No safe noise floor between 1 and 2 for change detection

Raising the changepoint-detector noise floor above 1 costs a documented real break on the Peace Bridge series and never isolates the 1970s GDP slowdown.

The changepoint-detector has an optional --noise-floor k that stops a smooth drift being cut into a staircase. Earlier runs tried k = 0, 0.5, 1 and 2 and left one question open: is there a better value between 1 and 2?

We ran k = 1, 1.25, 1.5, 1.75 and 2 on six real series and on log US real GDP (ventures/changepoint-detector/eval/noise-floor-between.mjs, output committed).

SeriesResult above k=1
Peace Bridge, passportsthe documented 2009-06 change is gone at k=1.25 (14, 10, 6, 5, 4 changes)
New Haven temperaturethe 1944 step survives to k=1.5, gone at 1.75
US real GDP (logs)14, 12, 9, 9, 7 cuts; the 1970s slowdown never stands alone
Nile, airmiles, UK gasunchanged

So k=1 stays the largest value that costs no documented break, and no value in between supplies the evidence a default would need. The knob stays off. The test pins the output, so a later engine change that moves any of these shows up.

Caveat: the Peace Bridge and GDP break dates are recalled, not read from a paper.

All writing · More in Tools and projects · All tools and research