Paste Canada's total fertility rate for 1931 to 1968 (births per 1,000 women, 38 years) into a changepoint detector that only knows flat segments and you get seven changes: 1933, 1942, 1946, 1953, 1963, 1965 and 1967. Most of them are the detector describing a curve with stairs. The rate fell through the early 1930s, climbed slowly to a peak in the late 1950s and then fell; a model with no way to say "the level moved a little every year" has to pay for that slope in steps.
Our changepoint detector has a second model for this, added earlier today: between jumps the level is a random walk, so slow movement costs only its small per-year variance and only a step too big for that wander is listed (after Romano, Rigaill, Runge and Fearnhead, JASA 2022). On the same 38 years it lists three jumps:
| Jump at | Level before | Level after | Size (noise sd) |
|---|---|---|---|
| 1946 | 3017 | 3405 | +7.0 |
| 1965 | 3541 | 3079 | -8.3 |
| 1967 | 2878 | 2557 | -5.8 |
1946 is the first full post-war year and the start of the baby boom; 1965 and 1967 are the two steepest years of its end. Everything else (the 1930s dip, the 1950s climb) reads as wander, which is what it looks like on the chart.
What changed in the product
The detector is free: page, command line, tables and CSV. The planned paid extra is a branded change report, the analysis on an analyst's own letterhead to print or forward. Until today it refused a wandering-level run and pointed at the CSV. Now it builds a report for that model too: the lead (how many jumps, the largest), the same chart and tables the page shows, the heavy-tail check when it fires, the benchmark score, a note naming any page options the model ignored, and assumptions written for a wandering level, among them that a slow shift is never listed as a jump and that a jump does not show a cause.
You can see the sample report, generated by the real code for an invented firm. A headless-browser test runs the whole page flow (paste, choose the model, open the report) and checks the opened report equals the Node output byte for byte.
How sure to be
On the 38 labelled series of the Turing Change Point Dataset this model scores F1 0.694 on 31 and 0.527 on 7 held out, against 0.477 and 0.296 for the plain level model, but reporting no change at all scores 0.663 and 0.639. It is a better tool for drifting series, not a solved problem, and the report says so on the page. The report is not on sale yet; nothing on the site can charge you.
Statistical analysis of data you supply, not advice.