FX Intervention Research InstituteFXIR.JP — DON'T TRUST, VERIFY.
Reference translation. The canonical text of every FXIR document is Japanese. This page is provided for readers and machines that do not read Japanese. Where this page and the Japanese canonical differ, the Japanese text governs. Canonical documents carry a version number and a SHA-256 hash; quote those, not this page. Rules of the institute (canonical): kaisoku_v1.0.md — SHA-256 a56916436249830dc5e866d656e2528a28b49d77927e54ce2c09c626e03ba597
METHOD

Methodology

Canonical: docs/methodology_v1.0.md (METHOD-v1.0) · SHA-256 426c93bfc5e8dcfa04c2d727ad67cdf287b4a45e26b0ecd65562f4ab3d023c30

Definitions are fixed by version. A research paper states which METHOD version it uses. Changing a definition creates a new version and the reason is recorded (Art. 32).

Time and business days

All times are recorded in JST and stated in ISO 8601 with the +09:00 offset. A business day is a weekday excluding Saturday and Sunday. Japanese public holidays are counted, because the price series exists on those days.

Episode variables

VariableDefinition
CostTotal intervention amount over the episode. MOF confirmed figures; where the period is not yet confirmed, press estimates are labelled ESTIMATE and kept separate.
ImpactHigh on the first intervention day minus the lowest low of the episode, in yen.
EfficiencyCost divided by Impact.
Prior velocityMean absolute value of the four differences taken from the five closes preceding the first intervention day.
50% recovery lineEpisode low plus unrounded Impact multiplied by 0.5. Display is rounded; computation is not.
Effect half-lifeBusiness days from the day after the last intervention day until the close first crosses back above the 50% recovery line. If it never does, the observation is right-censored and excluded from correlation analysis.

The general rule for grouping nearby intervention days into a single episode is not yet formalised. The current six episodes are taken as enumerated in WP-001 §2.2. This gap is stated rather than concealed.

Survival window

From the business day after the last intervention day to the day before the half-life is reached. An event occurring on the day the half-life is reached is not counted as having occurred while the effect was alive. If the window is empty, the judgement is recorded as indeterminate.

Statistical policy

With fewer than eight observations we do not use p-values as grounds for a conclusion. "Statistically inconclusive" is treated as a legitimate research result (Art. 22). Criteria are preregistered before the resolving data arrives; post hoc changes and multiple-comparison searching are not performed (Art. 19).

Boundary of AI involvement

AI does not participate in generating the published data JSON or in numerical aggregation (Art. 12). AI is used for drafting and audit assistance only, and no AI output is accepted as fact without independent recomputation (Art. 14). A machine-audit PASS certifies internal consistency, not correctness of content (Art. 53).