Defining random for a CustomDist time series

Questions v5
,
3

Thanks @ricardoV94, after sleeping on it I realized I needed to take a step backward.

In pymc3, when I used a GaussianRandomWalk, the sampler would “predict the future” for me if I passed missing data to the observed. Please see this gist.

In the new PyMC, this no longer works. Please see this gist. In that gist, I also show that when the GRW is a latent var, the sampler can predict it’s future; it just refuses to do so when the GRW is observed.

So my questions are:

  1. Am I right, or am I simply defining the model incorrectly?
  2. Why is this the new behavior?

In the 2nd gist, I go on to show that just as with the GRW, PyMC can predict the future of my CustomDist for day-of-week seasonality, but only when it is latent, not observed directly. This was the source of my original question in this thread: is it possible to have PyMC predict the future for my CustomDist class, just as it did for the GRW in pymc3?