Model for finding arbitrary number of changepoints

General
1

Hello,

I am trying to implement a changepoint model. Perhaps a time series example would make it clearer:

Screen Shot 2020-03-31 at 4.47.08 PM
Screen Shot 2020-03-31 at 4.47.08 PM1998×1250 288 KB

Each color is a different segment, where each segment is modeled as a normal distribution with mean zero and a different standard deviation (in the picture, the y-axis are the displacements sampled from such normal distribution for the segment). The number of segments is arbitrary.

I implemented a model for this using maximum likelihood based on literature specific for my problem. However, I need to add a bit more complexity to the model so I want to move on to a Bayesian framework.

In the past, I’ve implemented change points using reversible-jump mcmc. I don’t think this is possible with Stan correct? Any suggestions?

Thank you.

2

Hi!
Have you tried the prophet package?! It has some methods for estimating change points! And You can always use their stan code is a bit tricky to follow and personalize but maybe the package itself is a good start :)

3

Thanks @asael_am! I will look into it. I am just concerned that it would not be flexible enough for what I need but will check it out. I am just at a loss as to whether just try to implement a model using plain RJ-MCMC in R, or whether there is a package out there that takes advantage of HMC for these kinds of problems.