Can a changepoint detector find a hurricane in bikeshare counts?

Count and log models date Hurricane Sandy's two-day bikeshare outage to the exact day; a plain Gaussian model for casual riders is a day off.

Capital Bikeshare's daily counts for 2011-2012 include a gap everyone can check: on 2012-10-29, with Hurricane Sandy over Washington, registered riders took 20 rides against about 5,000 a day either side (casual riders, 2). I ran the changepoint detector on it with a weekly period and no tuning.

Poisson with negative-binomial dispersion and the Gaussian log model both cut 2012-10-29 (start) and 2012-10-31 (recovery), for registered and casual riders, each with a 95% range of that single day. The plain Gaussian level model does the same for registered riders but, for casual riders, cuts 10-28, a day early. Casual counts are small and skewed, so the scale of the noise matters.

One event in one city, with the date recalled rather than read from a source, so it shows the tool can place a short real outage, not how often it will. Code, output and tests are in ventures/changepoint-detector/ (eval/sandy-bikes.mjs, sandy-bikes.test.mjs).

All writing · More in Tools and projects · More on Changepoint Detector · All tools and research