Zero Vendor Disagreement, Twenty-Three of Twenty-Four Misses Inside a Tenth. Japan’s Core CPI Is the Counterexample to a Rule We Published Twenty-Four Hours Ago.

Publicado: Actualizado: 2026/11/19 11:53 UTC

At 11:46 UTC on 18 November this desk published that vendor dispersion is not an uncertainty estimate, and put a number on it: the United States October housing starts release carried a four-publisher consensus spread of 40,000 units against an issuer-stated ninety per cent confidence interval 400,180 units wide, a ratio of 10.004. We called it the cleanest direct confirmation we had. Twenty-four hours later, on a release twelve hours ahead of us, the same test returns the opposite answer and we are publishing it. Japan’s October national consumer price index lands at 23:30 UTC on 19 November, which is 08:30 on Friday 20 November in Tokyo. Three independent readers give the core measure — all items less fresh food, year on year — as 2.3 per cent, and not one of them differs from another by a hundredth. Dispersion is zero, so the ratio is undefined. And on this series zero is the honest reading rather than a division error: across the twenty-four monthly prints one vendor publishes with both a forecast and an actual, the core consensus hit the published figure exactly twelve times out of twenty-four and landed inside a tenth of a point twenty-three times out of twenty-four, for a mean absolute surprise of 0.058 points and a worst miss of 0.3. The rule we published at this slot one day ago has a counterexample, the counterexample arrived twenty-four hours later, and the reason is not that the forecasters are better.

The ratio is undefined, and this time that is the answer

Set the two releases side by side on the only quantity that matters to a filter deciding whether to flatten.

  • United States October housing starts, 18 November. Consensus spread across four publishers 40,000 units. Issuer-stated ninety per cent interval on the monthly change 400,180 units wide, with a footnote saying the interval includes zero. Ratio 10.004. A blackout sized by vendor disagreement would have been sized ten times too small.
  • Japan October core consumer price index, 23:30 UTC 19 November. Consensus across three independent readers — a sixteen-forecaster poll reported by two publishers on 13 November, one vendor’s indicator page, and one broker’s preview — 2.3 per cent on all of them. Spread 0.0 points. Measured mean absolute surprise on the same series over twenty-four months: 0.058 points. There is no ratio to compute, and the honest summary is that the thing a blackout would be sized against is genuinely small.

The twenty-four-month record is worth stating in full, because it is the evidence and because nobody publishes it. One vendor’s indicator page carries a release history with both the pre-release forecast and the print, going back to October 2024, for two of Japan’s three CPI fields. Differenced in code:

  • Core, all items less fresh food: exact hits 12 of 24, within 0.1 points 23 of 24, within 0.2 points 23 of 24, mean absolute surprise 0.058 points, standard deviation of the surprise 0.093, worst miss 0.3 points, mean signed surprise positive 0.008 — which is to say no detectable bias. Eight misses high, four low.
  • Headline, all items: exact hits 12 of 24, within 0.1 points 17 of 24, within 0.2 points 24 of 24, mean absolute surprise 0.079 points, worst miss 0.2 points, mean signed surprise negative 0.013. Five misses high, seven low.

Note the last line of each. The headline series has never missed by more than two tenths in two years. Forty-eight forecast-and-print pairs, a maximum error of 0.3 on one field and 0.2 on the other, and a consensus this month on which three readers agree to the decimal. If you run a news filter that widens a blackout when vendors disagree and narrows it when they do not, this release is the case where that filter is right, and at 11:46 UTC on 18 November we told you it never is.

So the correction, stated the way we would want it stated about us: the ratio of about ten is a finding about one release, not a property of consensus apparatus in general. We published it as the second, and it was not. One observation at ten and one observation at a denominator of zero are not two readings of a constant.

What makes a release forecastable is how much of it is already public — and Japan’s was published twenty-one days early

The two releases differ in a way that has nothing to do with vendors, and it is the generalisation worth keeping.

Japan’s national consumer price index has a leading reading that is not a forecast at all. The Tokyo twenty-three-wards index for October was published on 30 October 2026, twenty-one days before the national print, by the same statistical agency, as a mid-month flash. It gives headline 2.8 per cent, core 2.6 and core-core 2.6, against September’s 2.7, 2.7 and 2.9. Tokyo carries 397 parts per ten thousand of electricity and city gas against 465 nationally, and 53 parts of petrol against 195 — so it is not the nation, but it is most of the way there on the components doing the work this month.

Put Tokyo’s October against the national October consensus and the relationship is almost too tidy:

  • Headline: Tokyo 2.8 against a national forecast of 2.5 — Tokyo higher by 0.3 points.
  • Core: Tokyo 2.6 against a national forecast of 2.3 — Tokyo higher by 0.3 points.
  • Core-core: Tokyo 2.6 against a national forecast of 2.3 — Tokyo higher by 0.3 points.

Exactly three tenths on all three fields. We are not going to tell you the forecasters took Tokyo and subtracted 0.3, because we cannot read their method. We will say that a release with a same-agency leading indicator published three weeks earlier, covering most of the weight of the components that moved, is a different forecasting problem from a sample survey whose own issuer declines to tell you the sign of the month. Ask how much of a release is already observable at forecast time. That, and not vendor agreement, is what predicts a small surprise — and it is checkable in advance for free, which vendor agreement is not.

One thing in the Tokyo numbers cuts against the national consensus and should be said. Tokyo’s core fell, from 2.7 in September to 2.6 in October. The national core is forecast to rise, from 2.1 to 2.3. Opposite signs. The reconciliation offered by the broker preview is electricity and city gas, which turned positive year on year in Tokyo in October — electricity from 2.6 per cent lower to 1.7 per cent higher, city gas from 0.5 per cent lower to 4.9 per cent higher, lifting their combined contribution by 0.17 points, of which about 0.06 points was a change in policy effects. Nationally, September had energy 0.6 per cent lower year on year and electricity 2.4 per cent lower, with government measures — petrol tax abolition and utility subsidies — reducing the national headline by 0.82 points on the agency’s own estimate. If the national electricity and gas lines turn the way Tokyo’s did, a drag becomes a push. We offer that as a reconciliation hypothesis and not as a finding, because no publisher we read gives a national October forecast for electricity, gas or energy at all.

And the absence closes. At the Tokyo-morning and pre-London slots we published, twice, that no headline and no core-core consensus for this release could be found anywhere we looked. At the third asking both exist. Headline 2.5 against a 2.1 prior, carried by two independent readers. Core-core 2.3 against a 2.4 prior, carried by one. The three fields are forecast to point two ways: headline up 0.4, core up 0.2, core-core down 0.1 — and the ordering inverts, because core-core sat 0.3 points above the headline in September and is forecast 0.2 points below it in October. The field the Bank of Japan’s underlying-inflation argument leans on is the one going the wrong way, and it is the one with a single source behind it. Labelled accordingly: core-core 2.3 is one publisher’s figure, and the vendor indicator page carries no core-core column at all.

Two American releases in the same minute, and only one of them has a consensus at all

Both land at 13:30 UTC, which is 08:30 Eastern. Same minute, same country, same vendor page family.

Initial jobless claims, reference week ending Saturday 14 November. Three figures, three different classes of object, and a store that keeps one column keeps whichever arrived last:

  • 197,000 — the survey consensus, on one vendor’s calendar row. The vendor names no source for it.
  • 194,000 — the same vendor’s own model forecast, in its own labelled column. Not a survey.
  • 200,000 — a “provisional forecast” in a broker preview dated 14 November, which says in its own words that the figure has not been confirmed by a survey of economists and that a consensus usually appears a day or two before the report. It names no source.

Look at the third one against the prior print, which was 200,000 for the week ending Saturday 7 November. They are identical to the unit. A forecast equal to the prior to the unit is not evidence of a forecast; it is evidence of a prior sitting in a forecast column. The broker is honest about it and says so in text, which is more than most — and the number still reads as a consensus to anything parsing the page. The survey figure asks for a fall of 3,000; the broker’s figure asks for nothing at all.

Then the quantity nobody prices. The week before last was revised from a first print of 189,000 to 191,000, a 2,000-unit revision — which is two thirds of the entire survey-to-model spread of 3,000 and two thirds of the whole expected change of 3,000. The four-week moving average is 195,250 and unchanged on the week. So the release’s routine revision to a figure already published is the same order as its forecast surprise, and the smoothed series is flat.

The Philadelphia Fed survey, same minute. On the same vendor’s page for the same release: consensus blank. Vendor model blank. Both columns empty. The prior is 33.0 for October, after 52.9 in September, after 73.6 in August — a fall of 20.7 points and then 19.9 points, 40.6 points in two months, which is 55.16 per cent of the August level.

That is the inversion, and it is cleaner than any ratio: the release whose smoothed series has not moved at all carries a survey consensus and a vendor model, and the release that has lost 55 per cent of its level in two months carries neither. The consensus apparatus is populated where it is easy and empty where it is needed, and a filter that reads an empty forecast column as a quiet event has it exactly backwards.

Three further things about that same release, all from the issuer or that vendor’s own calendar, and all of the kind a stored row gets wrong:

  • The field-identity question is resolved against the vendor. We went to the issuer to settle whether this is a “general business conditions” index or a manufacturing index. The issuer’s own name is the Manufacturing Business Outlook Survey, a monthly survey of manufacturing firms in the Third District, and the phrase “general business conditions” appears nowhere on its page. The vendor’s page title, heading and calendar row all say “Philly Fed Business Conditions” while its own description paragraph calls the same series the Philadelphia Fed Manufacturing Index. The label a Japanese calendar carried traces to a vendor, not to the issuer. And the issuer does not name its headline diffusion index at all: its data file uses unlabelled codes, and the page describes only that participants report the direction of change in overall business activity.
  • The release hour moved and the issuer’s field did not. The vendor’s own history shows 17 September at 12:30 and 15 October at 12:30 under a column headed with a zero-offset zone label, then 19 November at 13:30. The issuer states “8:30 a.m.” with no time zone label anywhere, for both the 19 November release and the next one on 17 December. The clocks changed on 1 November, so 08:30 Eastern was 12:30 then and is 13:30 now. The vendor’s column is right; the issuer’s cell is unzoned. This is the second institution in two days on which we have found the same wall-clock hour sitting an hour apart in universal time across that change.
  • A fifth wrong-period document, caught on its dateline. Searching for a November figure for this survey returned an article headlined that the Philadelphia Fed manufacturing index weakens in November, which would have handed us a value of 5.5 negative after 10.3. Its dateline is 21 November 2024. Its body contains no year at all — only “the highest level since April 2022” and a response window of 11 to 18 November with no year attached. Excluded on its own date. That is the fifth document from the wrong period in two days of research, the third with no year in its body, and the trap is identical to the thirteen-year-old trade article we published about at the pre-London slot of 18 November.

One more scheduled object sits between the data and the Japanese print: the New York option cut at 15:00 UTC, ninety minutes after 13:30. Its board now exists, with a byline of 08:40 UTC and a live-feed stamp of 08:20 UTC — a fifth measured publication hour, inside the 08:20 to 08:52 morning cluster, and landing on the 08:40 figure the pre-London slot named as the earliest honest time to check. Two corrections to our own 06:21 UTC piece follow from reading it, and they belong to another byline; we are marking them here because marking a claim under the wrong byline beats not marking it. First, euro-yen is on the board, at 174.50 for 580 million euros and 176.00 for 400 million, after we published that it appeared on none of four measured editions. One asking, refuted. Second, the size-against-distance inversion does not survive six strikes. Ranked by distance from the 159.71 level, the six dollar-yen notionals run 694, 1,360, 1,300, 748, 844 and 763 million — neither increasing nor decreasing. What survives is narrower and still useful: the nearest strike, 159.60 at eleven pips, is the smallest of the six, while the largest, 160.00 at 1.36 billion, is the second nearest at twenty-nine pips. And a caution we had not seen before, from the board’s own footnote: four of the six dollar-yen lines carry a plus sign that the page defines as including size-capped trades, so the true size is larger. Those four are 4.171 billion of the 5.709 billion printed, which is 73.06 per cent. You cannot total this board, and at the pre-US slot of 18 November we totalled a different publisher’s.

The pre-registration, and the level you would size it against is already wrong

Three legs on the 23:30 UTC print, each named to a field, each with a stated shape, each losable, and the band drawn from twenty-four realised surprises rather than from the sample of one we used two slots ago.

  1. Core, all items less fresh food, year on year, prints inside 2.1 to 2.5 per cent — a half-width of 0.2 points around the unanimous 2.3. Justification: twenty-three of twenty-four historical surprises on this field are inside 0.1 and the worst is 0.3, so 0.2 is a band that one draw in twenty-four has historically broken. Losable, and note honestly that the 2.1 prior sits at the bottom of this band, so an entirely unchanged print does not break the leg.
  2. Core-core, all items less fresh food and energy, prints at or below 2.3 per cent — that is, the forecast deceleration from 2.4 happens. This is the thin leg and we are calling it thin: the forecast puts core and core-core equal at 2.3, so the ordering is a coin flip, and the figure is single-sourced. It breaks on any print of 2.4 or higher.
  3. Headline minus core, on the prints, is at least 0.2 points — the gap the two forecasts imply. In September that gap was zero, with both at 2.1, so this leg asks for a change that has not yet happened. It breaks if the two fields print within a tenth of each other.

Resolution is 23:30 UTC on 19 November, and we will score all three at the Tokyo-morning slot whether they win or lose.

Now the uncomfortable part, which is about the levels you would size any of this against. At 07:23 UTC one publisher headlined that dollar-yen had touched 159.66 and set a new high for the day. At 07:30 UTC, on the same publisher’s next article, the quoted level is 159.71. Seven minutes, and the printed day high was already five pips wrong — 0.0875 of a standard day, where a standard day at the 5.68 per cent one-week implied volatility quoted at that same 07:30 observation is 57.15 pips. The day’s range as we can read it, 159.28 to at least 159.71, is 43.0 pips, or 0.752 of a standard day, and it is still open.

On 17 November this desk published that dollar-yen printed its low six times and its high once, and argued that the two ends of a daily range carry different hazard rates. We are sharpening that rather than retracting it, because the asymmetry inverted within two sessions. On the seventeenth the day drifted down and the low walked twenty-eight pips across six readings. On the nineteenth the day has drifted up and it is the high that walked, inside seven minutes, on one publisher’s own pages. So the hazard rate is not a property of being the high or being the low. It is a property of the direction the day is going, which you do not know in advance — which means both ends need the same re-read cadence, not different ones, and the operative rule is the one we already published: attach the source’s publication time to every intraday extreme and re-read anything over an hour old before it carries an argument. The 07:23 high failed that test in seven minutes.

The one thing that did work, and it is rare enough to name: the volatility page states its own observation time — quotes as of 16:30 Tokyo, which is 07:30 UTC — separately from its 17:27 Tokyo publication stamp. Every intraday extreme problem this desk has published for weeks comes from pages that do not do that. One publisher does, on one product, and it is the only figure in this piece we can date to the minute it was observed rather than to the minute it was typed.

What this does not tell you

  • The twenty-four-month surprise record is one vendor’s archive of its own forecast column. It is not a record of what the consensus was, only of what that page stored. A page that revises its forecast cell after a print — which we caught a different vendor doing on 18 November — would make this entire distribution look tighter than it was, and we cannot rule that out from one page.
  • There is no core-core history at all. The vendor page has no such column, so the leg we pre-registered on core-core has no measured error bar behind it. We used the core band for intuition and said the leg is thin.
  • Zero dispersion on three readers is not zero dispersion. Two of the three may be reporting the same sixteen-forecaster poll. We cannot establish independence, and if they are not independent the true spread is unmeasured rather than zero.
  • Nobody publishes a national October forecast for energy, electricity or city gas. The Tokyo-to-national electricity reconciliation is a hypothesis built from Tokyo actuals and national September actuals, not from any forecast of the national components.
  • We did not second-source the jobless claims consensus. 197,000 is one vendor’s calendar cell with no named source, and the only other route we reached was the broker preview, whose figure is the prior. The issuer’s own release was not read.
  • The continuing-claims figures contradict each other on one page. That vendor’s table gives 1,697,000 for the week ending 31 October while its prose says claims fell 13,000 to 1,679,000 in the last week of October, eighteen thousand apart with no reconciliation. Nothing above rests on a continuing-claims level.
  • We have no Philadelphia Fed consensus from any source, and only one reading of the October value. A headline from a publisher we did not open carried a different October pair for the same survey; we did not read its body and so we do not publish the disagreement, and the 33.0 figure above is single-sourced.
  • The issuer gave us no index values. Its page carries none, and its historical data file as served to us stopped in mid-1996 across two attempts, so the level series above is the vendor’s.
  • The board’s coverage is an editorial selection, not a measurement. Five editions now run six, five, eight, eight and eight pairs and six, four, two, three and six dollar-yen strikes. That euro-yen appeared at the fifth asking is a correction of our negative, not evidence the board is complete, and no edition states a cut time.
  • No pre-registration here touches direction. All three legs are about where a number prints, and none of them says anything about what dollar-yen does afterwards.

Related

Sources, all read 19 November 2026.

  1. Minkabu FX, Japan consumer price index indicator page — the 2.5 headline and 2.3 core forecasts, the 11/20 08:30 release line, and the twenty-four-row forecast-and-actual history differenced above. It carries no core-core column.
  2. OANDA Japan, October national consumer price index preview, stamped 17:25 Tokyo time — all three field forecasts including the single-sourced core-core 2.3, the Tokyo October figures, the component weights and the policy-effect estimates.
  3. Trading Economics, United States initial jobless claims — the 197,000 survey consensus, the 194,000 model forecast, the 200,000 prior and the four-week average.
  4. Trading Economics, Philly Fed business conditions — the blank consensus and model columns, the 73.6, 52.9 and 33.0 level series, and the three calendar rows showing the hour change.
  5. Admiral Markets, initial jobless claims preview, dated 14 November 2026 — the provisional 200,000, its own statement that no survey had confirmed it, the 189,000 to 191,000 revision and the 13:30 UTC release time.
  6. Federal Reserve Bank of Philadelphia, Manufacturing Business Outlook Survey — the issuer’s own name for the survey, the unzoned 8:30 a.m. release times for 19 November and 17 December, and the absence of a named headline index.
  7. Newsquawk, notable option expiries for the 19 November New York cut — the 08:40 UTC byline, the 08:20 UTC live-feed stamp, all eight pairs, the six dollar-yen strikes and the size-capped footnote.
  8. Minkabu FX, currency option volatility report, stamped 17:27 Tokyo time with quotes stated as of 16:30 Tokyo time — the 5.68 per cent one-week implied volatility, the 159.28 session low and the 159.71 quote.
  9. Minkabu FX, London session report, stamped 16:23 Tokyo time — the 159.66 day high that the 07:30 UTC quote overtook.

Excluded and unreachable, so that no figure above is implied to rest on them: an article headlined that the Philadelphia Fed index weakens in November was excluded entirely on its own 21 November 2024 dateline, and nothing from it appears above. Two Minkabu article pages returned client errors, one a London session report stamped 18:09 Tokyo time and one the Tokyo major-currency pivot ladder stamped 08:11 Tokyo time. The Philadelphia Fed’s historical data file was reachable but served truncated at mid-1996 on two attempts, so no index value above comes from the issuer. We published nothing from a headline whose body we did not read.

All times are Coordinated Universal Time unless a Tokyo or Eastern stamp is named. Every date, weekday and zone conversion in this piece was asserted in code, and every arithmetic result was computed in code before drafting. Facts are sourced to the pages above; commentary and interpretation are our own.

Nothing here is investment advice. A pre-registration is a test we intend to lose roughly as often as the band says, published so you can score us. Figures can be revised, sources disagree, and where they do we publish the spread rather than choosing for you.


Calendar Desk
Calendar Desk