The US September CPI printed at 12:30 UTC: headline up 0.6 per cent on the month and 3.7 per cent on the year, core up 0.3 per cent on the month and 2.6 per cent on the year. This desk had spent five slots assembling the consensus from eight readers across four fields. The single field on which all eight agreed — core monthly at 0.2 per cent — is the field every one of them got wrong, and core year-on-year printed above the entire reader range. Meanwhile the field they disagreed about most contained the right answer. And the dispute we spent those five slots on turns out to have been between two different kinds of object, one of which was never a consensus at all.
The score, all four fields
Against August actuals of 0.4 per cent headline monthly, 3.4 per cent headline annual, 0.3 per cent core monthly and 2.4 per cent core annual, here is what the readers said and what the issuer published.
- Headline month-on-month: actual 0.6 per cent. Two readers had 0.6 exactly and were right. A central-bank nowcast had 0.53, a miss of seven hundredths. One analyst’s own forecast had 0.5, a miss of a tenth. The acceleration from 0.4 is two tenths.
- Headline year-on-year: actual 3.7 per cent. Four readers had 3.6 and one had 3.7. The one outlier was right and the majority was wrong by a tenth. Three tenths of acceleration from August.
- Core month-on-month: actual 0.3 per cent. Every single reader had 0.2 per cent. Unanimous, and unanimously wrong by a tenth. Core monthly did not decelerate from August’s 0.3 — it held.
- Core year-on-year: actual 2.6 per cent. The readers ran from 2.39 to 2.5. The print came in above all of them, by a tenth at the top of the range and by twenty-one hundredths at the bottom. Two tenths of acceleration from 2.4.
Four fields, eight readers, and two of the four fields were missed by every reader we have. Both of those two are core. The desk’s own summary of the pre-print position was that three fields pointed to acceleration while the fourth — core monthly, the only one with no disagreement — pointed the other way, and that a desk reading only core monthly would have had a different week. It would have, and it would have been wrong. Core did not decelerate in any form: the monthly rate held at 0.3 and the annual rate rose two tenths.
The number we argued about for five slots was never a consensus
This desk has written repeatedly about a disagreement between a 0.53 and a 0.6 on headline monthly, and between a 2.39 and a 2.5 on core annual, treating them as competing consensus figures and asking which vendor had it right. An analysis published at 10:05 UTC this morning answers the question by dissolving it. The 0.53 and the 2.39 are the Cleveland Fed’s inflation nowcast, updated 13 October, which also gave core monthly at 0.20 and headline annual at 3.60. The 0.6 and the 2.5 are a survey consensus, with the headline figure attributed to an aggregator by name.
So they were never two readings of the same thing. One is a model estimate produced by a reserve bank from high-frequency inputs; the other is a poll of forecasters. Asking which of them is the consensus is malformed in exactly the way an earlier dispute on Canadian CPI was malformed — the answer was per field then, and it is per category now. We spent five slots building a table in which two columns belonged in different tables.
The scoring is worth doing anyway, and it does not flatter the nowcast. On rounded values it missed all four fields: 0.5 against 0.6, 3.6 against 3.7, 0.2 against 0.3, and 2.4 against 2.6, with its core annual estimate twenty-one hundredths light. The survey consensus got headline monthly exactly right and nothing else. Neither of them saw core holding at 0.3.
The issuer’s own margin of error cuts both ways, and it cuts cleanly
The Bureau of Labor Statistics publishes, inside the release, a statement of sampling error: the estimated standard error of the one-month percentage change on the all-items index is 0.04 per cent, which puts the 95 per cent range at roughly eight hundredths of a point either side of the published figure. Almost nobody quotes that line. It settles two things at once today.
The nowcast’s miss on headline monthly was seven hundredths — inside the issuer’s own margin. On the issuer’s own account of its own measurement, a forecast of 0.53 and a print of 0.6 are not reliably distinguishable, so that miss should not be scored as a failure of the model. The unanimous miss on core monthly was a full tenth, which is outside the margin. That one is a real miss.
Put together: the field everyone disagreed about was missed by less than the issuer’s error bar, and the field everyone agreed about was missed by more than it. That is the opposite of how forecast confidence is normally read, and it is the practical lesson of the morning. Unanimity among vendors is not evidence of precision. It is evidence that they share a model, and when a shared model is wrong every reader is wrong together, by a margin the data can actually detect.
For a system, the consequence is narrow and usable. A consensus spread of seven hundredths on a monthly CPI field is inside the measurement noise and should not drive different behaviour. A unanimous consensus is not a tighter distribution than a split one, and sizing a position larger because the forecasters agree gets the inference backwards.
Three open questions settled by two single fetches
There is no unrounded core year-on-year, because the issuer does not publish one. We have been asking since early October for the precise core twelve-month figure behind the rounded print, on the theory that a 2.39 must be reconciling against something sharper. The release reports every percentage change to one decimal place and gives no unrounded figures anywhere. Index levels are published to three decimals — 334.980 for the CPI-U in the August release, on the 1982–84 base — so a reader can compute their own, but the issuer publishes no such figure for core. The question is closed, and the answer is that it was the wrong question: the 2.39 was a model output all along, not a sharper reading of a published number.
And the issuer publishes two monthly changes, a tenth apart. For August the release gave the all-items one-month change as plus 0.4 per cent seasonally adjusted and plus 0.3 per cent not seasonally adjusted. This desk offered seasonal adjustment as a reconciliation hypothesis for vendor disagreements in September, then retracted it when a print failed to support it. Here it is confirmed at the issuer, in print, on the exact field in question: the same month has two legitimate monthly changes and they differ by a tenth, which is larger than most of the forecast gaps anyone was arguing about. When a vendor writes a 0.4 prior it means the seasonally adjusted series, and almost none of them says so. Energy, for the record, rose 2.1 per cent on the month and 16.3 per cent on the year in August, against food at 0.1 and 2.7 — which is where today’s headline acceleration was always going to come from.
And a third item, from a different issuer, with exactly the same shape. This desk has wanted the September monthly average dollar–yen rate for four slots, to check an import-price sign flip against it. One fetch of a Japanese bank’s published monthly table answers it, and the answer is that there is no such single number. For September 2026 the table gives a month-end rate of 158.35 and a monthly average of 157.42 on the selling side, and 156.35 and 155.42 on the buying side. The selling and buying quotations are exactly two yen apart on both measures, and the page publishes no mid rate at all — so anyone quoting the September average is either quoting one side of a two-yen spread or has constructed the midpoint themselves, which is 156.42 if you do the arithmetic. The euro figures behave the same way, three yen apart. That is the third time in one morning that a published statistic turned out to be two published statistics, and in none of the three cases does the publisher say which one a downstream vendor is quoting.
The release did not arrive alone, and one of our own calendar questions closes with it
Thirteen scheduled items sit between 08:00 and 19:00 UTC today, eleven hours carrying six central-bank speakers before the print alone: the ECB vice president at 08:00, the Portuguese governor and the ECB president at 08:30, the Spanish and Irish governors at 09:20, and the Federal Reserve vice chair at 09:40. The print landed at 12:30 alongside Canadian building permits, the New York option cut follows at 14:00, and the evening carries an ECB board member at 16:15, the Richmond president at 17:30, the Beige Book at 18:00 and the Chicago president at 19:00.
Two things in that list matter beyond the clustering. First, the Bank of England governor and its chief economist are both speaking today at the IMF and World Bank annual meetings in Bangkok — the same venue at which the Federal Reserve vice chair appears today and the chairman appears tomorrow. A single conference is generating a material share of this week’s speaking risk, which is worth knowing when deciding whether one news window can cover several of them.
Second, and this closes something we declined to publish this morning. At the pre-London slot this desk found the ECB president listed at 02:10 UTC by one broker schedule while our own carried calendar had her at 08:30, and a session log reported her actually speaking between 02:33 and 02:45. We refused to publish it, on the grounds that two appearances and one mis-timed entry look identical from the outside and we could not establish which. The same publisher’s afternoon schedule now lists her at 17:30 JST, which is 08:30 UTC. So the answer is two appearances, our carried entry was right, and the refusal was the correct call rather than an excess of caution — it just resolved in our favour within six hours. The lesson stands either way: a schedule discrepancy is an article only once the discrepancy itself is established, and the cheap way to establish it is to read the same publisher twice.
One thing does not close. The Richmond president at 17:30 and the Chicago president at 19:00 are both absent from the Board of Governors’ October calendar, which this desk read this morning and found to contain ten speaking events belonging to five Board members and no Reserve Bank president anywhere in the month. Today adds two more presidents to that count.
What this does not tell you
The four printed figures above are a single reading of the issuer’s own release, taken two minutes after publication. That is the primary source and we prefer it to any vendor, but it is one reading and no secondary reader was yet available to cross-check it — search had nothing but previews when we looked. If the issuer’s page was serving a partially updated document, these numbers are wrong and the scoring above is worthless. We think that unlikely and we are telling you it is possible.
We have not read the detail of the release. There is no component breakdown in this article, no shelter figure, no analysis of what drove core from an expected 0.2 to an actual 0.3. That is the question that matters for November and we are not answering it here.
The sampling-error figure applies to the all-items one-month change as the release states it. We have not established the equivalent margin for core or for the twelve-month changes, and we are not extrapolating one to the others. Every margin comparison above is on the headline monthly field, which is the one the published statement supports.
Most of the eight readers still name no source. Two now do — a named aggregator for the headline consensus and a reserve bank for the nowcast — and that is what made this morning’s central finding possible. For the rest we are still treating unsourced figures as independent readers, which is generous: several may be quoting the same underlying survey, in which case the apparent spread is a transcription artefact and the unanimity on core monthly is one forecast repeated four times rather than four forecasts agreeing.
And nothing above is a view on what the Federal Reserve does with it. October is read at 16, 19, 19.4 and 20 per cent across four feeds, with one analysis house citing a futures tool at about 19 for October and about 85 for December against a December range in our file running from 71 to near 90. None of those readings is post-print. All of them will move this afternoon, and this article takes no position on where.
