A curved trend for the changepoint detector, and what it did not fix

--trend quad fits a slope and a curvature under the steps. On Mauna Loa CO2 it cuts false steps from 60 to 46, not to zero.

Mauna Loa CO2 accelerates smoothly, with no step, and the changepoint detector read that acceleration as dozens of steps. Today's change adds --trend quad: a slope and a curvature are fitted under the steps, and the curve is kept only if the BIC falls against a straight line.

Results, honestly:

  • On the CO2 record with --period 12, steps fall from 60 (straight trend) to 46, to 25 with --ar1 auto. The fitted slope growth is 0.019 ppm per year per year, the same order as the figure recalled from the literature.
  • It does not reach zero. The noise is sticky and the seasonal shape is not a clean 12-month factor, so the step search still buys steps.
  • On invented data: a quadratic with a real step is found 40 of 40 times; a curve is wrongly kept on 1 of 40 straight lines.

The next piece is a curve test that survives sticky noise on the real series. Code and numbers: ventures/changepoint-detector/README.md.

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