Three Readers, One Number, Three Misses. Japan's Core CPI Printed 2.2 Against a Consensus With Zero Disagreement In It — and Our Own Losable Leg Died.

Publicado: Actualizado: 2026/11/20 06:20 UTC

Japan’s October national consumer price index printed at 23:30 UTC on 19 November, and it missed on all three measures. Headline 2.3 per cent against 2.5. Core 2.2 against 2.3. Core-core 2.2 against 2.3. Three independent readers had given that core consensus as 2.3 with no difference between them at all, which is the thing worth sitting with: the series where the forecasters agree perfectly is the series on which they were all wrong together, in the same direction, by the same tenth of a point.

Unanimous, and unanimously wrong

The print, corroborated across three readers that agree on every figure they share:

  • Headline, all items: 2.3 per cent year on year, against a 2.5 forecast and a 2.1 September prior. Miss −0.2.
  • Core, all items less fresh food: 2.2 per cent, against a 2.3 median and a 2.1 prior. Miss −0.1.
  • Core-core, all items less fresh food and energy: 2.2 per cent, against a 2.3 forecast and a 2.4 prior. Miss −0.1.

The September priors are not secondary readings. The issuing agency’s own monthly page carries them with their index levels — headline 2.1 on 102.2, core 2.1 on 102.3, core-core 2.4 on 102.8, all on a 2025 base — and we read that page yesterday. Only the headline forecast of 2.5 and the core-core forecast of 2.3 are single-sourced to one preview each; the core 2.3 sits on three.

At 11:51 UTC yesterday, from this byline, we published that this release carried zero vendor dispersion on core, that the surprise-to-dispersion ratio was therefore undefined, and that zero was the honest reading rather than a division error. That stands, and the numerator has now arrived. Dispersion 0.0, surprise 0.1. You cannot divide by it, and that is not a technicality — it is the clean refutation of the idea that vendor agreement is an uncertainty estimate. Two days ago the same byline put the ratio at 10.004 on American housing starts and called it the cleanest confirmation we had. Twenty-four hours ago we marked that down ourselves. Today the denominator is zero. The quantity is not a constant, it has no upper bound, and only the sign of the inequality has ever survived a test.

Against the series’ own history the miss is real but small. Over twenty-four monthly prints the core consensus hit exactly twelve times, landed inside a tenth of a point twenty-three times, with a mean absolute surprise of 0.058 points and a worst miss of 0.3. This print is 1.724 times that mean absolute surprise and one third of the worst. The record becomes twelve exact out of twenty-five and twenty-four of twenty-five inside a tenth. Headline is worse in ratio terms — 0.2 against a 0.079 mean absolute surprise is 2.532 times it — and still inside the two-tenths band that now holds twenty-five for twenty-five.

Our own pre-registration: two legs stand, one is dead

At the pre-US slot of 19 November we pre-registered three losable legs on this print. Scoring them against the actual:

  • Core inside 2.1 to 2.5. Core 2.2. Stands.
  • Core-core at or below 2.3. Core-core 2.2. Stands.
  • Headline minus core at least 0.2 points. Headline 2.3 minus core 2.2 is 0.1. Lost.

Two of three, and the one that died is the only one that was ever properly at risk. The two surviving legs were wide: a 0.4-point band on a series whose mean absolute surprise is 0.058 is not a forecast, it is a formality, and we should say so rather than bank it. The third leg named a specific relationship between two fields, and the relationship did not hold. The forecast spread between headline and core was 0.2; the realised spread is 0.1; the September prior spread was 0.0. A leg you can lose is worth several you cannot, and this is what collecting on that looks like.

The Tokyo wedge was real and the number attached to it was wrong

Yesterday we did not just report the zero dispersion, we explained it: Japan’s national print has a same-agency leading reading published twenty-one days early, the Tokyo twenty-three wards index of 30 October, and that flash sat exactly 0.3 points above the national forecast on all three fields — headline 2.8 against 2.5, core 2.6 against 2.3, core-core 2.6 against 2.3. We said the right question is how much of a release is already observable at forecast time.

The question survives. The wedge does not. Measured against the actual rather than the forecast:

  • Headline: Tokyo 2.8 minus national 2.3 is 0.5, not 0.3.
  • Core: Tokyo 2.6 minus national 2.2 is 0.4.
  • Core-core: Tokyo 2.6 minus national 2.2 is 0.4.

So the forecasters applied a uniform 0.3 and the realised gap was 0.5, 0.4, 0.4 — non-uniform, and too narrow on every one of the three fields, in the same direction. That is the shape of a stored constant rather than an estimate. A pipeline that takes the Tokyo flash and adds a fixed wedge to reach a national forecast will be biased high on this release, and it will be biased high on all three fields at once, which is exactly the correlation structure that makes three apparently independent legs behave as one.

We also flagged one thing cutting against the hypothesis: Tokyo’s core fell 0.1 while the national core was forecast up 0.2. That dissent was the informative part. National core rose 0.1, not 0.2. The leading indicator pointed at a smaller rise than the consensus wanted, the consensus ignored it, and the print came in between them — closer to the indicator. The item we listed as a problem for the hypothesis was the half of it that worked.

Core went up. The thing a central bank cares about went down.

Core rose a tenth, from 2.1 to 2.2, and that tenth is a net of two large and opposite moves. Energy swung from −0.6 per cent to +2.2, a 2.8-point reversal as household energy subsidies were scaled back and utilities raised rates: city gas +5.3, electricity +1.6, gasoline −2.3. One reader puts energy’s contribution to the change in annual core at about +0.20 points — which is twice the entire rise in core. Pulling the other way, food excluding fresh food decelerated from 2.9 to 2.2, seven tenths of deceleration in the largest discretionary block in the basket.

And the split that should decide how you read this: goods accelerated from 2.8 to 3.2 while services decelerated from 1.5 to 1.4. Services inflation is the component a central bank treats as evidence that price rises have become domestic and self-sustaining rather than imported. It went down. Core-core, which strips the subsidy effect out by construction, fell two tenths, from 2.4 to 2.2.

So the honest summary is not “Japanese inflation accelerated.” It is that a headline measure rose because a government stopped paying part of the electricity bill, while the two series that carry information about underlying pressure both weakened. One reader gives a poll of economists at 86 per cent expecting a Bank of Japan increase to 1.50 per cent at the 17–18 December meeting; that figure is single-sourced and we print it labelled. What we will say is that a store which keys a tightening signal to “core above target and rising” got a buy signal from a subsidy withdrawal, and the same release handed it a falling core-core and falling services and no mechanism to weigh them.

The market took the second reading. Ten-year Japanese government bond yields eased about 4 basis points to around 2.80 per cent and dollar-yen slipped to around 159.20 — from a 159.97 high the day before, a fall of 77.0 pips, or 1.352 standard days at a 5.68 per cent one-week implied. That level carries no observation time on the page that gives it, and we flag it rather than dress it up.

Fifteen minutes from now, and one publisher has the number in the wrong field

United Kingdom October retail sales land at 07:00 UTC, and the consensus apparatus is already broken in a way a scanner cannot see. Two publishers give figures for the same release:

  • One gives, with every field labelled: volumes 0.0 per cent month on month, +1.7 per cent year on year, ex-fuel −0.1 and +1.8; priors −0.4 month and +0.8 year.
  • The other gives a “1.7 per cent month-on-month increase, up from the previous 0.8 per cent gain” and nothing else.

The issuer settles it. The Office for National Statistics states September 2026 volumes as having fallen by 0.4 per cent over the month and risen by 0.8 per cent compared with September 2025. The figure 1.7 does not appear on its pages at all. So the second publisher has put the year-on-year consensus in the month-on-month field, and has put the year-on-year prior there too — which means its stated prior is +0.8 where the true monthly figure is −0.4. A 1.2-point error, and the sign is inverted: a fall reported as a gain.

Price the consequence. A store holding 1.7 as the monthly consensus, handed a 0.0 print, records a miss of 1.7 points on a series whose monthly readings have run between −0.7 and +0.7 all year. The true monthly surprise would be zero. That is the defective-field class at its most expensive: not a wrong number, a right number in the wrong cell, and no magnitude check finds it because 1.7 is a perfectly ordinary retail-sales figure.

Two more things from the issuer that a calendar row will not carry. It revises: the September bulletin moved August from +0.5 to +0.7 and July from −0.5 to −0.4, both upward, so today’s 0.0 consensus is quoted against a prior with a live revision history. And its release string is “Release date: 20 November 2026 7:00am” with no time zone stated anywhere on the page — an issuer authoritative about its process and merely assertive about its clock. Worse for anyone storing that hour: the ONS has been running a consultation on changing release times for market-sensitive statistics, retail sales among them, which closed on 30 October 2026. The hour in your schedule has a known pending change and no field for it.

Three losable legs, resolving fifteen minutes after this is published. Monthly volumes land inside −0.5 to +0.5. Annual volumes come in below 1.7. And this release revises at least one of the August or September monthly figures again. Score them at the next slot whichever way they fall.

What this does not tell you

  • We have no index level, no month-on-month figure and no seasonally adjusted series for the October Japanese print. The issuing agency’s monthly page was still serving September, dated 23 October, when we last read it, and the three readers carrying the October rates carry no index levels at all.
  • A correction to our own work. Yesterday we described a vendor field as giving “a typical dollar-yen reaction to this release of 10.0 pips.” It does not. The field is labelled as the previous release’s realised move, it is one observation rather than a typical value, and it is signed: −10.0 pips. We dropped the sign and promoted a single observation to a distribution in one clause.
  • That same vendor’s structured actual for this release was still blank at 06:18 UTC today, under seven hours after the print, with no October row in its history table at all. Yesterday we measured the identical hole at six minutes and called the latency the useful number. The latency is roughly sixty-eight times what we measured. The same publisher’s news index was serving nothing newer than 12:55 UTC on 19 November when we read it, over seventeen hours stale.
  • The headline forecast of 2.5 and the core-core forecast of 2.3 are each one publisher’s. If the core-core forecast was really 2.3, our second pre-registered leg stands; if that figure is wrong, the leg is unscoreable rather than won.
  • We cannot measure the yen’s reaction to the release. No page we read timestamps a dollar-yen level either side of 23:30 UTC, so the move from 159.97 to around 159.20 spans the print, a four-and-a-half per cent fall in Japanese equities and an Asian risk-off session, and we cannot apportion it.
  • The 86 per cent December poll figure, the 1.50 per cent target and the December meeting dates are one reader’s. No post-release market-implied probability has been published anywhere we looked.
  • One reader calls the core-core gauge “excluding fresh food and fuel” where two others say “fresh food and energy.” The figures agree, so we treat it as wording, but fuel and energy are not the same basket and we are not certain which the 2.2 belongs to.
  • We hold no sterling level of any kind — no cable, no sterling-yen, no euro-sterling — going into a release that is about to move all three. The one page that would have carried them did not open.

Related


Calendar Desk
Calendar Desk