WORKING PAPER · REFERENCE TRANSLATION

The Intervention Reaction Function Has a Two-Layer Structure

Separating trigger conditions from execution conditions, and evaluating intervention effect by half-life — WP-001 · v1.3 · published 2026-08-08

Abstract

For the six yen-buying intervention episodes since 2022 (n=6), we measure intervention effect by its half-life: the number of business days until the daily close recovers half of the episode's peak impact, against thresholds fixed in advance. Half-lives range from 1 to 431 business days, with no clear relation to the amounts spent. The simple hypothesis "effects persist if a macro shift (+α) arrives" is rejected by episode E5; what separates long-lived from short-lived episodes is whether the +α arrived within the effect's survival window. All findings are descriptive observation, not statistical inference.

Key findings

Amounts E1–E5 are Ministry of Finance official figures (FACT). The E6 total is a press estimate until MOF's monthly release on 2026-08-28. The largest episode by amount (E5, ¥11.73T) lasted 5 business days; the longest-lived (E4) was among the smaller ones — persistence is governed by whether a macro reinforcement arrived in time, not by size.

The two-layer claim

Japan's yen-buying reaction function separates into (i) a trigger layer — state conditions under which intervention becomes likely (level, speed, political urgency), and (ii) an execution layer — conditions governing when and how it is carried out. Conflating the two produces most public misreadings of intervention "failure" or "success." Effect duration belongs to neither layer: it is governed by whether a macro shift (+α: a BOJ rate hike ≥25bp, a ≥50bp fall in the US 2-year yield versus the final intervention day, or a realized-volatility spike) arrives before the effect's half-life expires.

E6 addendum (post-publication, display layer)

WP-001 v1.3 froze its analysis with E6 right-censored (data through 2026-08-07). After publication, the preregistered E6 threshold — 159.483 JPY, the midpoint-recovery line fixed in PR-001 before the outcome — was crossed by the daily close of 2026-08-13 (159.518). The half-life is 8 business days, made unique by the Day-7 canonical close (159.426 ≤ threshold). Judgment used closes only; intraday touches on both days did not count. The canonical text is not retroactively edited; the determination will be incorporated in v1.4 after MOF confirms the E6 amount. Evidence data (SHA-256-pinned CSVs) are published on the live page (JA).

Limitations

n=6. Descriptive observation only; no statistical inference is claimed. Execution rates are not identifiable from public accounting (WP-004). Nothing here is investment advice. License: CC BY 4.0 — the data and text are free; verify them rather than trusting us.

Verify

Every claim resolves to a primary source or a SHA-256-pinned file. Preregistrations existed before outcomes; a Bitcoin timestamp (OpenTimestamps) of the stamped commit is pending confirmation. See Verify (EN) · GitHub.

English reference translation of the full paper: wp-001_en.md SHA-256 8caba200304a43495b7a0393766dfe5e41993c015eb5d44fa45e45426a9c5665. The Japanese canonical text governs: wp-001.md SHA-256 139025fd6e9584e4d6990e3df6c2080b7c8dcb0d4c564b81a5c5a3e57a17bd2f. Cite the canonical file and its hash, not the translation.