Repository navigation
Author decision: replacement for the stacking step-error metric (blocks P1) #50
Description
Author decision needed; blocks the step-error item in #39 (S-RE.3, SCI-11).
The stacking figure's step-error metric is the minimum over 120 post-event epochs minus the pre-event median, minus the same for the truth. Two consequences the review found: a negative value means the recovered drop is deeper than the true drop (over-estimated), the opposite of how both drafts read it; and at short stacks the minimum over noisy epochs is biased negative by the extreme-value effect, so the -5.4 percent at 1-day stacks is a property of the metric, not of the step. Choose the replacement:
- post-window median minus pre-window median (robust, simple, slightly smooths the step);
- a fitted step (amplitude of a Heaviside plus exponential recovery fitted over the post-event window; matches the earthquake truth's form);
- keep the minimum but report it as such with its extreme-value bias stated.
The chosen metric goes in the Fig 9 caption and Section 3.6 is rewritten from the regenerated sidecar.
Activity
Record the iteration-3 GitHub issues in the plan log and the collabor…
mdenolle commentedon Sep 13, 2026
P1 (PR #52) proceeds under option 1 as the default: the step amplitude is the median of the 120 days after the event minus the median of the 120 days before, on the recovered series and on the truth alike (
synthetic_demo.step_amplitude, also used by the one-at-a-time and multiverse drop metrics). Consequence worth knowing before you decide: with this definition the stack length has no first-order effect on the step in the multiverse (a stack shorter than the 120-day windows barely moves the medians), whereas the withdrawn minimum-based metric had ranked it first. A fitted step (option 2) would recover sensitivity to smearing at the cost of a model assumption. Switching is a one-function change.