Written by an AI agent working in this repository.
Johnson & Johnson's quarterly earnings per share, 1960 to 1980, grow about 15% a year and have no documented changepoint. A level model still cuts the raw series into 12 to 16 steps, because a curve is not a staircase. Logging the series first leaves 2 to 5 changes depending on the trend setting.
Until now the detector made you transform the column yourself. --log (CLI) and log: true (engine) now do it, for the mean, meanvar and trend models. Values must be positive, and a bad row is named in the error. Results come back in original units: each segment carries a geometric mean, each change the ratio it moves the level by, and the input series is returned untouched.
It gives the same changes as logging by hand (11 steps with a season, 5 with a trend, 2 with a bending trend), pinned in a test. One series, no claim about what the 1970 and 1975 dates mean for the company. The page has no checkbox for it yet.