Wednesday at 12:30 UTC the United States publishes September consumer prices, and this desk has spent four consecutive weeks discovering after the fact that it did not know what the consensus was. So here it is, established before the print. The headline is 3.6 or 3.7 per cent depending on which day you read the same aggregator. Core is 2.39, 2.4 or 2.5 against a 2.4 prior — three readers, three different signs. And the interesting part is that all three of them agree on what September itself did.
The headline consensus is 3.6 or 3.7, and the difference is five days
One aggregator’s own page, dated 11 October, calls the street consensus 3.6 per cent year-on-year against a 3.4 per cent August. A preview published 8 October quotes the same aggregator at about 3.7 per cent as of 6 October. Two other readers sit at 3.7: a research house dated 29 September, and a bank’s week-ahead note dated 9 October. The spread between the highest and lowest published consensus is ten basis points, and at least part of it is not disagreement at all — it is the same source at two different moments.
That is the metadata failure this desk keeps relearning. A probability needs its event, its horizon and its feed stored with it. A positioning figure needs its trader category. A consensus needs its as-of date, and a consensus quoted second-hand has a hidden one. Anyone who reads “consensus 3.7” in a preview on Wednesday morning is reading a figure that was current on 6 October, five days and one oil move ago.
Around the surveys sit the model estimates, which are a different animal and should not be averaged in: a regional Federal Reserve nowcast at 3.60 as of 7 October, a prediction-market price at 3.63 as of 6 October, and the aggregator’s own ensemble at 3.56 with its in-house model at 3.46. Counting those, the full published band for one number is 3.46 to 3.7 — twenty-four basis points. Counting only things that call themselves a consensus, it is ten.
Core has three readers, three signs, and one agreed monthly change
On the year-on-year core, the nowcast says 2.39, the research house says 2.4 and the bank says 2.5, against a 2.4 per cent August. Those are not three estimates of the same size. They are three different directions: down a hundredth, unchanged, and up a tenth. Eleven basis points separate the ends, which is larger than the headline consensus spread, on the field the Federal Reserve actually watches.
Now look at the monthly figures from the same three readers. Core month-on-month: 0.20, 0.2 and 0.2, against a 0.3 August. All three agree, to the decimal they publish, on what core prices did in September — and all three agree it decelerated by ten basis points on the month. A 0.2 core month annualises to 2.4 per cent, which is exactly the prior year-on-year rate. On their own numbers, September was a core-neutral month and they say so unanimously.
The headline monthly figures spread a little: 0.53 from the nowcast against 0.6 from the other two, seven basis points, with a 0.4 prior. Subtract core from headline and you get between thirty-three and forty basis points of non-core price change in a single month. Whatever Wednesday prints, the gap between the two fields is where it lives, and crude traded across a five-dollar band of published readings today alone.
The core disagreement is about a month that was published a year ago
Here is the arithmetic that makes the core spread interesting rather than annoying. A year-on-year rate is last year’s rate, plus this month’s change, minus the change in the month dropping out of the window. Take the August 2.4 as given and the agreed September 0.2 as given, and each reader’s core year-on-year implies a specific value for September 2025.
The nowcast’s 2.39 implies the base month was 0.21. The research house’s 2.4 implies 0.20. The bank’s 2.5 implies 0.10. That is three different values, eleven basis points apart, for a month that was measured, published and revised a year ago. September 2025 is not a forecast. It is history, and it has one value.
So the three readers are not disagreeing about September 2026 — they already agree about that. They are disagreeing about the rounding on either side of it. Both the prior and the base are published to one decimal, and an unrounded August core year-on-year anywhere in roughly 2.35 to 2.44 will produce all three of these answers from the identical 0.2 monthly forecast. The spread is the rounding of two numbers nobody forecast, propagating forward.
State the limit plainly: this is our arithmetic, not any publisher’s statement. We did not obtain an unrounded August core year-on-year or the unrounded index levels from the issuing agency, and until we do, the decomposition is a demonstration of how the spread could arise rather than proof of how it did. But the direction of the conclusion does not depend on the missing digits, because the alternative is that two of three professional forecasters are wrong about a published historical month, which is less likely than rounding.
What to do with this at 12:30 on Wednesday
Score the monthly fields, not the annual ones. A forecaster scored on core year-on-year is being graded on the rounding of a month they did not forecast, which is why three of them can agree about September and still look eleven basis points apart. The month-on-month figure is the only one of the four that is wholly about the period being released.
Expect the two fields to tell different stories, and decide now which one your rule reads. The headline is forecast to rise two or three tenths from 3.4; core is forecast to do nothing, in three directions. That is not a contradiction, it is the definition of the two series: the entire headline move is outside core, which means it is food and energy, which means it is the oil that is sitting in a five-dollar band of published readings. A filter that fires on “inflation beat” without naming the field will fire on the energy pass-through and call it a policy signal.
And carry the as-of date. If you store one consensus figure for Wednesday, store the day it was read with it. Three of the four consensus readings above were published between 29 September and 11 October, and the oldest and the newest come from the same house at different numbers. On Friday a consensus dispute this desk closed three times in a month still failed to predict the print; the point of establishing the figures beforehand is not that they will be right, it is that scoring them afterwards is only honest if you wrote them down first.
What this does not tell you
It does not tell you what September consumer prices did. Nothing here is a forecast, including the arithmetic.
It does not give you a core year-on-year consensus from a survey. The three core readings are a regional Federal Reserve nowcast, a research house’s published estimate and a bank’s own forecast — three different kinds of number that we have deliberately not averaged, because a nowcast and a house view are not survey participants. If the true street consensus for core differs from all three, every comparison above changes.
Two of the three core readings reached us through one preview page rather than from their own publishers, so they are a single fetch away from being single-sourced. The bank’s figures were read directly at the bank. The headline 3.6 was read directly at the aggregator; the 3.7 attributed to that same aggregator was read in the preview and we did not see the aggregator’s 6 October page itself, which is precisely the hidden-timestamp problem we are complaining about, committed by us.
And it does not tell you what the Bureau of Labor Statistics will call core. Core is not one series; prefer the issuing agency’s own figure and say which basket it is. We have not yet read the issuer’s release page for this print, and we will before we score anything.
Related
Release times and event windows: FX events calendar. Scoring prints against consensus: signals. News-window and filter configuration: EA presets.
