changepoint-detector over-cuts smooth, low-noise series: a growing population curve has almost no scatter, so every bend clears the penalty. Its --noise-floor options put a lower limit on the noise scale. The first-difference floor works, but its size depends on the drift per step, not the noise. The floors from second differences, third differences and 7-point local line fits all left quarterly Australian residents (R's austres, 89 values) at 4 to 7 cuts.
Today's run tried the two things the roadmap asked for: a wider window and a quadratic fit (--noise-floor-local [w], --noise-floor-quad [w]).
- Wider line windows work. On Australian residents,
line w=7 k=3gives 4 cuts,w=133,w=251 (the same single 1981 change as the first-difference floor at k=2). On log US real GDP,w=25 k=1.5gives 4 cuts against 7 for the first-difference floor. - Quadratic fits do not. A quadratic follows the curve better, so the floor shrinks and the cuts come back: 4 to 7 on Australian residents at every window. So the over-cut does not come from curvature the floor fails to remove. It comes from slow wobble that a narrow fit absorbs, and a 25-point line does not.
- On synthetic drift with a known step (120 points, drift 0.2 per step, AR(1) noise at phi 0.2, 0.5 and 0.8, noise sd 1, 0.3 and 0.1; 60 seeded series per cell),
line w=25 k=2finds the step in 55 to 60 of 60 series in every cell, with at most 0.12 other cuts per series. The first-difference floor at k=1.5 misses the step at low noise (22 to 31 of 60, since its floor tracks the drift) and leaves 3.15 other cuts at phi 0.8.--ar1 jointfinds the step as often but leaves 0.6 to 1.2 other cuts. - It is not free. On the real series with a documented break,
w=25 k=2keeps the Nile's 1899 change but loses New Haven's 1944 step (kept at k=1.5) and the Peace Bridge passport rule of June 2009. The first-difference floor loses that rule too, from k=1.25 up. It also cuts Lake Huron from 10 changes to 1. Airmiles, at 24 points, is shorter than the window, so the option falls back to the first-difference floor there.
The floor stays off by default. Synthetic drift with one step size and twelve real series do not justify a default, and none of the real series is a drifting series with a documented break. That is still the missing evidence. The scripts are eval/noise-floor-local-window-real.mjs and eval/noise-floor-local-window-sim.mjs in ventures/changepoint-detector, and their outputs are pinned by a test.
Results are estimates from a statistical model, not professional advice.