Yesterday morning this desk published that the Institute for Supply Management’s own page had the wrong timezone on its own report, and that an aggregator’s calendar had it right. The report has now landed, and we were right about the time and wrong about the conclusion. That is the smaller of two findings. The larger one is that US initial jobless claims beat their consensus by three thousand, and the gap between two published values of the actual — from the same publisher, on the same release — is also three thousand. If your system computes a surprise from that row, the disagreement about what printed is exactly as large as the thing you are measuring.
The time was right. The diagnosis was too confident.
We said the September ISM manufacturing report would land at 14:00 UTC, that the aggregator calendar carrying that time was correct, and that the Institute’s own page describing a 10:00 a.m. EST release was an hour wrong on 1 October. The first two held. FXStreet’s report of the print is stamped 14:06:10 GMT, six minutes after the hour, and the Institute’s own press release is stamped 10:00 ET.
Read that stamp again, because it is the correction. The press release says ET. Not EST. The same institution, on the surface that actually carries the release, uses the season-proof label and gets it right — and tells you the October report lands on Monday 2 November at 10:00 a.m. ET, which is also right. What we published yesterday as “the issuer is the one with the wrong timezone” is more precisely this: one institution runs two timezone labels on two surfaces, and the wrong one is on the page a human reads while the right one is on the page the wire picks up.
That is a worse problem than an issuer being wrong, not a better one. A rule of the form “trust the issuer” cannot be applied to an institution that disagrees with itself, and you now have to know which of its pages you are on.
Two published actuals, three thousand claims apart
Initial jobless claims for the week to 26 September printed at 12:30 UTC. Minkabu gives 197,000 against a 200,000 forecast. Fisco’s market-reaction note, timestamped 21:57 JST, gives 197,000 against 200,000. Fisco’s own New York session summary, published at 04:13 JST this morning, gives 194,000 against 200,000.
So the actual is 197,000 at three readings and 194,000 at one, and the one is the later article from a publisher that had already printed the other figure. We are not going to adjudicate it beyond saying that 197,000 has three readings and 194,000 has one. What we will do is the subtraction, because it is the whole point: 197,000 against a 200,000 consensus is a three-thousand beat, and the distance between the two published actuals is three thousand. The measurement error on what happened equals the signal.
This is the same-publisher check earning its keep again. If you had read only the New York summary — the natural thing to read at a Tokyo desk, because it is the one published overnight in your own timezone — you would have logged a six-thousand beat on a release that beat by three.
And the prior moved underneath both of them
Minkabu’s row carries something the reaction notes do not. The previous week’s initial claims figure is given as 198,000, revised from 197,000. Continuing claims are 1,701,000 against a prior of 1,712,000, revised from 1,719,000.
That 1,719,000 is the number this desk published in its own week-ahead two days ago. It no longer exists. On Tuesday we gave the revision hazard a magnitude for the first time, on non-farm payrolls: the prior was revised in three of three checkable months by a mean absolute 41,300. We framed it as a monthly-survey problem. Two days later it has fired on a weekly administrative count, in both series, in the same release, by 1,000 and 7,000. Weekly claims are supposed to be the clean one.
The consequence for a filter is mechanical and it is not about accuracy. A surprise is a difference between two numbers, and on this row both of them move after you have stored them.
Our own correction was wrong by the same amount, in the other direction
We published a claims consensus of 199,000 in our week-ahead. At the pre-US slot yesterday we corrected it to 201,000, on two readers that agreed exactly, and said in print “we were the outlier and we were wrong.” Three print-day readings give the consensus as 200,000.
Our original figure was 1,000 low. Our correction is 1,000 high. We moved the same distance to the other side of the number and called it a fix. The correction was still worth making — two agreeing readers is better evidence than one — and it still did not land on the figure the print was judged against. There is no version of this where the desk comes out well, so we are writing it down rather than leaving it in a log.
The useful generalisation is not about us. It is that a consensus has now had four published values on one release — 199,000, 200,000, 201,000 and whatever the absent fourth reader held — across a two-thousand span, on a print whose own actual is ambiguous by three thousand. On Tuesday we argued that chasing consensus precision on payrolls resolves about nine per cent of the uncertainty. This row makes the same argument at one twentieth of the scale.
Thirty-five hours on, one calendar still presents a finished release as a future event
The ADP employment change for September was released at 12:15 UTC on 30 September and three readers carried a figure that afternoon. At 23:30 UTC today, thirty-five hours and fifteen minutes later, Investing.com’s release history for that indicator shows the row with a date, a 12:15 release time, a previous value of 38.00K — and both the actual and the forecast cells empty. Every earlier month in the same table is fully populated.
We recorded this cell as blank at eleven hours, at eighteen hours, at twenty-three and a half, and now at thirty-five. It has not filled. It has got worse: the forecast is gone too, so the row no longer reads as a release awaiting its number. It reads as an event that has not happened. A reader coming to that page cold would conclude exactly that, and would be wrong by a day and a half.
The pairing with the ISM finding above is the whole lesson of this slot. The aggregator had the ISM time right and the issuer’s own public page had it wrong; the aggregator has had this actual missing for a day and a half while every wire has it. Judge a source per endpoint. Not per publisher, and not per institution either.
What the ISM print itself was, for the record
54.5 for September against 54.6 in August. Our published consensus readings were 54.8 at two publishers and 55.0 at two others; the print came in below all four, between 0.3 and 0.5 below, which is one and a half to two and a half times the whole vendor disagreement we had logged. New orders 55.3 from 53.7, production 56.7 from 58.3, employment 52.7 from 51.2, prices 77.9 from 71.1. The prices index is the only move worth a second look, and it is not a small one.
Our Calendar desk wrote yesterday that no plausible ISM figure in that neighbourhood changes a rate view. 54.5 is in that neighbourhood and the dollar went up anyway, on crude and the euro rather than on the survey. We will call that claim intact and note that it was not tested hard.
What this does not tell you
We could not settle the third item we came for. This desk has published twice that one publisher’s US session preview carried every timed row sixty minutes early, and said in print that a third instance would be the condition for calling it a pattern rather than two bad pages. We looked for that publisher’s preview of today’s session and could not find it. That is not evidence it does not exist — it is a search coming up empty on a page that may simply not be indexed yet. The item is open, not closed, and not advanced.
The 194,000 figure is single-sourced and we have not established how it arose. A transposition, a different reference week, a revision read from a later table and a plain error are all live, and we are not guessing between them.
The ISM timezone finding rests on two of the institution’s own surfaces plus one wire timestamp. One print does not establish what any of those pages will say in November, when the clock change makes the EST label correct for the first time in seven months. That is a date we have now committed to three separate predictions about, and it is a month away.
And nothing here is a view on the dollar. Every number above is about whether a row in your calendar is a row of facts. On the evidence of one Thursday, it is a row of estimates with timestamps.
Related
- FX events calendar — the release schedule, in UTC, with the times we actually hold.
- Signals — how a news window gets built around a release time.
- EA presets — news-filter configuration, including what to do when a source disagrees with itself.
