At 05:00 UTC this morning the Bank of Japan released its Core CPI indicators for August. If your calendar is Forex Factory, the row said forecast 1.5 percent against a 1.6 percent prior. If your feed is Newsquawk, the print came over as 2.6 percent against a 2.3 percent prior. If you keep history on Investing.com, that page’s actuals for the same series run 2.7, 2.7, 2.7 and 2.8, with forecasts of 1.7 and 1.6 sitting beside them. If you use MQL5’s calendar, the same release is called BoJ Trimmed Mean Core CPI and its July actual is 2.0 against a forecast of 3.1. Four vendors, four mutually incompatible histories, of a series whose values were published months ago and are not in dispute at the source. Nobody here is wrong. They are quoting different numbers.
The spread is not on the forecast. It is on the prior.
A consensus disagreement is ordinary and this desk has published several of them. What is not ordinary is a disagreement about the previous value, because the previous value is a settled historical fact that the issuing institution published a month ago. Here it is settled three different ways: Forex Factory carries 1.6 percent, Newsquawk carries 2.3 percent, and Investing.com’s most recent prior is 2.7 percent. That is a 110 basis-point spread on a number nobody is forecasting.
Once the prior is contested the surprise is meaningless. Today’s headline arithmetic — 2.6 against 1.5, a 110 basis-point beat — is a subtraction performed across two vendors’ tables. It is not a beat. It is one publisher’s series minus another publisher’s series, and the answer would change if you swapped either one.
Each vendor’s forecast column is detached from its own actual column
This is the part that makes the field-mismatch reading unavoidable rather than merely plausible. Investing.com’s page shows a March actual of 2.2 percent against a forecast of 1.6, and a May actual of 2.8 against a forecast of 1.7 — forecasts running about a full point below that page’s own actuals, month after month. MQL5 has the opposite sign: July actual 2.0 against forecast 3.1, June 1.8 against 2.7, May 1.7 against 2.9 — forecasts running about a full point above its own actuals, month after month.
Forecast error does not look like that. Forecast error is noisy and centred. A persistent one-point offset in a fixed direction, inside a single vendor’s own table, is two different series stacked in two adjacent columns. Both pages are internally broken in the same way and in opposite directions.
The Bank publishes fifteen numbers under this heading
The mechanism is documented at the source, and it is six months old. On 26 March 2026 the Bank of Japan announced that it now publishes estimated CPI inflation excluding institutional factors, and it set out five indicators — CPI excluding institutional factors, trimmed mean, weighted median, mode, and a diffusion index of increasing and decreasing items — across three baskets: all items less fresh food; all items less fresh food and energy; and all items less food and energy. Release time is 14:00 JST, two business days after Japan’s official CPI. That is exactly the 05:00 UTC slot your calendar shows.
So the single row labelled BOJ Core CPI is a pointer into a grid of up to fifteen published series, and no vendor tells you which cell it has picked. The trimmed mean and the weighted median are not close to one another; that is the entire point of publishing both. A vendor that switched cells at any point in the last six months would produce precisely the history discontinuities on display above.
One more line from the same notice deserves to be read by anyone treating this as a fixed series: the Bank states that the definition of institutional factors may be subject to change without prior notice. A measure whose exclusions can be redefined without an announcement is not a measure you can build a stable surprise metric on.
We could not settle it at the source, and we are saying so
The honest limit of this article is that we did not close it at the Bank. The Core CPI page carries the data as a chart PDF and an Excel file. The PDF extracts cleanly enough to confirm the series names — trimmed mean, weighted median, mode, excluding fresh food and institutional factors, excluding fresh food, energy and institutional factors, excluding food, energy and institutional factors — but it is a chart deck and carries no table of monthly values. We therefore have the Bank’s taxonomy from the Bank and the numbers only from vendors, and we are not going to pretend otherwise. Today’s 2.6 percent is one wire, read in full, with no second reader found.
Note also what did not happen. Minkabu’s round-up of indicators and news carried during the Tokyo session, timestamped 14:40 JST — forty minutes after the release — records no economic indicators for the session at all. A release that four vendors cannot agree on is also a release that at least one Tokyo desk did not log.
What to do with the row
Delete the surprise, keep the event. If you have a news filter keyed on this release, the window is still real — 05:00 UTC, two business days after the national CPI, and it is a scheduled time you can trust. What you cannot trust is any rule of the form “if actual minus forecast exceeds N, do X”, because the subtraction is being done across incompatible series and the sign of your offset depends on which vendor your feed happens to be.
If you want an inflation input for Japan that you can stand behind, take it from the national CPI two business days earlier, where the Statistics Bureau publishes one headline and one ex-fresh-food figure and the vendors agree. This release is a research product with a grid behind it, being distributed through a pipe built for single numbers.
This desk has now published four shapes of the same underlying problem in twelve days: a vendor carrying a core measure in a headline field, a previous column holding a flash where the release compares to a final, a release filed under the wrong calendar day because the page renders in New York, and now four vendors carrying four histories of one settled series. The first three were errors. This one is not an error at all — it is a taxonomy being flattened, and it will not be fixed by any vendor correcting anything.
What this does not tell you
It does not tell you which cell of the Bank’s grid any particular vendor is quoting. We have inferred a field mismatch from the shape of the discrepancies; we have not proved it, because we could not read the Bank’s own monthly table. It does not tell you that 2.6 percent is the right figure for August — that is one wire and we found no second reader. It does not tell you whether any vendor changed series in March, only that a change at the source occurred in March that would explain the pattern. And it tells you nothing about the yen: this release moved nothing we can see, and we are not going to manufacture a reaction for it.
