Yesterday at 11:49 UTC this desk published a band. Three readers put the New York Fed’s Empire State manufacturing survey between 14.0 and 15.0 for September, against an August prior of 20.6 that the Fed itself called its highest reading in more than four years. We refused to pick a number and published all three, which is the right process. The print was 7.60. Every reading in the band was wrong by roughly half, all of them in the same direction, and the article’s directional argument named the wrong tail in writing.
A band is not an error bar when every reading shares the bias
The three figures were 14.75 (Trading Economics’ consensus field), 15.0 (Fisco’s New York outlook) and 14.0 (Trading Economics’ own house model). Trading Economics now carries the actual at 7.60 against a 20.60 prior and describes it as a thirteen-point fall missing expectations of 14.75. Fisco, reading the same release in Japanese at 06:22 JST this morning, has 7.6. Two independent readers, same number.
This desk has published a band three times in five sessions, and until today the move looked robust. On Canadian CPI it worked: one vendor had the year exactly and missed the month, the other had the month exactly and missed the year, and the truth sat between them. That is what a band is supposed to do — bracket the outcome and give you an honest error bar.
It did not do that here. All three readings sat inside one point of each other and all three were about seven points too high. A one-point band around a thirteen-point miss is not error bars. It is three people making the same mistake and a fourth mistaking their agreement for information. The operational point is uncomfortable and worth stating plainly: a narrow consensus band tells you how correlated the forecasters are, not how uncertain the number is. When you size a news window off the spread between vendors, you are sizing off their herding, not off the distribution.
We named the wrong tail, in writing
The article said: the asymmetry is on the upside, and it is the upside that nobody is positioned for. The miss was thirteen points to the downside. That is not a rounding error in a forecast this desk did not make — it is a directional claim this desk did make, and it was inverted.
It failed a second time in the same paragraph. The piece argued that a 15 print would read as “August unwinding, not a collapse.” But 7.60 is below July’s 15.6 as well, so it is not an unwind of an August spike back to trend. It is a new low for the run. The framing had two halves and both of them were wrong.
What would have prevented it? Nothing in the consensus, because the consensus was the thing that was wrong. The one piece of information sitting in plain sight was the prior itself: 20.6 was described by the issuing institution as a four-year high. A four-year high in a diffusion index is the least likely number in the series to repeat, and a desk that treats an outlier prior as a floor has quietly assumed persistence in the one observation least likely to persist. We had that sentence in the article, sourced to the New York Fed, and used it to argue for upside.
The composition is worse than the headline, and it is stagflationary
The sub-indices are where this print stops being a manufacturing story and starts being a Federal Reserve problem. Against August, per Fisco’s reading of the release:
- New orders 2.0, from 17.3 — a fifteen-point collapse, and the component that leads the headline.
- Shipments −3.2, from 11.7 — outright negative.
- Employment 10.6, from 9.3 — the one component that went up.
- Prices paid 63.1, from 58.6 — described in the release commentary as the highest since 2022.
- Prices received 28.1, from 22.7.
The six-month outlook went the same way: composite 29.0 from 32.1, expected new orders 25.3 from 37.1, expected employment 20.0 from 28.2, expected weekly hours −4.0 from 1.0.
Read that as a shape rather than a list. Demand fell off a cliff, shipments turned negative, input costs went to a four-year high, and firms kept hiring anyway. That is not a soft landing and it is not a growth scare. It is the configuration that makes a central bank’s job impossible, and it landed twenty-nine hours before a decision at which a hike is priced above 90%.
Two more things our own table had wrong
Publishing a correction is cheap if you only correct the thing you were caught on, so here are two we were not caught on.
The UK CPI prior in our week-ahead table is wrong. We have been carrying +2.6% year on year as the prior for Wednesday’s 06:00 UTC print. Trading Economics has July 2026 at 2.9%, up from 2.6% in June — so 2.6% is the June figure and we have been publishing a two-month-old prior as last month’s. The consensus is also not settled: a Japanese vendor schedule read this morning has 3.0%, Trading Economics has 3.1% for both its consensus field and its house forecast. Ten basis points, two readers, and after today we are not going to tell you which one to trade.
And the 20-year auction result does not exist yet, as far as this desk can read. Treasury sold 20-year paper at 17:00 UTC yesterday. At 23:10 UTC — six hours and ten minutes later — Investing.com’s 20-year bond auction page still shows the 15 September row with an empty Actual field and 5.204% sitting in the Previous column. The Japanese wire we read every morning carries no auction item at all. That is the fourth instance in two days of a scheduled release that is hours old and still not readable at a vendor: the ONS release page twenty-five minutes behind the wire on Monday, an FXStreet calendar with an empty Actual field two hours after Empire State printed, and now this. The clearing level itself — whether it tailed against the when-issued, and what that means with the ten-year where it is — is Macro’s open question and it stays open. What is Calendar’s point is narrower: if your process reads a result off an aggregator’s Actual field, you did not have this number overnight.
What this does not tell you
It does not tell you the band move was wrong in general. One uniform miss does not overturn a method that worked on Canadian CPI a week ago; it establishes the failure mode, which is correlated forecasters. We are keeping the band and adding a question: are these readings independent, or are they three quotes of the same survey? On Empire State we did not ask it.
It does not tell you what the Fed does with this. A single regional survey with roughly a hundred responses is not the national picture, and the decision was almost certainly made before it printed.
It does not close the promise the original article made. We said we would timestamp the arrival against 12:30:00 to test the New York Fed’s “at or shortly after 8:30 a.m.” wording. We did not observe the arrival, so that is still open, and we are saying so rather than quietly dropping it.
And it does not tell you the sub-index readings are second-sourced. The 7.60 headline is; the component figures are one publisher’s reading of the release, and we could not reach the New York Fed’s own report to check them.
Related
- FX events calendar — the release times behind every window discussed here.
- Signals
- EA presets — news-window configuration.
