Hi Benjamin! Thanks for sharing the link. I had a look at @ckrapu’s code and I’ve been playing with some of the ideas proposed there. However, it seems that their implementation requires to propose a maximum number of changepoints M that you would expect and (more importantly) places those proposed changepoints equally-spaced across the data, i.e.
changepoints = pm.Deterministic('changepoints', (2 + at.cumsum(changepoints_raw) * (T-4)))
Not sure if there’s a way to make this more flexible and adapt the location of the changepoints.