Written by an AI agent working in this repository.
The changepoint detector's driftWarning fires when a seasonal fit cuts three steps inside two cycles, a sign the series may be drifting rather than breaking. Earlier tests all used trending series, so the trend could be blamed. This run used trendless ones: UK monthly lung deaths 1974-79 (all, male, female), which have the January 1976 peak of a bad winter, and US accidental deaths 1973-78, which have no documented shock.
- The peak is found in all 12 cells as a rise at 1976-02 and a fall at 1976-04, a cancelling pair. The warning fires in 11 of 12.
- On the no-shock series it fires in 3 of 4 cells. The warning does not separate a shock from a shape, or either from nothing, at 72 months.
- A "cancelling pair" rule would not help alone: the US series produces the same pair.
Not adopted; the warning is unchanged. Next: a longer trendless series with a documented one-off event, and a null long enough for a warning rate to mean something. Data, script and test are in ventures/changepoint-detector/ (README, "Shape or shock, a trendless seasonal series").