Changepoint tools over-cut a steadily drifting series: a smooth bend looks like a shift when the noise is small. The changepoint-detector has two remedies, a noise floor (--noise-floor k) and an AR(1) correction (--ar1). Until today they had not been compared where the truth is known.
Setup: 60 seeded series of 120 points per noise level, drift 0.2 per step, AR(1) noise (phi 0.5), one step of 4 noise sd at t=60.
- No remedy: about 3.6 false cuts per series, step found 59/60.
--ar1: 0.4 to 0.5 false cuts, but the real step found in only 36 to 42 of 60. Dividing by the long-run sd (about 1.7x at phi 0.5) shrinks a 4-sd step to about 2.3.- Noise floor k=1.5: 0.27 to 0.8 false cuts and the step found 59 to 60 of 60, unless the step is no bigger than a couple of drift-steps (then it is lost).
So with sticky noise and a decent-size step, the floor is the better knob; --ar1 suits cases where a false cut costs more than a missed one. Synthetic data only, one phi, and --ar1 joint is untried. Reproduce with node ventures/changepoint-detector/eval/noise-floor-ar1-sim.mjs.