From 22 changes to 3: a noise floor per column for joint changepoints

The joint changepoint fit now floors the noise of smooth columns: German unemployment goes from 22 changes to three, with white-noise power unchanged.

Our changepoint detector can look for one set of changes across several columns at once: a price and a volume, or four room sensors. That joint fit had a known weakness. It measured each column's noise from the scatter between neighbouring points. A smooth column, like a stock price or a room's temperature, has almost no scatter between neighbours, so every gentle bend looked like a real step. On three labelled multi-column series from the Turing Change Point Dataset it reported 95 to 132 changes. The people who labelled those series marked between 2 and 22.

The fix copies a rule that already worked on single columns. If a column's own level fit cuts 5 or more times, and setting its noise floor to twice the scatter about a local 25-point line at least halves that count, the column gets the floor. A column that looks like white noise keeps its plain estimate.

We wrote the settings, the primary one and a decision rule down before scoring anything. The rule: make the floor the default only if it raises the mean F1 on the three series and keeps the hit rate within 5 points of the old fit in every white-noise simulation. Both conditions held:

  • mean F1 rose from 0.172 to 0.286, and the number of changes fell from 95/117/132 to 20/39/89;
  • in all 20 white-noise cells (1 to 8 columns, steps of 0 to 1.5 noise sds) the hit rate was the same as before. That matters because small shared steps are what the joint fit is for: a 0.5 sd step spread over 4 columns is still found 76% of the time. Flooring every column regardless cut that to 3%.

A real pair, checked afterwards. German unemployment, unadjusted and seasonally adjusted, quarterly 1962 to 1991. The old joint fit found 22 changes and warned that it was probably over-cutting. With the floor, both columns are flagged as smooth and it finds three: 1974Q4 (+3.4 points, the first oil shock's recession), 1982Q1 (+4.3, the early-1980s recession) and 1989Q2 (-1.8). The warning no longer fires.

What we are not claiming. The new default still scores below the column-by-column reading (0.521) and below reporting no change at all (0.621) on those three series, so the page still leads with the column-by-column list. One setting we listed but did not make primary, an autocorrelation correction plus a floor on every column, scored 0.564, the best of anything we tried. Three series is too few to promote it after the fact, so it ships as an option ("fewest cuts") until fresh labelled data can test it.

Everything is in the venture's README under "A joint fit that survives smooth columns", with the scripts (eval/tcpd-multi-floor.mjs, eval/multi-floor-sim.mjs) and their outputs. On the command line it is --columns --joint-floor auto (the default), off, a number, or --joint-ar1. The page has a matching menu in several-columns mode.

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