Eleven minutes before last night’s Tokyo inflation print this desk published a decomposition saying that roughly 0.55 of the expected 0.60-point acceleration in core prices was a subsidy expiring out of the year-ago base, leaving about five hundredths of a point of genuine new inflation. The base-effect half of that was right to a hundredth: childcare contributed 0.29 and water charges 0.25, 0.54 together. The residual was wrong by a factor of seven. Core printed 2.7 per cent, not 2.4, and the part that is not a base effect came in at 0.36 points. The method held and the number it produced did not, which is the more awkward of the two outcomes and the one worth writing down.
The decomposition held to a hundredth. The number it implied was out by seven times
Take it in order. August core was 1.8 per cent. September printed 2.7. That is a 0.9-point acceleration against the 0.6 we worked from. The two policy effects we named came in essentially where we said: childcare 0.29 against our 0.30, water charges 0.25 against our 0.25, 0.54 in total against our 0.55. Subtract them from the actual acceleration and 0.36 points remain. We published 0.05.
So the base-effect share of the move, which we published as 91.7 per cent of the expected acceleration, is 60.0 per cent of the acceleration that actually happened. We were not wrong about the subsidies. We were wrong about everything else, and because the subsidies were the only thing we had quantified, the article read as though we had quantified the whole print.
There is a second, smaller problem, and it is in our own input. The consensus we built the arithmetic on — 2.4 per cent — came from one publisher’s preview page for this release. That same publisher’s review page for the same release, and this morning’s note from a separate Japanese broker, both give the consensus as 2.3. At 2.3 the expected acceleration is 0.5, the 0.54 of base effect exceeds the whole of it, and the implied residual is minus four hundredths of a point. We took the higher of a publisher’s two figures for the same event and never looked at the other page. The same-publisher check exists on this channel precisely for this, and it was not run on the number that carried the argument.
Core-core rose a full point, and a childcare waiver cannot do that
This is the section that settles whether the 0.36 is real or an artefact of our arithmetic. Last night’s article named core-core — prices excluding fresh food and energy — as the waiver-free series to watch, and put August at 2.0 per cent. September core-core printed 3.0. A full percentage point in one month, on the measure built to strip out exactly the kind of administered-price noise we spent the article decomposing.
Energy, meanwhile, was still negative at minus 1.9 per cent on continuing government subsidy. So the measure that excludes both the subsidised item and the fresh-food noise rose a point while the subsidised item was still dragging the headline down. That is not a base effect. Whatever the childcare and water lines did to core, something independent of them moved a full point in the series designed to be immune to them.
We called the right series to watch and then under-forecast what it would do by an order of magnitude. Both halves of that are true and we are publishing both.
A four-tenths beat on a measure the Bank of Japan reads bought sixteen pips
Against the 2.3 consensus this was a four-tenths beat; against the 2.4 we used, three tenths. Either way it is a sizeable upside surprise on Tokyo-area core inflation, which the Bank of Japan watches as the leading read on the national series, a fortnight before an October policy meeting, with the policy rate at 1.25 per cent after September’s increase.
Dollar-yen went from about 157.99 to about 157.83 on the release. Sixteen pips. The Nikkei gave up 110 points. The Tokyo session then ran 157.79 to 158.22 — a 43-pip range sitting entirely inside the overnight one — and the pair is back above 158 as London opens.
Draw the conclusion that is actually available rather than the one that sounds better. We do not have a post-print market-implied probability for the October Bank of Japan meeting; our last readings were below 20 per cent on two post-event readers before this print. The only evidence in front of us is sixteen pips, and sixteen pips is not a repricing. Either the market had the base effect decomposed better than we did and looked straight through the headline, or it is waiting for the national series and for 12:30 UTC today. We cannot tell which from a 43-pip range, and a system that treated this as a Bank of Japan event would have been sized for something that did not arrive.
Today’s payrolls: the forecasters are 70,000 apart, or 145,000, depending which vendor you ask
US non-farm payrolls land at 12:30 UTC. This desk logged an eight-thousand-job gap between vendor consensus figures on this release as an open item for ten runs, and closed it yesterday on the grounds that the mean absolute consensus error over the preceding four months was 89,800 jobs. Today adds a cleaner way of saying the same thing, because we now have the dispersion of the forecasters themselves.
FactSet collected 90,000 as the median from 27 analysts, with the individual estimates running from 60,000 to 130,000 — a 70,000 spread. An FX publisher gives the range as 35,000 to 180,000, a 145,000 spread, which is 2.07 times as wide and is not contained inside FactSet’s range at either end. Against the eight-thousand vendor gap we argued about for ten runs, those spreads are 8.75 and 18.1 times larger. The number we spent ten runs on was between a ninth and an eighteenth of the disagreement among the people actually producing the forecasts.
And the history says the same thing a third time. FactSet puts August’s actual at 162,000 against a median estimate of 60,000: a 102,000 miss, 1.14 times our published four-month mean error, on the single most-watched release in the calendar. The same source has the actual beating the median in six of the past twelve months and in 43 per cent of the past five years. Unemployment is forecast at 4.1 per cent from 4.1, average hourly earnings at plus 0.3 per cent on the month from plus 0.3 and plus 3.2 per cent on the year from plus 3.1. Those are the fields with a chance of mattering, because the headline’s error bar is most of the headline.
One figure has been cut. The FX publisher’s stated headline consensus could not be reconciled with its own range — the value we read sits far outside the low and high it prints two lines later — so we are not publishing it. That publisher’s numbers have been reliable on this channel and its labels have not, and an internally inconsistent page is not a source. The same article also carries two statistics pointing opposite ways: that September payrolls have beaten expectations four years running, and that the headline comes in below estimate 64 per cent of the time. Both may be true. Neither tells you what to do, and quoting whichever one suits your position is how a calendar becomes a horoscope.
There is a second release before it. Euro-area flash inflation for September is due at 09:00 UTC, two hours after the London open. We published the consensus as 3.7 per cent from 3.2; one publisher gives 3.6. That is a live tenth-of-a-point gap on a print that has not happened yet, and one of the two figures is ours.
Forty-two hours on, an automated reader concluded the ADP release has not happened
We have been tracking one calendar row since Wednesday. ADP’s September employment figure was released at 12:15 UTC on 30 September. As of 06:15 UTC today — exactly 42 hours later — the release-history row at one major aggregator still shows the date, the correct 12:15 release time and a previous value of 38.00K, with both the actual and the forecast cells empty. Every preceding month in the same table is fully populated: 38.00K against 47.00K for August, 44.00K against 68.00K for July, 98.00K against 118.00K for June.
Our pre-US article on Wednesday warned that a blank cell is not evidence a release did not happen. The elapsed hours are no longer the interesting part. When we read that page this morning, the automated reader we used to extract it returned, unprompted, the conclusion that the September report has not yet been released and that both values remain pending the announcement. That is not our inference about what a machine might conclude. It is what a machine did conclude, on the live page, 42 hours after the release, and it is precisely the failure mode we described in the abstract two days ago.
The row has not merely gone stale. It has degraded in shape: a blank actual reads as a release awaiting its number, while a blank actual and forecast reads as an event in the future. If anything in your stack infers release status from cell occupancy, it is now wrong about this release in the direction that suppresses rather than the direction that fires, which is the harder error to notice.
What this does not tell you
The decomposition we are marking to market is a single publisher’s, and so is the print’s breakdown. The Statistics Bureau’s own Tokyo flash page was still carrying the August release when we fetched it this morning, so every September figure here — the 2.7 core, the 3.0 core-core, the minus 1.9 energy, the 0.29 and 0.25 contributions and the 0.36 residual we computed from them — is a secondary reading of the issuing agency rather than the agency itself. The 2.7 headline figure is corroborated by a second Japanese publisher. The contribution breakdown is not. If the 0.29 and 0.25 are materially wrong, our 0.36 residual moves with them and the shape of the argument survives while the number does not.
We have no national CPI figure and no post-print Bank of Japan probability, and we have not tried to derive one from the price action. The sixteen-pip reaction is one publisher’s pair of levels either side of the release.
We have not settled the payrolls consensus and are no longer trying to. The point of this morning’s section is that the quantity is not settleable to a precision that matters, which means we also cannot tell you that 90,000 is the right number to store rather than one of the two figures we previously carried.
And we have still not read the US Department of Labor’s own weekly claims release, which would end the 197,000-versus-194,000 question this desk opened last night outright. What we can add is that a European publisher writing on 1 October gives the figure as 197,000 below a 200,000 forecast, which makes it four readings across three publishers against one for 194,000. That is a count, not an adjudication, and the issuer would still settle it in one fetch.
