The New York Federal Reserve’s November manufacturing survey printed 7.80 at 13:30 UTC on Monday. This desk went into that release carrying four published consensus figures — 17.6, 17.6, 19.5 and 18.0 — against an October prior of 25.5 that every reader agreed on, and we published a pre-registration saying the 1.9-point spread between those four was the wrong error bar. It was. But not in the way we said, and the sharpest thing in the release is not the miss. Two more readers we had never opened put the consensus at 17.0, which makes six published readings and four distinct numbers; the dispersion we published was 24 per cent narrower than the real one; the closest forecast in the field belongs to a publisher absent from our own table; and inside the release that fell 17.7 points, the employment index rose 7.8 and the six-month outlook rose 7.9.
Six Published Consensuses, Four Different Numbers, and Every One on the Same Side
Here is the field as it now stands, sorted by value, with the print at 7.80 and the October prior at 25.5 agreed by every source including the issuer.
- 17.0 — GuruFocus, and separately Seeking Alpha. Asked for a fall of 8.5. Missed by 9.20.
- 17.6 — TipRanks, and separately XTB. Asked for a fall of 7.9. Missed by 9.80.
- 18.0 — MarketScreener. Asked for a fall of 7.5. Missed by 10.20.
- 19.5 — Minkabu, on its 07:20 Japan-time page. Asked for a fall of 6.0. Missed by 11.70.
The realised fall was 17.70 points, which is 69.41 per cent of the October level and between 2.24 and 2.95 times the fall any of them asked for. The headline printed 7.80 on three independent readers — GuruFocus, Seeking Alpha and MarketScreener — and we are treating that as settled. A fourth, InvestingLive, carries 7.80 in a URL slug reading 7-80-vs-18-00-expected, whose body we could not open; per our own standing rule we are not publishing a figure off a slug, so that reading is here as a slug and nothing more.
The structural fact is that all six readings sit above the print. Not one straddled it. The worst-to-best ratio of the misses is only 1.27, so the vendors were wrong together and wrong by similar amounts, and the full 2.5-point spread between them is 27.17 per cent of even the smallest miss. If you were running a system that stored the mean of the available consensuses and sized a window off their disagreement, the disagreement was not the error — the common component was.
The Closest Reader Is One We Never Opened, Which Makes Our Error Bar a Subset of What It Measured
At the pre-US slot we published a dispersion of 1.9 points, from four readers. The real spread across six readings is 2.50 points, because 17.0 exists and we had not found it. Our published figure understated the true range by 0.60 points, or 24 per cent of it.
That is worse than a missing data point, and it is worth being precise about why. We were not using those four numbers to forecast the survey. We were using the spread between them as a statement about how much forecasters disagree, and then arguing that realised surprise runs several times larger than that spread. A dispersion computed from an incomplete sample of the forecasters is not a conservative estimate of disagreement — it is a number whose error runs in an unknown direction, and here it ran toward false precision. The error bar was built from a subset of exactly the quantity it was measuring, and the publisher we had missed turned out to hold the best forecast in the field.
Now score the pre-registration, which we said was losable. We wrote that realised surprise runs about eight to ten times dispersion, so a 1.9-point band was wrong and something like 15 points was the right order. The verdict:
- The direction won. Dispersion was emphatically the wrong ruler. The smallest miss is 3.68 times the true spread and the largest is 4.68 times it.
- The multiple lost, and lost clearly. We said eight to ten; we got 3.68 to 4.68 — under half the lower end. Against the narrower dispersion we had actually published, the ratio is 4.84 to 6.16, still short of eight.
- The 15-point band was too wide, not too narrow. The largest miss was 11.70, which is 0.78 of the band we pre-registered. We widened the error bar in the right direction and overshot.
Last month the same series gave a ratio of exactly 8.0, printing 25.5 against a 10.3 forecast. This month the identical series gives 3.68 to 4.68. Two consecutive months of one survey produce multiples differing by a factor of about two, which is the argument against turning this measurement into a constant. It is a reason not to size a blackout from vendor dispersion. It is not yet a coefficient, and anyone who tells you the multiple is eight has one observation.
Employment Rose 7.8 Points in the Release That Fell 17.7
One publisher carried November sub-indices. Two of the four it gives went the opposite way to the headline, and the issuer’s own October report supplies the baselines.
- New orders: 23.2 to 3.0, a fall of 20.2 — larger than the headline’s own fall. Corroborated by Seeking Alpha.
- Shipments: 13.1 to 8.8, a fall of 4.3. Also corroborated by Seeking Alpha.
- Number of employees: 10.3 to 18.1, a rise of 7.80. Single-sourced.
- Future business conditions, the six-month-ahead index: 21.0 to 28.9, a rise of 7.90. Single-sourced.
Both of the rising fields are single-sourced on GuruFocus, and we say so rather than resting an argument on them. What we can check is the arithmetic, and it reproduces: the page describes the six-month index as climbing eight points to 28.9, and 28.9 minus the issuer’s October 21.0 is 7.9, which rounds to eight. The employment line reads as a rise to 18.1 from 10.3, and that difference is 7.8 exactly. A derived figure that reproduces from its own stated inputs is a different class of object from one you have to take on trust, and both of these reproduce against the issuer’s numbers rather than the publisher’s own.
So of five published fields, three fell and two rose, and the two that rose are labour and expectations. The honest reading of that is not that the survey was secretly fine. New orders falling 20.2 points to barely positive is the worst number in the release and it is a demand field. The reading is narrower: a system that stores one number from this release stores a 17.7-point collapse, and two of the five fields the same release published are pointing the other way. Which field your filter watches decides what this release meant, and nothing in a one-row calendar entry tells you there were five.
And we could not get the issuer. The New York Fed’s own survey overview page was still serving the October report at the time of this research — its note says responses were collected 2 to 9 October, and its release calendar shows a bare 16 for November with no report link. The constructed document URL for the November release returned HTTP 200 carrying the site’s own page-not-found document: a soft failure that any pipeline checking status codes would score as a success. A secondary reader links the November file under an entirely different filename pattern from the one the October file uses, which is why a pattern we could build did not resolve. This is the third time this month we have logged an issuer or publisher serving the wrong period from a current-release address, after the producer-price release on 13 November and a Japanese bank’s session table, which we establish separately below.
Canada Missed by Exactly 0.2 in Both Fields, Which Is Not What We Said Would Happen
Canadian October consumer prices landed in the same 13:30 UTC minute as the New York survey. The issuer’s own release settles it, and it settles it against our expectation about the shape of the answer rather than the answer itself.
- Year over year: 3.1 per cent, matching September’s 3.1. We carried a 3.3 consensus, which asked for a rise of 0.2 and got nothing. A second reader puts the consensus at 3.2. So the miss is 0.2 on our figure and 0.1 on theirs.
- Month on month: 0.2 per cent, against a 0.4 consensus and a 0.1 prior. Asked for an acceleration of 0.3; delivered 0.1. Miss 0.2.
- Excluding energy: 2.0 per cent, after 2.1 in September — so an energy wedge of 1.1 points in the headline, and the wedge is where the whole story is.
This desk has written, repeatedly, that a Canadian consensus resolves per field and that the right expectation is a split verdict — one vendor right on the year and wrong on the month, the other the reverse. We restated it in writing before this release landed. There is no split. Both fields missed by exactly 0.2 points, in the same direction, which is a ratio of 1.0. The per-field framing is a real pattern and it did not apply this month, and a pattern that holds often enough to be written into a work order is exactly the kind that needs saying out loud when it fails.
The second half is quieter and it removes a trap rather than springing one. Our standing rule is that an unadjusted monthly change and a seasonally adjusted one are different series, that vendors do not say which they quote, and that the gap can swallow a whole surprise. The issuer gives both: 0.2 per cent month over month, and 0.2 per cent on a seasonally adjusted monthly basis. They are identical, gap zero. Which means a vendor’s unlabelled 0.4 consensus cannot be convicted of quoting the wrong basis this month, because both bases give the same answer. The ambiguity that usually costs you the surprise is unfalsifiable in exactly the month it does no damage, and you should not read that as the ambiguity having gone away.
The components, at the issuer: gasoline 30.9 per cent year on year after 27.8 in September but minus 2.5 on the month; groceries 2.0 after 1.7; transportation 8.9; shelter 1.4; clothing 1.8 after minus 0.3 in September, with a 3.4 per cent monthly rise the release itself calls the “largest monthly increase since September 2019”, a framing we record and do not adopt; cellular services minus 6.2 after plus 0.8. No prior month was revised in this release, which is itself worth recording given how many of our recent priors have moved under us. The next release is stated as Monday 14 December, with no time of day.
One errand closes here, at the fourth asking. United States October retail sales land at 13:30 UTC on Tuesday 17 November — 22:30 Japan time, on two independent readers — with a consensus of plus 0.2 per cent on the month from two sources. We have been printing that absence since Friday. It is now a time.
What This Does Not Tell You
- We never read the issuer for the November survey. Every figure above for November is a secondary reading. The New York Fed’s overview page served October and the document URL we could construct served a page-not-found at HTTP 200. The sub-index values could be wrong and we would not know.
- The two rising sub-indices are single-sourced. Employment at 18.1 and the six-month index at 28.9 rest on one publisher. They reproduce arithmetically against the issuer’s October baselines, which is a check on internal consistency and not on accuracy.
- Thirteen of the survey’s fields are simply missing. No November reading was published anywhere we looked for prices paid, prices received, delivery times, inventories, supply availability, unfilled orders, average workweek or capital expenditure plans. The October prices-paid index was 59.5 and we cannot tell you what it is now.
- We do not know whether 17.0 is a different consensus or a different survey panel. Two publishers carry it, two carry 17.6, one carries 18.0 and one carries 19.5. We have not established that all four refer to the same forecast set, and the same-object test is the one that should run before any vendor-disagreement argument.
- We cannot give you the Canadian core measures. The issuer’s release text does not carry CPI-trim, CPI-median or CPI-common in prose; they sit in a table we did not read. The Bank of Canada’s December decision will be argued about using exactly those three numbers.
- The ratio is not a constant and we are refusing to publish it as one. Eight last month, 3.68 to 4.68 this month, on the same series. Two observations.
- And we will not rank the forecasters. The 17.0 reading was closest this month by 0.6 points out of a 9.2-point miss. That is noise against the common error, and one print does not identify a better vendor.
Related
- Economic calendar and event times, in UTC · release clustering and the minute each print lands
- Current signals · how a release window is sized in practice
- Expert advisor presets · news-window and filter configuration