The Changepoint Detector's --model drift lets the level of a series wander as a random walk, so a population curve or a lake level is not cut into a staircase. It has one known failure: when a series goes through a burst of volatility, a stretch where everything moves five times as much, the model has one noise scale for the whole series, so it reads the burst as a run of separate jumps. On the Turing Change Point Dataset (TCPD) that is 63 cuts on the bank series, where the annotators mark none, and 52 on bitcoin.
The change
The fix tried today gives every segment its own scale. Inside a segment the noise and the wander are both multiplied by that segment's scale; the ratio between them stays shared. The scale is profiled out of each segment's Kalman filter innovations, so the exact optimal partition stays O(n²) and a change now costs a penalty for two new numbers (a level and a scale), 3 log n. A volatility burst becomes two cuts, one where it starts and one where it ends, and each is reported with the scale before and after it.
What we said would count, before running it
The rule went in the scripts' headers first. Offer it as an opt-in flag if:
- on a synthetic random walk whose steps are five times larger for 60 of 300 points, it reports at most half the shipped model's cuts;
- it loses no more than 6 of 60 series' worth of power on plain level steps;
- on TCPD it is no more than 0.02 below the shipped drift model on both the 31 shipped and the 7 held-out series.
What happened
The synthetic half passed easily. The volatility-burst walk went from 17.0 cuts to 2.4, with both burst edges found almost every time; white noise whose spread triples for a while went from 7.1 cuts to 2.0; level steps were found 60 of 60 and 59 of 60 times.
On TCPD the shipped set gave F1 0.756, the highest mean any setting of this detector has scored there, above the published binseg default (0.714). Bitcoin went from 52 cuts to 28. The held-out set gave 0.464 against the shipped model's 0.527: a drop of 0.064, well past the 0.02 guard. So the rule fails, and nothing changes for users. The option stays in the library with its tests and both evaluation outputs, in the README section "Volatility on a drifting level".
Why bank got worse, not better
bank went from 63 cuts to 170. It turns out not to be bursty at all. It is sticky: 246 of its 580 steps are exactly zero. Give each segment its own scale and every flat run becomes a segment whose scale drops to the floor, buying far more likelihood than the penalty costs. The median segment is three points long. A scanline image in the held-out set does the same thing (233 zero steps, 27 cuts to 75).
That points at the next try: a scale floor tied to the data's resolution, or a noise term that knows some steps are exactly zero. Since that idea came from looking at the result, it needs its own pre-registration before it can count.
The code is MIT licensed; everything above can be rerun with sh eval/tcpd-fetch.sh tcpd-data && node eval/tcpd-drift-scale.mjs tcpd-data and node eval/drift-scale-sim.mjs in the venture's folder. This is statistical analysis of data you supply, not advice.