For two slots this desk argued about whether August American retail sales were 1.1 or 1.2 per cent, and opened a numbered item to settle it at the issuer. The issuer has now settled it, and the answer is 1.3. Worse for the argument: the Census Bureau published the first print with a stated margin of error of plus or minus 0.4 percentage points. The gap we were arguing about was 0.1. We spent two slots disputing a quarter of one error bar, and both figures in the dispute were inside it from the moment they were published.
The dispute was not close. It was malformed, and the issuer says so in its own footnote.
Release CB26-165, Advance Monthly Sales for Retail and Food Services, September 2026, released 15 October at 8:30 a.m. Eastern — 12:30 UTC — states August sales at 739.3 billion dollars, up 1.3 per cent from July, with a margin of error of plus or minus 0.1 percentage points. It also states, in the same document, that August was previously reported as up 1.1 per cent, plus or minus 0.4.
So the two readers this desk had in conflict were a vendor calendar carrying 1.1 and a third party computing about 1.1 from a data service’s levels. Both were faithful to what had been published. Neither was a measurement of August, because the first print of August was never precise to a tenth and the issuer said as much on the page.
The generalisation is cheap and we should have applied it a week ago: when two vendors differ by less than the issuer’s stated margin of error, there is no disagreement to adjudicate. Check for the error bar before opening the argument. We did not, and the cost was two slots and a carried item.
September beat by less than its own error bar too
September advance retail sales printed up 0.8 per cent on the month, 745.0 billion dollars, plus or minus 0.4. Consensus readings this desk holds: 0.3 from a Japanese calendar and from one American reader, 0.4 from a second American reader, and 0.3 headline with 0.5 ex-autos from a 9 October collation of two vendor calendars.
Run the error bar. The print’s own 90 per cent interval is 0.4 to 1.2. The lower bound equals one vendor’s consensus exactly. A headline reported everywhere as a beat of four or five tenths is, at the issuer’s own stated confidence, not distinguishable from the number it beat.
The internals are the part nobody led with. Ex-autos was up 0.9 per cent and 7.2 per cent on the year. Ex-autos and gasoline stations was up 0.2 per cent. And one reader gives the control group at 0.1 per cent against a 0.3 forecast, from a 1.2 prior — a miss, in the same release as the beat, on the measure that feeds the consumption accounts. Total sales were up 6.3 per cent on the year, plus or minus 0.5.
A release that beats on the headline, misses on the control group, and shows two tenths once you strip cars and fuel is not one fact. If your filter reads the headline field only, it read the strongest version of a mixed report.
Core producer prices came in three different numbers, and the issuer prints only one of them
Headline final-demand PPI for September rose 1.3 per cent on the month and 6.3 per cent on the year, the Bureau of Labor Statistics states, released the same minute. Estimates in this desk’s file: 0.5 and 0.6 on the month, and 5.3, 5.4 and 5.5 on the year. Against the mid of those, the monthly print was seven to eight tenths high and the annual print eight tenths to a full point high. On any reading it is a large miss in one direction, and the energy component explains much of it — final-demand goods up 2.6 per cent with energy up 10.8 per cent, and trade services margins up 1.4.
Then there is core, and this is the finding. The BLS news release gives final demand less foods, energy and trade services at 0.5 per cent on the month and 5.0 per cent on the year. It does not state a monthly change for final demand less foods and energy anywhere in the release text.
Published readings of “core PPI” on the month: 0.7 per cent at two readers, 0.5 per cent at two others, and 0.8 per cent at one reader which corrected itself to 0.7 at 16:21 GMT, thirty-five minutes after publishing. That is three distinct values for one field, and one vendor publishing a correction on it inside the hour. The two tenths between 0.5 and 0.7 is not an error by anybody. It is two different baskets being given the same one-word name, and the stricter of the two is the one the issuer actually prints.
This file has a standing rule that core is not one series. This is the sharpest instance of it yet, because the measure most vendors call core is the one the issuing agency does not put in its own news release.
Two Federal Reserve banks released manufacturing surveys in the same minute, 35 points apart in direction, and each calendar carried one of them
At 12:30 UTC the Philadelphia Fed published its October business outlook survey at 20.9, from 37.8 in September — down 16.9 points, with new orders collapsing from 29.2 to 4.3 and the six-month forward index from 52.9 to 33.0. In the same minute the New York Fed published Empire State at 25.5, from 7.6 in September — up 17.9 points. Two regional surveys of the same sector, published simultaneously, moving 34.8 points of combined swing in opposite directions.
The surprises were opposite too. Philadelphia missed every published expectation we hold: 26.4, 28.0, 28.4 and 32, a 5.6-point spread among four readers, with the print 5.5 to 11.1 points below each. Empire beat its two published expectations of 10.25 and 12.0 by 15.25 and 13.5 points.
Yesterday this desk published that three calendars gave three different rosters for this minute and opened an item to settle which bank released. The answer is both, and each vendor calendar carried exactly one: the Japanese morning schedule listed New York and no Philadelphia; the Japanese evening schedule listed Philadelphia with a forecast and a prior and no New York. The week-ahead table this desk carried from 13 October listed both and was the only one that was right.
Then the issuers, and the contrast is almost too neat. The Philadelphia Fed’s release page states, in terms: released October 15, 2026, at 8:30 a.m. ET. The New York Fed’s October survey document states no release date, no release time and no timezone anywhere in it, and gives no next release date. It does state that responses were collected between 2 and 9 October — which closed four and a half days before the inflation print the market is trading. The survey that beat by fifteen points was taken before the event.
The hottest wholesale print in four years cut the October hike probability by fourteen points
This is the number that should bother you. Producer prices beat on the headline, on both monthly and annual, on the year-on-year core, and on the goods side. And the probability of an October increase fell.
The readings, with their hours, because this desk published a probability as a level two days ago and had to retract it: 43.7 per cent at 13:13 UTC on 14 October; 31.6 at 23:23 the same evening; around 40 or less at 08:31 on 15 October; 26 per cent at 22:48 on 15 October, attributed to Prime Terminal, down from 40 a day earlier. A fifth account, published 15:46, says October odds reached about 60 intraday on 14 October before closing near 40 — a figure no other reader in this file gives, and we are logging it rather than using it. December, at the same 22:48 reading, is near 92 per cent.
So the ratio is roughly three and a half to one between December and October on the day the inflation data ran hottest. We are not going to explain that with confidence, and we would rather name the candidates. It could be that the market reads an energy-driven producer print as a tax on growth rather than a policy input, which is what a retail control group at 0.1 and a Philadelphia collapse to 20.9 would support. It could be that the October meeting was never seriously live and the 40s were a reaction artefact that has decayed. It could be that four of the eight officials on the executive side are, per one reader this morning, cautious about moving early, and the probability is tracking them rather than the data. Those are three different trades and we cannot separate them from here.
Claims, for completeness: initial claims 199,000 against an estimate of 200,000 and a carried prior of 197,000 — the one field in the block that landed on its consensus. Continuing claims 1.710 million, a fifth consecutive week below 1.73 million.
What to do with a release block that scores like this
Four rules, all of them about measurement rather than direction.
Put the issuer’s margin of error in the same field as the print. A beat smaller than the error bar is not a beat, and two of today’s six releases qualify. If your calendar table has no error-bar column, the beats it reports are not scoreable.
Name the basket on every core field. If your rule fires on “core PPI above forecast”, it just fired or did not depending on which of three published values your feed carries, and the issuer prints only the smallest of them.
When two releases share a minute, make sure both are in the schedule, and check that they cannot contradict each other. Today they did, by 35 points of combined swing, and a filter holding only one of the two measured half the minute.
And stop treating a published probability as a level. Four readings of the October meeting across 34 hours span 43.7 to 26. Carry the timestamp in the same field or do not carry the number.
What this does not tell you
It does not tell you the consensus for any of these releases, because there is no such object. Every forecast figure above is one vendor’s or one collation’s, the spreads are wide, and we have named the spread rather than picking a midpoint.
The core PPI story is a claim about naming, not about anybody’s arithmetic. If a reader quoting 0.7 per cent labelled its basket explicitly and we read past the label, the fault is ours. We read the issuer’s release and reported what it does and does not print.
The probability figures reached us through publishers rather than from the futures tool itself, which this desk has not read directly at any point this week. The 26 and the 92 are one reader’s at one hour. The 60 is one reader’s account of a day that had already closed.
The Empire State survey-window point is a fact about the collection period, not a claim that the index is wrong. A survey taken before an event is still a measurement of the weeks it covers. It is simply not a measurement of the thing the market repriced on 14 October, and the two get used interchangeably.
The retail control group figure is single-sourced and we have labelled it as such in the text. And the ten-year yield, which several of these moves ran through, still has no published observation at the Federal Reserve’s own series for 14 or 15 October — a stated release date that has now passed with the newest value still 13 October at 5.21 per cent, against two publishers quoting 5.035 and 5.06 today. We are not reconciling that here.
No directional view, on anything. A block in which the inflation field ran hot and the hike probability fell is a reason to widen your error bars, not to pick a side.
Related
- FX event calendar — the 12:30 UTC block and Friday’s schedule
- Signals — scoring a print against a consensus that has a spread
- EA presets — release-window filters when two releases share a minute
- We published 43.7 per cent. Seven minutes later the same feed read 31.6. — the probability retraction this piece extends
- Eight readers agreed on one field. It is the one they all got wrong. — the CPI print, and the sampling-error argument applied again here
