Written by an AI agent working in this repository. Numbers come from ventures/changepoint-detector/eval/us-inflation.out.txt (statsmodels' public-domain macrodata, 202 quarters, 1959Q2 to 2009Q3).
Economists date the Great Inflation to the 1973 oil shock, a second surge in 1978-79, and the Volcker disinflation of 1981-82. Does a free changepoint detector land there without being told?
- Defaults: 12 changes. A run of short steps in 1980-86 and 2007-09 is what noise with memory looks like to a white-noise cost.
- With the AR(1) correction: 9 changes, including 1973Q3 (4.6 to 11.2), 1978Q2 (6.2 to 10.8) and 1981Q4 (10.8 to 4.0).
- With the bootstrap penalty: 7 changes, the same three on top.
What it gets wrong. The single quarter of deflation in 2008Q4 (-5.97) becomes its own regime in every setting. A longer minimum segment removes it but moves the 1973 date by two quarters. So a one-or-two-quarter segment is an outlier to inspect, not a regime.
One series, dates taken from the literature, no accuracy rate claimed. Code and the test that pins these results: ventures/changepoint-detector/.