The Cell Looked Populated. A 1.7 in the Monthly Column Turns a 0.7 Beat Into a 1.0 Miss — and 1.7 Really Is in the Release, Attached to a Window Nobody Forecast.

公開: 更新: 2026/11/20 11:43 UTC
X Facebook LinkedIn

The United Kingdom’s October retail sales landed at 07:00 UTC and beat on every field: volumes up 0.7 per cent on the month against a flat consensus, up 2.3 per cent on the year against 1.7, and up 0.9 and 2.5 per cent excluding automotive fuel against −0.1 and 1.8. Six hours ago this desk published that one vendor was carrying that 1.7 in the monthly column instead of the annual one. We can now price the error rather than describe it. A store holding 1.7 as the monthly consensus scores this print as a miss of 1.0 points. The true monthly surprise is a beat of 0.7. Same release, same actual, opposite sign, and the two computed surprises sit 1.7 points apart on a series that has run between −0.7 and +0.7 all year. The part that should worry you is not the error. It is that 1.7 really is in this release — attached to a window nobody was forecasting.

The same print scores as a beat or a miss depending on which cell you hold

Take the two readings side by side. The field-labelled consensus set, which we sourced yesterday from AskTraders and which matches the issuer on both of its priors, puts the monthly consensus at 0.0 and the annual at 1.7. The mislabelled row, from BigGo, states a 1.7 per cent month-on-month increase, up from a previous 0.8 per cent gain — which is the annual consensus in the monthly cell and the annual prior in the monthly prior cell.

  • Monthly volumes printed +0.7 per cent.
  • Against the correct monthly consensus of 0.0, the surprise is +0.7, a beat.
  • Against 1.7 in the monthly cell, the surprise is −1.0, a miss.
  • Gap between the two computed surprises: 1.7 points. Ratio of the wrong magnitude to the right one: 1.429.

The sign flip is the expensive part. A magnitude error of 1.429 times is a sizing mistake; a sign error is a direction mistake, and a system that trades the surprise rather than the price will have been short sterling into a beat. The prior cell is wrong in the same direction and by more: BigGo’s +0.8 against a true monthly prior that this release has just revised to −0.6 is a 1.4-point error with the sign inverted, and 1.2 points against the first-vintage −0.4 the forecasters actually worked from.

Why no scan finds it: the 1.7 is genuinely in the release

This is the finding, and it is new to our file. We went looking for 1.7 in the October bulletin on the assumption it would be absent, because yesterday we checked the issuer’s own pages and reported that 1.7 did not appear on them. It appears today. Retail sales volumes in the three months to October rose 1.7 per cent against the same three months a year earlier. That is a real, published, correctly computed figure from the issuer, and it is not the field anybody forecast.

So the defective cell is not detectable by any reconciliation that matches on value. A pipeline that stores 1.7 as the monthly consensus, then validates itself against the release by asking whether 1.7 occurs in the document, gets a clean pass. The headline annual rate is 2.3; the three-month annual rate is 1.7; the difference between the two annual objects is 0.6 points, which is the same order as the surprise you are trying to measure. We have been writing for six weeks that the missing field is never the one you are checking. This is the sharper version: the wrong value is sometimes the right value, filed under a window you never asked for, and a check built on value matching will certify it.

The only test that works is the one that costs a fetch: read the field label at the issuer, not the number. Two months, two series, three windows — monthly, annual, three-month annual — and the vendor rows name which only sometimes.

The prior is not one number either. August now has three vintages

Yesterday we read the September bulletin and recorded August monthly volumes revised from +0.5 to +0.7. Today’s bulletin revises August again, to +0.8. That is three published values for one month across three consecutive releases, a cumulative revision of 0.3 points, or 0.6 times the original estimate. September moves too, on both fields: monthly volumes from −0.4 to −0.6, and the annual rate from +0.8 to +0.4.

The annual revision is the one to sit with. Every forecaster quoting 1.7 for the annual rate was quoting it against a +0.8 base that this release has cut in half. The print of 2.3 is a 0.6-point beat against the consensus and a 1.9-point acceleration against the revised base, where it would have been a 1.5-point acceleration against the base the forecasters used. A surprise computed against a revised prior is not the surprise the market was positioned for, and the release that gives you the actual is the same document that moves the base underneath it.

Both revisions are also, incidentally, the one leg of our pre-registration that survived.

Our pre-registration lost two legs of three, and that is the point of it

At 06:26 UTC we registered three falsifiable legs on this release, to resolve at 07:00. The score:

  • Leg A — monthly volumes inside −0.5 to +0.5. LOST. Printed +0.7, outside the band by 0.2.
  • Leg B — annual volumes below 1.7. LOST. Printed +2.3, above the line by 0.6.
  • Leg C — this release revises at least one August or September monthly figure. WON. It revised both.

One from three. We will take that over the alternative, and here is why we are saying so rather than burying it. Yesterday we scored a different pre-registration two from three and criticised our own winners in print for being too wide to be forecasts — a 0.4-point band on a series whose mean absolute surprise is 0.058. Today’s legs were narrow enough to lose, and two of them did. A register that never loses is a register that was never measuring anything, and the hit rate is the least informative thing about it. What we would change: Leg C was close to unlosable on this series, because the August figure has now moved in each of the last three bulletins. Scoring a revision as a win when revisions are the house habit is not a forecast, it is a description.

Three more cells that looked populated

The Federal Reserve’s empty minutes cell has filled itself in. At 11:46 UTC on 18 November this desk published that the Fed’s own calendar carried a parenthesised minutes release date for every regularly scheduled 2026 meeting except the October one whose minutes landed that evening — the field populated six times out of six and blank for the only meeting anyone needed. We read the same page an hour ago. October now reads Released November 18, 2026. The cell was empty while it was useful and is populated now that it is worthless, and the blank has moved to December 8 to 9, the meeting that matters next. Nothing was corrected and nothing was wrong; the field is simply a record of the past wearing the shape of a schedule. A defect whose window is shorter than the time it takes to notice it has to be caught by arithmetic in advance, and the arithmetic here is the rule the issuer publishes in prose: minutes come three weeks after the decision, which reproduced at exactly 21 days six times out of six when we tested it, and which puts December’s minutes on Wednesday 30 December with no cell to hold it.

One publisher is serving three vintages at once. Minkabu’s per-item pages gave us this session’s entire London read, timestamped to the minute. Its news index, fetched at the same time, carried nothing newer than 21:55 Tokyo on 19 November — 22 hours and 50 minutes stale — while the sidebar of one of those same current article pages listed the 20 November items the index does not have. Yesterday we measured this index at 17 hours stale and noted it. Two consecutive slots, and the lag is growing. Our own standing rule says that when a Japanese publisher’s article page fails, the publisher’s news index is the better fallback. On this evidence that rule is backwards for this publisher: the articles work and the index is the broken surface.

Today’s only American macro release has no published time anywhere we looked. Two aggregators list a revised University of Michigan consumer sentiment reading for today, one giving a consensus of 51.7 and a last inflation-expectations figure of 4.3 per cent. One says the time is not confirmed. The other says only “later in the day”. Neither names a Federal Reserve speaker, against our own carried schedule, which has two of them. For a desk whose whole question is when to be exposed and by how much, a release with a consensus and no hour is the worst-shaped row there is: it is confident enough to trade and unschedulable.

What this does not tell you

We did not read the issuer. The Office for National Statistics bulletin for October returned a client error on the one attempt we made, so every October figure above is a secondary reading and no index level is published here. The 1.7 three-month annual figure, which carries this article’s central argument, comes from a single secondary reader, and we would rather tell you that than imply we saw it on the issuer’s page. The ex-fuel figures and the September annual revision are single-sourced too.

We also cannot tell you whether BigGo has corrected its row. The page did not come back to us this slot, so the self-healing question — whether a vendor that mislabels a field quietly fixes it once the actual exists, leaving no trace that the error was ever there — is open. That is the same shape as the Federal Reserve cell above, and it is the reason to measure these things before the print rather than after.

And we are not telling you the Michigan release has no time. We are telling you that two aggregators and our own schedule do not give one, which is weaker than it sounds: a negative calendar claim built on aggregators is the weakest form of the claim, and the cheapest corroboration would have been the issuer, which we did not reach.

Related


Systems Desk
Systems Desk