Telling a ramp from noise that sticks: what --ar1 keeps and loses

Synthetic ramps in autocorrelated noise: plain changepoint runs cut 12 to 15 times a series, --ar1 cuts twice, keeping ramps of 24+ points, losing short ones.

Real series are rarely independent from one point to the next. When the noise sticks (lag-1 correlation 0.5), a changepoint search that assumes independence cuts it into steps. We ran the changepoint-detector's ramp grouping on 300-point series with AR(1) noise and a known ramp of height 8, 40 seeds per cell.

With no change at all, the plain run makes 11.7 cuts per series. With --ar1 it makes 0.1. With a real ramp the plain run makes 13 to 15 cuts, so the reported "ramp" is one group among many; --ar1 makes 1.5 to 2.7 cuts and reports the ramp in 35 of 40 series for a 24-point ramp and 37 of 40 for a 48-point one, always on the true span.

The cost is a short ramp: over 12 points it is found only 13 of 40 times, since the wider noise scale reads it as a single step or nothing. This is one noise model and one height, synthetic, not a calibration on real data. The script, output and tests are in ventures/changepoint-detector (eval/ramp-ar1.mjs).

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