Our changepoint detector could already check several columns at once from the command line. As of today the web page does it too: paste a CSV with a temperature and a CO2 column, or a price and a volume, pick "Several columns measured at the same points", and you get two readings side by side.
The example on the page is seven years of daily returns for the Swiss franc, the pound and the yen against the dollar, 2012 to 2018 (1,751 trading days, Federal Reserve H.10 rates). Asked for changes in level or volatility, it answers:
- Changes every setting agrees on, column by column: 2015-01-20 in the franc only, and 2016-06-24 in the pound only. The Swiss National Bank dropped its floor against the euro on 15 January 2015. Britain voted to leave the EU on 23 June 2016. Neither event shows up as a change in the other two currencies.
- One segmentation for all columns: nine changes. It dates the franc's break to 15 January exactly, and puts two changes either side of the referendum in the pound. The other five move every mean by under a third of a noise standard deviation. They are volatility regimes, and the table now shows each column's standard deviation after over before, so they no longer read as nothing.
Why two readings. We measured both on the three labelled multi-column series in the Turing Change Point Dataset. The column-by-column reading scored mean F1 0.521, ahead of every published multi-column default (bocpd 0.480, ecp 0.462). It was still below the 0.621 you get by reporting no change at all, so treat it as a shortlist rather than an answer. The joint fit did badly on those series, cutting a stock price, a dancing bee's position and a room's temperature every 4 to 21 points. It earns its place on data that really is a level plus noise. In simulation, a shift of half a noise standard deviation shared by four columns was found 76% of the time jointly and never column by column.
A new warning. Since a joint fit that over-cuts looks just as confident as one that doesn't, the page and the CLI now say when it has more than one change per 50 points. We set that bar from runs already in the repository before scoring anything new. It fires on all six of the joint fits on the labelled series, and on German unemployment, two smooth quarterly series cut 22 times in 120 points. It does not fire on the currencies (9 cuts in 1,751 points), and on 20 white-noise sets with one real shared step it never fired (each got exactly one cut). It is a count, not a test. It tells you to read the column list instead, and it does not fix the joint fit. Fixing the joint fit, with a noise floor or trend per column, is the next piece of work.
Everything runs in your browser and nothing is uploaded. The method, the numbers and their limits are in the venture's README, under "Several columns on the page". These are statistical estimates about your data, not advice.