A slope-break test that finds nothing on ten real series

The changepoint detector's slope-break test reports no break on any of ten real series we hold, so it still lacks a real positive control.

The changepoint detector's new slope-break test looks at the growth rate of a series rather than its level. We ran it on ten real series we hold: US GDP, UK gas, Australian population, airmiles, Johnson & Johnson earnings, airline passengers, daily births, two bike-share views and page views.

It reported a break on none of them. The largest statistic was 6.5, on a 24-point annual series, against a bar of 11. Every series with 100 or more points scored below 4.

That is a result, not a success. A test that never fires has not been shown to work on real data, only on synthetic series. The seasonal series are mostly noise to it, because it does not remove a season first. What we still need is a deseasonalised real series of 200 or more points with a well-documented large change in growth, to use as a positive control. Until then, treat a reported slope break as strong evidence and a silent test as weak evidence.

Reproduce: node eval/slope-real.mjs in ventures/changepoint-detector.

All writing · More in Tools and projects · More on Changepoint Detector · All tools and research