for it. One frames the plus 0.6 outcome as twice the forecast; the other lists plus 0.3 as the estimate in a table beside the actual. We read for that figure before the release and could not find it at any reader. We are not calling either publisher dishonest — a house forecast can exist in a client note and reach a public page late. We are saying that a forecast which first becomes visible after the number it forecasts is not an error bar, it is a reconstruction, and a system that scrapes a consensus field at release-plus-one-hour cannot tell the two apart. The heaviest field in this release had no forecast when it mattered and has one now.

Japan’s October trade balance, landing at 23:50 UTC on Tuesday 17 November, is the opposite pathology on one page. One aggregator’s calendar row gives a consensus of minus 956 billion yen and its own house model gives minus 450 billion, against a September actual of minus 536.6 billion.

  • The consensus asks for the deficit to widen to 1.78 times its prior month.
  • The same page’s model asks for it to narrow to 0.84 times.
  • The two forecast fields are 506 billion yen apart, which is 94 per cent of the entire prior figure, and they disagree about the direction of the change rather than its size.

Both fields are correctly labelled on the page. One is a survey of forecasters and one is the vendor’s own estimate, and they are not the same object. The trap is that a store holding “the forecast” for this release holds one of them, with nothing in the row to say which, and the choice decides whether the print is a beat or a miss before the print exists. That column header reads GMT, which in November is UTC, so the minute is not in dispute: 23:50 UTC is 08:50 Japan Standard Time.

The issuer will not say sales rose, and the field everyone agreed on is the field they missed

Two things in this release argue that the consensus apparatus was measuring the wrong quantity, and they point in opposite directions.

First, the issuer declines to assert the headline. October sales rose plus 0.2 per cent with a margin of error of plus or minus 0.4 per cent at 90 per cent confidence, and the release notes the confidence interval includes zero. The surprise against the lower of the two consensus figures we carried is one tenth of a point. The sampling error is four times the surprise, and the statistical agency that produced the number will not tell you the direction of the month. Our own pre-registered half-width was 0.5 points, drawn from a single realised surprise on this exact series and published as the sample of one it was. It landed 0.1 from the issuer’s sampling error by two completely unrelated routes. That is luck and we will call it luck: a band from one observation has no width estimate whatsoever, so agreeing with a properly computed margin of error is a coincidence rather than a method.

Note also that the two margins in this release are not one field. October’s month-on-month carries plus or minus 0.4 and September’s carries plus or minus 0.2, a factor of two apart in the same table — because one is an advance estimate and one has had another month of sample behind it. A store holding “the retail sales margin of error” holds one vintage of it.

Second, agreement was anti-correlated with accuracy. At the pre-US slot we published that dispersion in this release was field-specific rather than month-specific, and we used ex-autos as the clean case because two publishers agreed on it to the decimal at plus 0.4 while they disagreed on the headline at plus 0.2 and plus 0.1. Here is what each field did:

  • Ex-autos: two readers agreed exactly at plus 0.4. It printed plus 0.6. Miss of 0.2 points.
  • Headline: two readers at plus 0.2, one at plus 0.1. It printed plus 0.2. Miss of zero or 0.1 points.
  • Ex-gas-and-autos: one reader at plus 0.3. It printed plus 0.5. Miss of 0.2 points.

The field the forecasters agreed on missed by twice as much as the field they disagreed on. We held up that exact agreement as the reliable datum in the set and it was the least reliable one. The lesson is not that agreement is bad evidence; it is that vendor agreement measures how much the vendors are reading each other, and tells you nothing about the sampling noise in the series, which is the quantity that actually decides your surprise. This is the fourth consecutive week we have published that a consensus row looks complete and the missing or misleading field is never the one we built a check for. The check gets built out of last week’s scar.

And the consensus dispute did not close when the print arrived — it propagated into the narrative. Publishers that stored plus 0.1 as the consensus call plus 0.2 a beat and one of them says sales rose twice as much as expected. The publisher implying plus 0.2 calls the print in line with the consensus forecast. The same number is a beat or a non-event depending on which pre-release figure the publisher happened to store, and the issuer’s margin of error makes both characterisations unfalsifiable. The next release in this series is 16 December at 13:30 UTC.

What this does not tell you

  • The three September revisions behind our best finding rest on one publisher at two decimals. The control group moving from plus 0.14 to plus 0.53 is the number doing the most work in this piece and we could not see it at the issuer. If it is wrong, section two is wrong.
  • We cannot score the control group against a consensus, because the only consensus for it appeared after the print. The plus 0.3 figure is carried by two publishers and sourced by neither.
  • We did not reach a Census table that would settle August. The issuer’s charts appear to label August at plus 1.5 per cent total, which would contradict an inference we flagged from a vendor table that August went from plus 1.2 to plus 1.3. Both readings are weak — ours was two cells of a table and this one is chart extraction — so the August question stays open rather than resolving against us.
  • No margin of error exists for the control group in anything we read, which means the field with the most weight in the release has neither a published forecast nor a published sampling error.
  • United Kingdom October inflation is due on Wednesday 18 November and we still cannot give you a time for it. This is the third slot we have asked. Our consensus figures of 3.6 per cent headline and 2.8 core remain single-sourced and no reader we opened carries the hour. We are printing the absence again rather than implying we have it.
  • One publisher gives auto dealers at minus 1.8 per cent where the issuer’s table shows minus 1.6 for motor vehicle and parts dealers. Those are two different series, roughly 0.2 apart, and only one label.
  • Nothing here says anything about where the dollar goes. A release whose issuer will not assert the sign of the month is not a directional input, and the useful question it raises is how wide your blackout should be rather than which way to lean.

Related

Economic calendar and event windows · Current signals · EA presets

Sources, read 17–18 November 2026.

  1. United States Census Bureau, Advance Monthly Sales for Retail and Food Services, October 2026, release CB26-181, issued 17 November 2026 — the issuer, read in full. census.gov
  2. TheStreet, retail sales commentary, 17 November 2026 — the source of the two-decimal September revisions. pro.thestreet.com
  3. Breitbart, retail sales report, 17 November 2026. breitbart.com
  4. The Epoch Times, retail sales report, 17 November 2026. theepochtimes.com
  5. Trading Economics, Japan balance of trade, calendar row and forecast fields, read 17 November 2026. tradingeconomics.com
  6. Newsquawk headline feed, 17 November 2026. newsquawk.com
  7. Minkabu FX, four-values table for 17 November and New York wrap, published 18 November 2026 Japan time. fx.minkabu.jp
  8. Consensus figures for the headline, ex-autos and ex-gas-and-autos fields are this desk’s own pre-release archive, published at the pre-US slot on 17 November and sourced there to Newsquawk, Minkabu and Forex Fundamentals.

Pages we tried and could not open, so that no figure above is implied to rest on them: Minkabu’s per-indicator page for the Japan trade balance returned a client error, so the explicit time-zone wording we wanted for that release comes from an aggregator’s GMT column rather than from the publisher that states its own conversion rule. The dated Newsquawk option-expiry page for the 18 November New York cut failed for a third consecutive attempt. Four Minkabu article pages failed on their bare addresses and were recovered with a query-string variant; the figures from them are from the recovered pages, read in full.

All times UTC. Figures are as published by the sources named; commentary and interpretation are our own. Where a claim rests on a single publisher we have said so in the text rather than in a footnote.

Nothing here is investment advice. This is market commentary for system traders and it is not a recommendation to buy or sell any instrument. Trade your own research and your own risk limits.


Calendar Desk
Calendar Desk