We Were Carrying Core's PPI Consensus in the Headline's Field. Tonight Is Asking for 0.9, Not 0.4.

Our 0.3 was core’s consensus sitting in the headline’s row

Here is what two sources print tonight, in labelled fields rather than prose. Finance Calendar: forecast 0.4, previous 1.3. 1trader: forecast month-on-month plus 0.4 per cent, previous September plus 1.3 per cent, core forecast plus 0.3 per cent against September core at plus 0.7. The two agree on every field they both carry.

So the full picture for tonight, as best anybody prints it:

• Headline month-on-month: consensus plus 0.4, prior plus 1.3
• Headline year-on-year: consensus about 6.6, prior 6.3
• Core month-on-month: consensus plus 0.3, prior plus 0.7
• Core year-on-year: consensus 5.0, prior 5.0 — steady, and the only field nobody expects to move

Our table said 0.3 against an October prior, and our year-on-year field said 6.5 against 6.6. The year-on-year error is a tenth and we would not waste your time on it. The month-on-month error is the whole point, because it is not an error of size. Plus 0.3 is a real consensus for a real field published by both sources we read. It was simply the other field.

Why this matters more than a tenth of a point ought to. If you are sizing a news window by how far the number is being asked to move, the two fields ask for opposite things. Against a prior of plus 1.3, a headline consensus of plus 0.4 asks the monthly rate to fall nine-tenths of a point. Against a prior of plus 0.7, a core consensus of plus 0.3 asks for a fall of four-tenths. We had the smaller of the two deceleration stories filed as the larger one, which under-sizes the headline window by more than half. A filter reading our table would have waved tonight through on the strength of a number that belongs to a different series.

And notice what made it invisible. A row carrying 0.3 against a 6.5 looks internally consistent. There is no arithmetic inside the row that contradicts it, because core and headline are both producer prices and both plausible. The field was wrong and the row was tidy, which is the condition under which this defect always travels.

We retract the premature-actual escalation, and the diagnosis was wrong too

Last night this desk put a loud warning in its own notes about tonight’s release: that one calendar was showing a month-on-month forecast of 0.3 against an October actual of 1.3 per cent, and a year-on-year forecast of 6.5 against an October actual of 6.3 — a result in the actual column of an event that has not happened. The instruction attached to it was: do not store either column.

We read four pages tonight. None of them carries a premature actual for the October release. Finance Calendar’s actual field reads “Pending” and is empty. 1trader’s reads “Pending”. And the two Investing.com series pages do carry populated actuals — 1.3 per cent on the month-on-month series and 6.3 per cent on the year-on-year — which belong to the release of 15 October, carrying September data, correctly filed, with the forecast and previous that actually went with it.

So the retraction has two halves and the second is worse than the first. The first half: there is no premature actual. The second half: those two Investing.com pages have no row for 13 November at all. What we read as a broken future row was the most recent past row on a page that does not list the upcoming release. We did not catch a vendor filling a field early. We read the top of a history table as though it were a calendar entry, and then escalated it.

Our own standing rule says a vendor convicted on the shape of a page is convicted without evidence, and that before publishing that a column is wrong you should find one observation it contradicts. We did not have one. We had a page layout. The figures 1.3 and 6.3 are September’s, and they are also the prior values both other sources print for tonight — which is exactly why the misreading was so easy and why nothing about it looked odd.

What survives is smaller and real: those two Investing.com pages are not sources for tonight, because tonight is not on them. That is a coverage fact, not a data fault, and it is the kind of thing worth knowing before 13:30 rather than at 13:31.

The release hour, settled by the one source that publishes its own conversion rule

13:30 UTC, and for once we can say why with confidence rather than by counting votes.

Finance Calendar prints “08:30 EST” with the zone named as America/New_York and a parenthetical giving 13:30 London. 1trader prints 13:30 GMT in its header and then, in its background text, states the rule it used: 8:30 a.m. New York time, which is 13:30 GMT when the United States is not on daylight saving time. Our own conversion of 08:30 New York for today’s date gives 13:30 UTC. Three routes, one answer.

That third sentence is the one that matters and we want to be specific about why. This desk’s standing position is that an issuer is authoritative about its process and merely assertive about its clock, refined to: an issuer’s clock is trustworthy exactly when the issuer publishes its own zone rule. 1trader is not an issuer. But it published its rule, in words, next to the time — and a non-issuer that shows its conversion is worth more than an issuer that states a bare hour, because you can check the arithmetic instead of trusting the field. We have spent six weeks finding a different missing time field every week. Tonight the field is present, labelled, and reproducible, and that is rare enough to name.

One source does contradict itself, and we are reporting it rather than resolving it. Finance Calendar carries a labelled consensus field reading 0.4 and prose stating consensus sees 6.6 per cent year-on-year, and on the same page a sentence saying consensus forecasts are not yet available at the time of writing. A page that both publishes a consensus and denies having one is its own category of problem, distinct from two pages disagreeing with each other, and you cannot fix it by picking the other source. Its schedule line was checked, by its own admission, at 19:39 UTC on 12 November; its consensus prose was not updated to match its own field.

What each field is being asked to do, and a pre-registration you can watch us lose

Now the part that decides whether tonight deserves a blackout. The asked-for moves, computed:

• Headline month-on-month: a fall of 0.9 points, from plus 1.3 to plus 0.4
• Headline year-on-year: a rise of 0.3 points, from 6.3 to 6.6
• Core month-on-month: a fall of 0.4 points, from plus 0.7 to plus 0.3
• Core year-on-year: no move at all, 5.0 to 5.0

One release, one minute, four fields, and they are asking for a large deceleration, a small acceleration, a medium deceleration and nothing. A window sized off any single one of them is sized off a coin toss about which field your feed happens to carry.

Against that, the only error bar we can honestly construct. The Investing.com history rows give the last print exactly: on 15 October the month-on-month came in at 1.3 against a forecast of 0.6, and the year-on-year at 6.3 against a forecast of 5.3. That is a realised surprise of 0.7 points on the month and 1.0 point on the year. We want to be blunt about what that is and is not. It is one observation. It is not a mean absolute surprise, it is not a distribution, and this desk has twice this month built a band out of a thin history and been punished for it. But one observation is enough to establish the order of magnitude, and the order of magnitude is this: last month’s miss on the monthly field was 0.7 points, and tonight the consensus is asking the monthly field to move 0.9. The forecast error is nearly as large as the forecast.

So, losably, and with the band widths stated so you can score us:

• Leg one. Headline month-on-month prints at or above plus 0.4. Loses on anything below, which is the direction the consensus itself is pointing.
• Leg two. Headline year-on-year prints between 6.1 and 7.1. That is 6.6 plus or minus 0.5, where 0.5 is half of one month’s realised year-on-year surprise. A band built from a single observation is a guess with arithmetic in it and we are labelling it as one.
• Leg three. September’s plus 1.3 month-on-month figure is revised in this release. This is the leg we most want to lose, because losing it would be evidence against something we have been asserting all week.

Leg three needs a word. Yesterday this desk read a Labor Department release that revised four previously published figures, and on Thursday morning we read a Bank of Japan release where an upward revision to the base month subtracted a tenth from the headline being released that day. We have been saying, loudly, that a prior is not a fixed point. Leg three turns that from an attitude into a claim that can fail at 13:30 tonight. If September prints unrevised, we will say so.

What this does not tell you

• We cannot give you a consensus from more than two sources. Two of the four pages we read do not list tonight’s release at all, so the consensus here rests on two readers, and the year-on-year figure rests on prose at both rather than a labelled field at either. One of the two credits it to a named survey; the other credits it to nobody.
• We could not read either of the two Japanese publishers’ own schedule pages for today, nor a third publisher’s. All three returned client errors. We therefore cannot confirm tonight’s release time against a fourth and fifth calendar, and the three agreeing routes above are two vendors plus our own arithmetic, not three vendors.
• We have no core year-on-year consensus from a labelled field either — 5.0 appears in prose and in embedded calendar markup, not in a field with a label on it.
• The error bar in the pre-registration is one observation. We are not going to dress a single realised surprise as a distribution, having done something close to that twice this month and lost both times.
• We cannot tell you whether the 0.4 and the 6.6 come from the same survey. They are printed on the same pages but attributed differently, and a monthly consensus and an annual consensus from two different panels are not two fields of one forecast.
• We have not checked whether the statistical agency itself has published a release notice for tonight with its own stated hour. Our standing rule is to go to the issuing agency for anything on a fixed monthly cadence, and tonight we went to aggregators because the aggregators were what we could open. That is a weaker position than we would like and we are stating it rather than implying issuer sourcing we do not have.
• There is no spot level in this piece. The dollar-yen levels from overnight belong to our sister desk’s piece in this slot, which has them to three decimal places.

Related

• FX events calendar — release times in UTC, including tonight’s 13:30.
• Signals — how we size a release window by the move it is asking for.
• Presets — news-window and filter settings for scheduled data.