Placebo test of the effect method

For each sampled port or chokepoint and each daily series, we drew random start dates from 2022-01-01 (1 January 2022) to the data date, 2026-10-25 (25 October 2026), with window lengths of 3, 7, 14, 28 days. A window was kept only if neither it nor the 120 days before it overlapped any approved event on that place, plus a 14 day margin after each event. Each kept window went through the same code that produces published figures, method wfm-effect-1.2, with the same baseline rules, holiday exclusions and noise floors. The random draw is seeded (20261028), so the run can be repeated. Dates before 2022-01-01 (1 January 2022) are left out because pandemic shutdowns were real, unrecorded disruptions.

A window counts as flagged by the same rule the anomaly detector uses: a standardised score of 3.0 or more in either direction, a change of 25 percent or more, and a baseline of at least 3.

Every window here is one where no event was recorded, so a flag is either a false alarm or a real disruption nobody has written up yet, such as an unreported closure, a schedule change, or a data gap. The share flagged is therefore an upper limit on the method's false-alarm rate, not a measurement of it. A large score on a published event is a reason to read the sources, and is not proof of cause.