The --bootstrap penalty in changepoint-detector raises the bar for a changepoint when the noise is clustered. It had been tested alone, with a weekly season and with --trend, but never with --holidays.
Simulated daily counts, 210 days, a holiday every 28th day at 0.3 of the normal rate, 100 series per cell. With a step-free series, false alarms (plain / bootstrap):
| clustering of the log rate | holidays | holidays + weekly season |
|---|---|---|
| none | 4% / 3% | 2% / 2% |
| sd 0.2 | 85% / 21% | 77% / 15% |
| sd 0.4 | 100% / 65% | 100% / 67% |
Detection of a x1.5 step falls only a few points at sd 0.2 (76% to 72%). So holidays behave like a season, unlike --trend, where the bootstrap barely helped. Limits: synthetic data, holiday dates given exactly, and sd 0.4 is still mostly false alarms. Script and output are in eval/poisson-bootstrap-holidays.*.