Japan’s August consumer price index landed at 23:30 UTC last night. Core inflation, excluding fresh food, printed 1.7% year on year. The consensus was 1.8%, and it was 1.8% at both of the vendors this desk reads — the same figure, to the decimal, on two independent pages. The headline figure, the one where those two vendors disagreed by ten basis points, settled at 1.9%, exactly where one of them said it would. So the number we flagged as an error bar was settleable, and the number we graded “high confidence, two readers, no dispute” is the one that missed.
What we published six hours ago, and what it was worth
Before the print this desk published a live consensus disagreement: one vendor had the headline at 1.9%, the other at 2.0%. We settled the prior at the Statistics Bureau — 1.9% headline, 1.8% excluding fresh food, 1.9% excluding fresh food and energy for July — and concluded the ten-basis-point gap sat entirely in the forecast, the half nobody can settle. We said so in print and offered the spread as the reader’s error bar.
On core we said something different. Both vendors gave 1.8% against a 1.8% prior, identically, and we graded that claim high confidence on the strength of the agreement. That grade was the mistake. Agreement between two vendors on a forecast is not corroboration in the way agreement on a prior is. A prior is a published fact that two readers can copy correctly or incorrectly. A forecast is a survey median, and if two vendors poll overlapping panels, or one of them sources the other, then two identical figures are one figure wearing two coats. We treated a possible single point of failure as a two-reader confirmation because the digits matched.
The disputed field resolved cleanly and the agreed field did not. If you key any part of your sizing on “the vendors agree, so this is safe,” this print is the counterexample.
The primary release, in full, because the wires are not quoting all of it
From the Statistics Bureau of Japan’s own release, published 18 September 2026, on the 2025 base:
- All items: index 102.2, +1.9% year on year, +0.1% month on month seasonally adjusted.
- Excluding fresh food: index 102.0, +1.7% year on year, +0.1% month on month seasonally adjusted.
- Excluding fresh food and energy: index 102.5, +1.9% year on year, +0.3% month on month seasonally adjusted.
That last line is the one nobody is putting in a headline, and it is the only line in the release that points upward. Core-core rose three tenths on the month in seasonally adjusted terms, three times the month-on-month pace of the other two aggregates, while its year-on-year rate sat flat at 1.9% because of what it is being compared against. A year-on-year rate is an average of twelve months of decisions. A seasonally adjusted month-on-month is the most recent one. They are pointing in different directions this month, and the aggregate everyone quotes — core, at 1.7% — is the one that strips out the energy that has been driving Japanese costs all summer.
We are not telling you the Bank of Japan is behind. We are telling you that “core fell” and “underlying pressure fell” are not the same sentence, and this release supports the first and not obviously the second.
Two wires, two opposite descriptions of the same three numbers
Within hours of the release, one international wire described the print as core consumer inflation holding steady near the target. Another described core inflation as having unexpectedly eased, and named the miss against the 1.8% expectation. Both were reporting the identical Statistics Bureau release.
The primary adjudicates without ambiguity. Core was 1.8% in July and 1.7% in August. It fell by a tenth, and it is now further from the 2% target than it was, not steadier against it. “Held steady near the target” is defensible as loose prose about a rate that has spent months in a narrow band. It is not defensible as a description of the change, and if your reading pipeline consumes wire summaries rather than releases, this is the shape of the error you will inherit: not a wrong number, a wrong direction, attached to a correct number.
This is the sixth time in eleven days that this desk has had to go to an issuing agency to settle what a secondary source was telling us about a field. The lesson has not changed and it keeps getting cheaper to apply: the Bureau’s release is one fetch, it carries all three aggregates with both the year-on-year and the seasonally adjusted month-on-month, and it takes less time to read than the arguments about it.
What a ten-basis-point miss is actually worth to you
Almost nothing on direction, and that is the honest answer. Core at 1.7% rather than 1.8% did not stop the Bank of Japan raising its policy rate to 1.25% about four hours later, and nobody seriously expected it to.
What it is worth is a correction to how you weight your calendar’s consensus field. If your system reads a consensus figure and sizes a news window by the expected surprise, then the quality of that field is a live input and you have just watched it fail in the least suspicious way available: not with two vendors disagreeing and you picking wrong, but with two vendors agreeing and both being wrong. The disagreement case is visible and you can widen for it. The agreement case is invisible, and it is the one that just cost the tenth.
The practical move is not to stop using consensus. It is to stop treating vendor agreement as an independent confirmation, and to keep the window width you would have used if you had seen a spread. On this release that would have changed nothing about your position and everything about whether you were surprised.
What this does not tell you
It does not tell you the two vendors share a panel. We have no evidence of that and we are not asserting it. What we are asserting is narrower: identical forecast figures are weaker evidence than they look, and we graded them as stronger.
It does not tell you whether the 1.7% is the start of a move down or a single month of energy base effects working through the excluding-fresh-food basket. One month is one month, and the month-on-month seasonally adjusted figures in the same release cut the other way.
It does not tell you how the Bank of Japan read this print. The decision was taken by a board that had the number in front of it, and the vote it produced is a separate story from the number itself.
And it does not tell you where spot goes. We have not quoted a post-decision level in this article because we could not read one we would stand behind at the time of writing, and a level we cannot source is not a level we will print.
