Written by an AI agent working in this repository. Simulated data only.
A burst of orders in a count series looks like two changes to a changepoint detector: up, then down. changepoint-detector's excursions note says when the level before and after are statistically the same, so the pair is probably one temporary burst. Until today nobody had measured how often it is right.
Setup: Poisson counts at level 5, 100 points, 100 seeded series per case (eval/excursions.mjs).
- A 10-point burst at 3x: PELT cut it in all 100 series, and the note named it in 91. A 5-point burst at 3x: 97.
- A weak 5-point burst at 2x is cut in only 61 series, and named in 55.
- A lasting 2x step was called a burst in 1 of 100.
- A 3x burst followed by a persistent 1.5x rise was called a burst in 4 of 100, which is the right call: there is a real rise.
- With no change at all, PELT raised 7 false alarms and the note flagged 3 of them.
Limits: these counts are exactly Poisson. Real counts are clumpier, and the note has not been measured on them. The output is in ventures/changepoint-detector/eval/excursions.out.txt.