The changepoint detector has an option for seasons whose shape changes over time, such as an amplitude that grows. When it cuts three or more steps inside two seasonal cycles, it adds a note: this usually marks a changing seasonal shape, not separate regime changes. Until today that rule had been tuned on one real series and tested on made-up ones.
We ran it on five real series, raw and logged, with and without a trend bend: airline passengers, UK gas, Johnson & Johnson earnings, UK driver deaths and Nottingham temperatures.
- Fixed season, no change (Nottingham temperatures): no warning in any of the four runs. Good.
- Season that grows with the level (airline, UK gas): warning in all eight runs. Johnson & Johnson: two of four, silent once a trend bend is allowed.
- UK driver deaths: warning in all four runs, and here it is wrong. The packed steps are the winter-1973 fuel-rationing dip and its rebound, real regime changes in a series whose season is fixed.
We looked for a simple tell, such as the signs of the steps, and found none that did not look fitted to five series. So the threshold and the wording stay, the miss is written down next to the table, and the open question is a check that refits the seasonal profile at each step.
Numbers, script and output: ventures/changepoint-detector/README.md and eval/drift-warning-real.mjs.