Written by an AI agent working in this repository.
The yearly count of "great" inventions and discoveries (1860-1959, World Almanac) runs about 2.5 a year, then reads 7, 12, 3, 10, 9 in 1884-1888, then settles near 3.7 and later 1.7. Plain Poisson changepoint search reports three level shifts: 1884, 1889 and 1933. The first two are really one burst.
What the changepoint-detector options do on this one series (variance is 1.64 times the mean):
| option | changes returned |
|---|---|
| plain Poisson | 1884, 1889, 1933 |
--overdispersion nb | same three, with the 1889 range widened |
--overdispersion (quasi-Poisson, dispersion 2.76) | none |
--bootstrap | none |
--bursts | 1884 (2.5 to 4.0) and 1933 |
--bursts flags 1885 (12, cut back to 3) but leaves a step at 1884, because it removes a single spike, not a five-year plateau. The cheap quasi-Poisson option declines the burst as the slower bootstrap does. Neither is a model of bursts, which is still open. This is one real series, so it shows how the options behave here, not how often they are right. The check is in great-discoveries.test.mjs.