Eleven hours ago this desk published that four independent readers carried the October payrolls consensus at 65,000 to the thousand, and argued that the unanimity was worth nothing because the readers were not four surveys. We were right about the second half and wrong about the first. A page published the day before ours says economists expect 59,000, and it names six houses whose individual forecasts run from 50,000 to 95,000. The unanimity we spent a section dismantling did not exist in the first place, and the evidence against it was already on the web when we wrote it.
The page that broke our claim is older than our claim
The figures, with the label each publisher actually prints:
- 65,000 — four readers as of 23:11 UTC yesterday, one of which attributes it to the median of Bloomberg contributors. A sixth reader carries 65,000 again this morning.
- 59,000 — a US personal-finance desk, published 5 November, under the label “Economists expect”. No survey is named and no producer is credited.
- 65,000 to 70,000 — a Tokyo market outlook timestamped 10:54 today, under the heading “Key market forecasts”, as a range rather than a point.
Six thousand jobs between the two point figures is not a lot. The failure is not the size of the gap, it is that we published a claim of exact agreement after a search that did not find a page carrying a different number, and that page was a day old. A negative claim about what every reader says is only as strong as the sweep behind it, and ours was not strong enough to carry the word “agree”.
There is a second label fault in the same batch, and it is the kind this desk has now logged six weeks running. One of this morning’s readers prints the 65,000 figure and the 4.2 per cent unemployment rate under the heading “U.S. September Nonfarm Payrolls”. September printed four weeks ago, at 29,000. A filter keyed to the month name in a consensus row would read today’s forecast as a stale September figure and ignore it. The number is right and the field it sits in is wrong, which is exactly the shape of the Canadian field-mislabelling we published in September and the euro-area consensus-column dispute we closed on Wednesday.
The houses are 45,000 apart, and the median nobody polled is in the middle of nothing
Here is the distribution the 59,000 page actually publishes, which is the most useful thing we have found this morning:
- Barclays, Marc Giannoni: 50,000, with unemployment falling to 4.1 per cent, and a three-month moving average through October of +71,000 against +28,000 for the three months through July.
- Deutsche Bank: 55,000 headline against the 29,000 prior, private payrolls 55,000 against 46,000, unemployment unchanged at 4.2 per cent, average hourly earnings +0.3 per cent against +0.1 per cent, three-month averages of 51,000 headline and 54,000 private.
- Wells Fargo: 65,000, unemployment holding at 4.2 per cent.
- RSM US, Joseph Brusuelas: a net gain of 95,000, unemployment 4.2 per cent.
Low to high is 45,000 on a consensus of 59,000 or 65,000. Yesterday we published that the named houses were 40,000 apart, with a high of 90,000; this morning’s named list has the high at 95,000 from RSM. We cannot settle from here whether that is a different house or the same figure transcribed differently in two places, and we are not going to pretend otherwise. The width is the point and the width got wider.
Two of the four disagree with each other about the direction of the unemployment rate on effectively the same payroll number — Barclays at 50,000 with the rate falling to 4.1, Wells Fargo at 65,000 with it holding at 4.2. If you are sizing off the unemployment print rather than the headline, the houses closest together on jobs are furthest apart on the thing you are watching.
And Barclays supplies the one figure in this batch that is a test rather than a guess: a three-month average of +71,000 through October against +28,000 through July. That is an assertion about a trend with an arithmetic consequence — it requires today’s print to be well above the recent run-rate. Deutsche’s +51,000 three-month average and Barclays’ +71,000 cannot both describe the same window, so at least one of them is anchored on a different three months, and neither page says which.
The prior we built yesterday’s argument on is also two numbers, and so is ADP’s
Yesterday we wrote that one month ago the same unanimity sat at 90,000, that September printed 29,000, and that the 61,000 miss was 68 per cent of the forecast it busted. The 59,000 page says September came in “well below a consensus forecast of 93,000”.
So the miss was 61,000 or 64,000, depending on whose September consensus you use, and our published 68 per cent becomes 68 or 69 per cent of the busted forecast. The correction makes our argument stronger, not weaker, and we would rather say that out loud than leave the discrepancy for a reader to find. What it does to the sizing arithmetic we published is minor: repeat the error in either direction from 65,000 and you land near 126,000 or 4,000 on the 61,000 version, or near 129,000 or 1,000 on the 64,000 version. Both still describe two opposite trades inside one month’s demonstrated error.
The ADP line has the same problem and it is sharper, because it is a revision question. On Wednesday we published that ADP printed 86,000 in October, that September was revised to exactly 86,000, and that the month-on-month change was therefore zero. The 59,000 page says October’s 86,000 was “down from 90,000 in September but above the 75,000 economists expected”. Read those together: the September ADP figure is 90,000 as originally published and 86,000 after revision, so the month-on-month change in private payrolls is zero or minus 4,000 depending on whether your feed carries revisions. And the October ADP consensus now has a fifth reading at 75,000, against the 67, 69 and 70 thousand we published.
This is the standing rule on this desk and it keeps paying: check whether the prior was revised before you quote a change. A zero and a minus four thousand are different stories about the same release, and both are correctly derived.
13:30 is two countries. 15:00 is four events, and one of them nobody is talking about
Today’s stack, in UTC, with the reader count behind each time:
- 12:15 — Huw Pill, the Bank of England’s chief economist, the day after an 8–1 hike. One reader, and see the caveat below on its time zone.
- 13:30 — US October payrolls, unemployment rate and average hourly earnings. Three readers agree.
- 13:30 — Canadian employment change and unemployment rate, the same minute.
- 15:00 — US preliminary University of Michigan sentiment, three readers; US preliminary University of Michigan inflation expectations, one reader; the Canadian Ivey PMI, two readers; and the New York option expiry cut, which this desk published this morning.
- 16:00 — the Bank of Japan governor at the IMF. One reader.
The item worth isolating is the inflation-expectations field at 15:00, and only one of our readers lists it separately from the sentiment headline. Yesterday the one-year US inflation swap fell about 25 basis points to a two-month low, on a crude collapse, and that move is the main reason December hike pricing came off. The preliminary Michigan inflation expectation is the first survey-based read on whether households saw the same thing, and it lands 90 minutes after payrolls, in the same minute as a Canadian PMI and an option cut. A window keyed to “consumer sentiment” will be open for it. A window keyed to the payroll release will have closed an hour earlier.
For sizing rather than direction: one desk puts the options market at about 1.49 per cent for today’s payrolls, quoting it at 14:30 CET, which converts to 13:30 GMT exactly and is a rare case of two notations on one page reconciling to the minute. One-week dollar–yen implied volatility is 7.78 per cent on a Tokyo reading, up from the low 6 per cent area yesterday — and this desk published 6.23 per cent and a move to the high sevens at 06:18 this morning, so that one is now closed in our favour.
The Canadian 6.5 per cent that is also 6.485 per cent
Two readers have the Canadian unemployment rate holding at 6.5 per cent. One of them, a research house writing on 25 October, says in its headline that unemployment should “remain at September’s 6.5 per cent rate” and in its body that it should “fall to 6.485 per cent from 6.53 per cent before rounding”.
That looks like a publisher contradicting itself and it is not. Both figures round to 6.5 per cent at one decimal place, so the headline and the body are the same forecast at two precisions. What the house is actually forecasting is a four-and-a-half basis point fall in the unemployment rate that is invisible in the field every calendar prints. We run this containment check before publishing two numbers as a dispute, and this one failed the dispute test, which is the correct outcome.
The rest of the Canadian picture: the consensus in our file is +12,200 against a prior of minus 68,300, with the same house carrying its own forecast at +20,000 and describing September’s decline as 68.4 thousand, more than fully explained by a 70,000 fall in public-sector employment. It expects a public-sector pause, private sector +15,000, self-employment +5,000, goods employment minus 10,000, private services +30,000 and a labour force up only 10,000 after a 52,900 fall in September. Average hourly wages for permanent employees are forecast to slow marginally to 2.3 per cent from September’s 2.3 per cent, against 2.0 in August, 3.7 in June and a year-to-date high of 5.1 in March.
That wage series is the Canadian field that has moved most and is quoted least. It has halved since March. A filter keyed to the Canadian headline employment change, in the same minute as US payrolls, will not see it at all.
What this does not tell you
We did not reach three of the readers we wanted. A FactSet payrolls page, a Morningstar jobs preview and an FXStreet Canadian employment preview all returned client errors. The FactSet page’s own title carries 65,000, and we are not publishing a figure off a title; the FXStreet headline carries the 6.5 per cent Canadian rate, which is why we describe that one as second-sourced by a headline rather than by a body. Four of this morning’s fetches failed, and every one of them was meant to be a corroborating read.
We cut a revision figure that did not reconcile. The 59,000 page carries a sentence on July and August payroll revisions whose numbers do not agree with its own stated total by roughly 220,000 in the wrong direction. We cannot tell whether that is the page or our reading of it, so no revision figure appears anywhere in this article. If the July and August revisions matter to your position, go to the Bureau of Labor Statistics release itself.
The 12:15 and the 16:00 are inferred time zones. The page carrying both prints “Japan time” explicitly on two items and omits the zone on the other two. Reading the unzoned items in the same zone puts the Michigan release at 15:00 UTC, which is exactly where two independent readers with stated zones put it. That reconciliation is why we trust the Pill and Ueda conversions; it is still a reconciliation and not a printed fact, and the Pill appearance is single-sourced either way.
The consensus is still not the print, and that has failed every week we have tested it. Nothing above tells you what 13:30 produces. The most defensible thing in this article is Barclays’ conditional about the three-month trend, because it can be scored. The least defensible is any of the point estimates, including the one four readers share.
And one release today already happened and nobody has priced it. Japanese September real wages rose 2.2 per cent year on year against a 1.3 per cent forecast, with a 1.6 per cent prior — a 0.9 point beat on a single reader, in a week when the Bank of Japan’s governor speaks at the IMF at 16:00. We have not second-sourced it and are flagging it rather than building on it.
Related
- FX event calendar — the 13:30 US and Canadian pair, and the four-event 15:00 including the Michigan inflation-expectations field.
- Signals — why a conditional about a trend can be scored and a point estimate cannot.
- EA presets — filters keyed to a month name in a consensus row, and what happens when the row says September.
