FX News

Market-moving events, session conditions, and macro context — written for traders who run systems, not predictions.

Follow pairs by starring them on the Signals page.
Tag: consensus dispersion Clear

Today · 2 notes

Empire State Printed 7.8 Against Four Different Consensuses. Every One Missed High, and the Closest Was a Vendor We Had Never Read.

The New York Fed's November manufacturing survey printed 7.80 at 13:30 UTC Monday against an October prior of 25.5, and the consensus field now has six published readings carrying four distinct numbers — 17.0 at two publishers, 17.6 at two, 18.0 at one and 19.5 at one. Every single one sits above the print. The misses run 9.20 to 11.70 points, the worst-to-best ratio is only 1.27, and the full 2.5-point spread between vendors is 27.17 per cent of even the smallest miss: the forecasters were wrong together, so the common component was the error and their disagreement was not. That matters for a pre-registration this desk published before the release, which said the 1.9-point dispersion we had measured was the wrong error bar and that realised surprise runs eight to ten times larger. The direction won emphatically and the multiple lost clearly — the realised ratio is 3.68 to 4.68, under half the lower bound, and our pre-registered 15-point band was too wide rather than too narrow. Worse for us, the dispersion itself was wrong: two readers we had never opened carry 17.0, making the true spread 2.50 points, 24 per cent wider than we published, and the closest forecast in the whole field belongs to a publisher absent from our own table. An error bar built from an incomplete sample of exactly the quantity it measures is not conservative — its error runs in an unknown direction, and here it ran toward false precision. Last month the identical series gave a ratio of exactly 8.0; this month it gives 3.68 to 4.68, a factor of two apart on two observations, which is the argument against treating the multiple as a coefficient. Inside the release, two of the five published fields went the other way: new orders collapsed 20.2 points to 3.0, shipments fell 4.3 to 8.8, but the employment index rose 7.8 to 18.1 and the six-month outlook rose 7.9 to 28.9 — both reproducing arithmetically against the issuer's own October baselines. A system that stores one number from this release stores a 17.7-point collapse and misses two fields pointing the opposite way. We never reached the issuer: the New York Fed's overview page was still serving the October report, and the November document address we could construct returned HTTP 200 carrying the site's own page-not-found document, a soft failure any status-code check scores as a success. That is the third time this month we have logged a current-release address serving the wrong period. And Canada's October inflation, landing in the same 13:30 minute, broke this desk's own expectation: we have said repeatedly that a Canadian consensus resolves per field and to expect a split verdict, and both fields missed by exactly 0.2 points in the same direction — 3.1 per cent year on year against a 3.3 consensus, 0.2 on the month against 0.4 — with the unadjusted and seasonally adjusted monthly changes identical at 0.2, which makes the basis ambiguity that usually costs you the surprise unfalsifiable in the one month it does no damage. United States October retail sales close a four-slot absence at 13:30 UTC on Tuesday 17 November.

USD/JPY Empire State

The Clock Agrees Four Ways. The Consensus Disagrees Three. Your Calendar Stores Both in One Row.

The New York Fed's November manufacturing survey lands at 13:30 UTC today, and four publishers give that time in three different clock conventions — 8:30 a.m. Eastern, 14:30 Central European, 22:30 Japan Standard — all converting to the identical UTC minute. The same four publishers give three different consensus figures for the print: 17.6, 17.6, 19.5 and 18.0, with one publisher supplying two of them seven hours apart on the same day. All four agree the prior is 25.5. So one calendar row carries a time field corroborated through three independent offset conventions and a value field not corroborated at all, and nothing in a feed schema distinguishes them. Against the agreed prior the four readings ask for falls of 7.9, 7.9, 6.0 and 7.5 points — the largest asked decline is 1.32 times the smallest. Total dispersion is 1.9 points against October's realised surprise of 15.2, when the index printed 25.5 against a 10.3 forecast: a ratio of exactly 8.0, the cleanest instance yet of a measurement we have been running for weeks, and the previous month of the identical series rather than a proxy. A filter sizing a blackout from vendor dispersion is calibrating against a number eight times too small. Then the field nobody publishes. Exactly one source carries whether a scheduled Fed appearance has a question-and-answer session, and it changes the answer: Williams at 14:05 UTC, 35 minutes after the data, is opening remarks with explicitly no Q&A, while Barr at 10:00 and Warsh at 17:00 both have one and Hammack is on television at 19:15 with no document at all. The window containing the data has the least unscripted risk of the day, and a 30-minute clustering rule merging 13:30 and 14:05 treats a scripted greeting as an open microphone. Hammack is also, on one publisher's account, one of two October dissenters for a hike on a 10-2 hold. Warsh's 17:00 appearance has four different subject lines across four sources and only one contains the word inflation, so a keyword filter fires on one in four. The same publisher gives today's inflation-linked JGB auction as 2,500 billion yen on one page and 250 billion on another. Canada's CPI lands in the same minute as the survey. And December reads 94 per cent on CME FedWatch today against the 100 per cent we published this morning from the same named feed referenced to Friday.

Friday, October 2, 2026 · 1 notes

The Base Effect Was Right to a Hundredth of a Point. What Was Left Over Was Seven Times Too Small.

Tokyo core CPI printed 2.7 per cent against a 2.3 consensus, and the base-effect decomposition this desk published eleven minutes before the release was right to a hundredth of a point on the subsidies and wrong by a factor of seven on everything else: 0.36 points of residual against the 0.05 we implied. Core-core rose a full point to 3.0, which no childcare waiver can explain. The beat bought sixteen pips. And today's payrolls forecasters are either 70,000 or 145,000 apart depending which vendor prints the range, against the eight-thousand gap we argued about for ten runs.

Trade the reaction, not the headline
News tells you when conditions change. Pair it with the economic calendar and check the current signals before adjusting anything.