The only "Bayes" I can see is the author starts with a uniform prior and applies Bayes rule in one of the examples, and in fact the whole thing is kind of all over the place.
His approach of dragging a cut point across the data and asking "is it significant here? how about here?" is going to be prone to all sorts of false positives: if you set your threshold to 0.05, you'll expect one positive every 20 looks at the data. (He's looking for overlaps of credible intervals; this is going to work out to be the same thing).
Thanks for commenting! The whole technique is foundational Bayes. Doesn't use MCMC, but still bayes even with the use of credibility intervals. I don't make much of a use of priors to your point and that can definitely be seen as somewhat controversial, but each window is aggregating information on all of the hypotheses given all of the data. That's why I qualify it as "bayesian". How do you typically qualify bayesian vs not?
To your point of the issue with false positives... yep! :) It's a very simplistic naive approach. I read over the article you linked to and I think the results are fairly similar except that my (again very simplistic) technique finds many change points if they exist... and possibly even none. So what I'm doing generalizes more.
The sequential nature of what I'm doing bothers me too, so I'd love pointers to other articles doing something similar.
2 comments
[ 4.9 ms ] story [ 19.4 ms ] threadHis approach of dragging a cut point across the data and asking "is it significant here? how about here?" is going to be prone to all sorts of false positives: if you set your threshold to 0.05, you'll expect one positive every 20 looks at the data. (He's looking for overlaps of credible intervals; this is going to work out to be the same thing).
Here are a couple useful links for doing this kind of problem: https://qualityandinnovation.com/2015/07/14/a-simple-intro-t... http://arxiv.org/abs/0710.3742
To your point of the issue with false positives... yep! :) It's a very simplistic naive approach. I read over the article you linked to and I think the results are fairly similar except that my (again very simplistic) technique finds many change points if they exist... and possibly even none. So what I'm doing generalizes more.
The sequential nature of what I'm doing bothers me too, so I'd love pointers to other articles doing something similar.