The Changepoint Detector has a branded report for several columns measured at the same points: a small-multiples chart, the changes every setting agrees on column by column, and one segmentation for all the columns with each column's shift in its own noise. It is meant to be printed or saved as a PDF and handed to someone. Until today a wide set did not fit the page.
What broke on paper
Paper cannot scroll. The joint table sat in a box that scrolls sideways on screen, and the chart is one image a printer cannot split. A headless-Chromium check (eval/multi-report-print.mjs) prints the report on A4 with 15 mm margins, which leaves 680 by 1009 CSS pixels, and measures the widest table and the tallest chart:
| case | joint table | tallest chart |
|---|---|---|
| UK road casualties, 7 columns, level and volatility | 814 px | 631 px |
| synthetic, 12 columns, level | 1,081 px | 1,078 px |
| synthetic, 20 columns, level and volatility | 2,156 px | 1,792 px |
Anything past 680 px wide was cut off at the right margin. Anything past 1009 px tall ran off the bottom of the page.
The fix
Past six columns (four when the volatility model adds an sd ratio to every cell) the report now splits the columns into balanced groups. Each group gets its own chart and its own joint table, captioned "Columns 1 to 4 of 7: DriversKilled, drivers, front, rear" and kept whole on a page. Every chart draws the same changes, so the groups line up by date. After the change all four cases fit: the widest table is 648 px and the tallest chart 542 px. Sets of two to four columns, the usual case, are laid out exactly as before; the three sample reports regenerate byte for byte.
What seven real columns say
The wide test set is real: R's Seatbelts table, the monthly UK series that Harvey and Durbin used in 1986 to measure the front-seat belt law of 31 January 1983. Its columns are drivers killed, drivers killed or seriously injured, front-seat and rear-seat passengers, distance driven, petrol price and van drivers killed. We left out the source's law column, since it is the answer.
- The joint level-and-volatility fit finds five changes, and one of them is 1983-02, the first full month under the law.
- The column-by-column panel, the reading that held up best on labelled data, agrees on only three changes, all in the petrol price: 1974-04, 1977-08 and 1979-06.
- The report itself warns that the joint fit is cutting more than once per 50 points, so its 1983-02 is a lead, not a finding. The casualty counts are strongly seasonal and the several-column mode has no season term yet. One column at a time with
--period 12, the detector does date the law (1983-01, with a range that holds February).
So the report prints whole now, and the data shows the next job: a season term per column. Until then, for monthly counts, run the important column on its own with its season.
The report is a planned paid extra and is not on sale. The detector, the CLI and the CSV output are free and MIT licensed. A detected change describes the data, not its cause. This is a statistical estimate, not advice.