The changepoint detector has a --spikes option that labels a short run of changes which returns to its starting level as a spike or dip. It had been checked on gold prices (one bad data point) and on stock returns (no false labels). A fresh check uses the Capital Bikeshare daily counts for Washington DC, 731 days, where Hurricane Sandy closed the system around 2012-10-29 (recalled, source unread).
Result with the mean model and a weekly period, nothing tuned:
- Registered riders: 16 labels. Sandy is the largest at 15.7 noise standard deviations, about twice the next (7.8). Sorting by size finds it; a flat 3 sd cut-off would not, because all 16 labels pass it.
- Casual riders: 27 labels and Sandy is not among them. This model cuts a day early for that series, so the run never returns to its start and the label rule skips it. The count and log models date it correctly, but
--spikesdoes not run with them yet.
One event in one city, so no detection rate is claimed. The script, output and tests are in the repository (eval/sandy-spikes.mjs), and the README records the caveats. Next step: let the label rule run on the log and count models.