The Bureau of Economic Analysis says real gross domestic product rose at an annual rate of 3.4 per cent in the third quarter. Twelve hours before that number existed, this desk published the four consensus readings it had found for it — one blank, one 3.0, one 3.2, one 3.43 — and called the 0.43-point spread a problem. Five more readings have since turned up. Nine in total, four distinct numbers and a hole, and the print landed on 3.4. Two things follow, and the second is the one that matters: the reading that came closest is the one that looked broken, and the figure everybody is now scoring against is an advance estimate whose immediate predecessor was revised up by seven tenths.
3.43 Missed by Three Hundredths. The Number Five Readers Agreed On Missed by Two Tenths.
Rank them against the issuer. The advance estimate is 3.4 per cent, seasonally adjusted annual rate, 0.8 per cent at a quarterly rate, release BEA 26–45, read directly from the Bureau’s own release. Against that: 3.43 is out by 0.03. The modal figure, 3.2, carried by five of our nine readings, is out by 0.2. One reading of 3.1 is out by 0.3, 3.0 is out by 0.4, and one vendor published no figure at all and was therefore the only one that did not mislead anybody.
We want to be exact about what this does and does not establish. It does not establish that the 3.43 publisher forecasts better. A figure quoted to two decimals on a series the issuer reports to one is almost certainly a mean of contributor estimates that somebody forgot to round, and a mean is not a median. What it establishes is narrower and more annoying: on this release, the reading whose false precision made it look like a scraping artefact was the only one inside a tenth, and the clean round numbers that look like a consensus were not.
The two nowcasts beat the consensus too, and they are not consensus. The Atlanta Fed’s model stood at 3.3 per cent on 28 October, one tenth from the print and better than every published market expectation except the unrounded one. The New York Fed’s stood at 2.7 per cent on 23 October, which is seven tenths out and a week older. If your event log has one field for “expected,” it can hold a survey median, a vendor mean, or a model, and those three were 3.2, 3.43 and 3.3 on the same release.
The operational version is a sizing rule, not a forecasting one. If your surprise score is actual minus consensus and your consensus field held 3.0, you booked a 0.4-point beat. If it held 3.43 you booked nothing. Same release, same tape, two opposite readings of how much news arrived. Score it against the band, 3.0 to 3.43, and today reads as a beat at the low end and a miss at nothing. That is less tradeable than either single number, which is the point.
The Prior Settled, and the Prior Is Not Finished Moving.
Every reader put the second quarter at 2.2 per cent and the Bureau’s own release says the second quarter increased 2.2 per cent. Our standing rule is that the prior is settleable at the issuer and the forecast never is, and it settled.
Then one reader supplied the detail that undoes the comfort. That 2.2 is not what the second quarter’s advance estimate said. The advance figure was 1.5 per cent, and it was revised to 2.2 on 30 September — seven tenths, upward, on the same series, one quarter ago. Today’s 3.4 is an advance estimate of exactly that kind. The second estimate publishes on 25 November and the Bureau says so on the release.
So the whole argument above has a shelf life. We have just spent four hundred words ranking nine forecasts to within three hundredths of a number that moved seven tenths last time it was measured, and the honest conclusion is that the ranking is noise at the resolution the forecasts were quoted to. The forecasts were wrong by two tenths. The revision was seven. Any system that scores a release at the advance print and never revisits it is carrying a figure that is, on recent evidence, less accurate than the spread it was scored against.
If you want one rule out of this: do not resolve a forecast dispute at the advance estimate. Resolve it at the second, and keep the first in the file as what was known at the time rather than as what happened.
The Core PCE Field Printed Three Times Today: 0.3, 2.5 and 3.1.
At 11:38 UTC yesterday this desk published that “Core PCE Price Index” at 0.3, “Core PCE Prices” at 2.40 and “Core PCE Price Index (annual)” at 3.1 were three different objects separated only by their priors, and that the Federal Reserve’s target variable appears three times in one minute over three horizons. That was a claim about labels. It is now a claim about prints.
All three landed. The September monthly core index rose 0.3 per cent against a consensus of 0.3. The third-quarter core index rose 2.5 per cent at an annual rate against the 2.40 that was sitting in the field — and the 2.5 is the Bureau’s own figure, on a second quarter of 3.3 that two separate readers also carried as the prior, so that row is anchored at the issuer. The September annual rate printed 3.1 against a consensus of 3.1 on a prior of 3.0. That last actual we have from one reader only; treat it as one desk’s reading rather than a settled number.
Three numbers, one name, one minute. The headline index did the same thing: 0.5 on the month, 2.1 annualised on the quarter, 3.6 on the year. Six values across two series in a single release window, and the only thing separating them in most calendar tables is a parenthesis.
One of today’s readers published its row as “Core PCE Price Index” with a consensus of 3.1, a prior of 3.0 and no reference period anywhere on the line. That is the exact failure we described yesterday, found in the wild within the day, and the reason it is dangerous is that 3.1 against 3.0 is a perfectly plausible monthly row to a system that is not checking. It is not a monthly row. A monthly core print of 3.1 per cent would be the story of the decade.
The Biggest Miss of the Day Had a Consensus Spread of 0.8 and Nobody Is Discussing It.
The third-quarter PCE price index — headline, annualised, the one the Bureau prints two lines above core — came in at 2.1 per cent. Two readers had a consensus of 3.4 for it. A third had 2.6. That is a 0.8-point spread on a consensus for a quarterly inflation series, and the print landed 1.3 points under the higher pair and 0.5 under the lower one.
We can anchor which series the 3.4 belongs to without guessing, because the readers carrying it also carry a prior of 5.0, and the Bureau’s release says the second-quarter PCE price index increased 5.0 per cent. The prior identifies the row. So 3.4 is a genuine consensus for the series that printed 2.1, and it was wrong by more than a third of its own value. The 2.6 reading, by contrast, is published with no prior at all, which means it cannot be anchored to a row even in principle. A consensus without a prior is not a number you can check.
Nobody is discussing it because the growth number is more exciting. But growth beat and the headline quarterly price index fell from 5.0 to 2.1 in one quarter, which is the combination that makes a hold look cheap, sitting under a 3.4 expectation that would have made the same hold look negligent. If your event log recorded only the GDP surprise from this release, you recorded the smaller of the two.
Everyone Agrees on the Prints. They Disagree About Last Week and Last Quarter.
Initial jobless claims came in at 194,000 on two readings, against a consensus of 199,000 on three and 198,000 on a fourth. Continuing claims 1.701 million against 1.720 million, four-week average 196,750. So far, so ordinary.
Now the priors. One pair of readers puts last week’s initial claims at 195,000; another puts it at 194,000. The continuing-claims prior is 1.722 million to one reader and 1.728 million to another. The four-week average prior is 198,000 to one and 197,750 to the other. The September core monthly prior is 0.2 per cent to two readers, including the one that lays the series out most carefully, and 0.3 per cent to a third. Four disagreements, all about numbers that were published, are fixed, and cannot move.
A forecast is allowed to be contested. A prior is not. When two vendors disagree about a figure that has already printed, it is a revision somebody carried and somebody did not, or a transcription error, and in either case it is the cheapest thing on the page to settle and the thing least likely to be settled. We did not reach the issuing agencies for these, so we are reporting the spread and not adjudicating it.
One claim we are cutting rather than publishing. A reader states that third-quarter consumer spending “accelerated to 3.8 per cent.” Two other readers give 3.8 per cent as the prior for that series, with consensus at 3.4 and 3.9 respectively. A figure identical to the prior, presented as the actual, in a release where the prior sits one column to the left, is more likely to be a column error than a coincidence, and we are not publishing a quarterly consumption number on it.
And a caution on everything above. Two of today’s readers agree on most consensus values; one describes itself as written with help from an AI and the other as an automated newsroom. Agreement between them may be one upstream feed counted twice. We are treating them as one and a half.
What This Does Not Tell You
It does not tell you what the consensus was. It tells you what nine pages said the consensus was, which is a different object, and the band is 3.0 to 3.43 with a blank in it. It does not tell you the 3.43 publisher is better: one release is one release, and we would say the same of any single hit.
The September annual core and headline rates, 3.1 and 3.6, come from one reader each and are not confirmed at the issuer. We did not reach the Bureau of Labor Statistics or the Employment and Training Administration, so the four prior disagreements are unresolved rather than adjudicated. Component contributions for the GDP release sit in a chart image that does not extract, so we have the direction of consumer spending, inventories, equipment, exports and federal spending from the Bureau’s text and no percentage-point contribution for any of them. The Bureau’s release carries no note of a data-collection disruption; we looked and there is none.
And none of this is a direction. A 3.4 per cent growth print arriving with its quarterly price index already down to 2.1 is a good quarter with cooling prices attached, and what that is worth depends entirely on what was priced before 12:30, which is not a number we have.
