Statistics Bureau of Japan

Statistics Bureau of Japan connector: the CPI at item depth.

Japan held four series from the IMF and, briefly, 21 COICOP categories taken from the OECD. The instruction on 2026-09-06 was to drop the OECD and take Japan from its own office, so this replaced it: 785 to 792 items per base, down to "Non-glutinous rice-A" and "Boiled Udon".

NOT e-Stat, and not for want of trying. e-Stat's JSON API needs a free application id -- asked without one it answers, in Japanese, "authentication failed, check your application ID" -- and its file listings are rendered by JavaScript, so neither is reachable from here. The Statistics Bureau publishes the same numbers as plain CSV under www.stat.go.jp/data/cpi/<base>/csv/, keylessly.

What that costs, stated because it is a real loss: the OECD had Japan back to 1955 at 21 COICOP categories; these files begin in 2015 at ~790 items. Depth bought with history.

Four things worth knowing before changing anything here.

The CSV is WIDE and Shift-JIS. Six header rows describe the columns -- Japanese name, English name, item code, serial number, weight, weight per 10,000 -- and then one row per month, YYYYMM followed by one value per item. So a "row" is a month and a "column" is a series, which is the transpose of every other connector here. The encoding is Shift-JIS because row 0 carries the Japanese names; reading it as UTF-8 does not fail, it silently mangles every label.

One file per base, and all three are real. 2015 covers 2015-01 to 2021-06 and is discontinued; 2020 covers 2020-01 to 2026-07; 2025 covers 2025-01 to 2026-07. They are not old-base reruns of one span -- 2015-01 to 2019-12 exists in the 2015 file alone -- so all three are taken and THE BASE IS PART OF THE SERIES ID. An index on the 2015 reference and one on the 2020 reference are different series, and merging them would be a splice this system does not make silently.

The rates are DERIVED. These files carry an index and nothing else.

No hierarchy is published in them. The columns are ordered hierarchically and the weights make the ten major groups obvious, but no column states a parent, so jpsb_parent is not emitted and Japan's categories arrive flat. That is a known gap rather than an inference: guessing a tree from column order is exactly the silent wrong answer this system avoids.