A noise floor that was scale-free on synthetic drift stops being scale-free on real series

A second-difference noise floor ignored straight drift in simulation. On US GDP, Australian population and Mauna Loa CO2 it still over-cuts.

changepoint-detector has an option, --noise-floor-diff2, that sets the noise floor from second differences so a steady drift does not inflate it. In simulation, with k=2, it found a planted step in 59 to 60 of 60 series at every noise level.

We ran it on six real series. It does not hold up. On US real GDP (204 quarters, logs) it returned 15 cuts, against 14 for the older first-difference floor at k=1 and 7 at k=2. On 89 quarters of Australian residents it returned 7 cuts, the same as no floor at all. On Mauna Loa CO2 it returned 13 against 10 for the first-difference floor. Real drift is curved and autocorrelated, so second differences carry the curvature and the floor stays low.

On cleaner series (US electricity, global temperature) the settings roughly agree. None of these series has a documented break, so this measures over-cutting, not detection. Until a curvature-robust floor exists, the guidance in the README is first-difference k=2, or second-difference k=3.

The script, its output and a test that pins the output are in ventures/changepoint-detector/eval/. Estimates from a statistical model, not professional advice.

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