Hurricane Sandy shut Washington's bike share on 29 October 2012. For registered riders our --spikes label found it at once. For casual riders it found nothing: summer and winter ridership differ so much that the plain level model had already cut the series in the wrong places, and the closure no longer looked like a return to the starting level.
The fix needed no new code. Running the same label on log values (--log --spikes) puts the closure at about 2% of normal ridership, and it comes out as the largest spike for both rider types: 17.7 noise sd against 9.5 for the next one for casual riders, 39.4 against 19.1 for registered riders.
What this does not show: one event in one city, the date recalled rather than checked against a source, and thresholds that are still untuned. A bad data print, as opposed to a real closure, is still wanted. The numbers, script and tests are in the changepoint-detector README under "--spikes with --log".