I suspected that the marginal likelihood computed from the SMC is not correct (see https://junpenglao.xyz/Blogs/posts/2017-11-22-Marginal_likelihood_in_PyMC3.html#Computing-marginal-likelihood-using-SMC for another example). However, I have not yet have time to diagnose and validate the implementation.
cc @aloctavodia. Maybe we should take down the doc before we can make sure the implementation is correct?