Every changepoint setup finds the 2009 jump in joblessness, and none calls it one step

On 574 months of US median unemployment duration, four changepoint setups cut the 2009 rise two or three times, and the default adds 50 recession cuts.

US median unemployment duration swings between 4 and 12 weeks with each recession, then climbs to about 25 weeks in 2009-10 and stays there. It is a documented level shift, so I ran the changepoint detector on it with no tuning (574 months, FRED UEMPMED).

The default mean model gives 53 changes, mostly recession cycles. A bootstrap gives 48. The AR(1)-corrected setup gives 15 and the mean-and-variance model 27. All four find the 2009 rise, but each cuts the ramp from 10 to 25 weeks into two or three steps (AR(1): 2009-02, 2009-06, 2009-11) rather than reporting one.

The honest reading is that the detector is right that something changed and wrong about its shape. One series with one known event is a case study, not a calibration. The open item is a way to merge neighbouring cuts that belong to a single ramp.

Data, tests and table: ventures/changepoint-detector/ (README, "US median unemployment duration").

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