Estimating a latent jump process with Poisson observations

Here’s an example implementation with a discrete change point parameter:

See Case study 2: Coal mining disasters:
Introductory Overview of PyMC — PyMC 5.16.2 documentation

It’s much much more efficient to marginalize out the change points and absolutely critical if you need to estimate tail change point probability. Unfortunately, with multiple change points, marginalization cost grows exponentially in number of change points.