A changepoint detector told West German unemployment had 24 regime changes

Run on 30 years of West German unemployment, a mean changepoint model finds 24 steps; allowing for autocorrelation leaves 2. What that does and does not show.

West German unemployment, quarterly, 1962 to 1991, is a good test for a changepoint detector because it is not made of steps. It sits near 1% until 1973, climbs through the oil-shock recession of 1974-75 to about 4.5%, and climbs again through the 1980-82 recession to about 9%.

The changepoint detector (PELT with a BIC penalty, MIT licensed) reads a ramp as a staircase. On the seasonally adjusted series it returns 24 level changes, on the unadjusted one 10. With --ar1, which allows for lag-1 autocorrelation, it returns two: 1974Q3 (mean 1.05 to 4.33) and 1981Q4 (4.33 to 8.28).

Three things to take from it, none of them a claim about the economy:

  • The two dates sit on the two recessions, but the rate had been rising for a year or more before each, so they mark the middle of a ramp, not a day something happened.
  • --ar1 is not free. It drops the real 1966-67 recession bump the default finds, and on the unadjusted column, giving the seasonal period as well (--period 4 --ar1) loses the 1974 step and keeps only 1981Q3.
  • This is one series. It adds a second real case, after Gallup approval, of --ar1 taming an over-cut at the price of a smaller real change.

The series is in the repository (data/german-unemployment.csv) with its source, and german-unemployment.test.mjs pins every number above. The data are AER's GermanUnemployment (Franses 1998). This is statistical analysis of public data, not economic advice.

All writing · More in Testing and red team · All tools and research