I just encountered this problem, so thanks @lucianopaz and @ricardoV94 for the great explanations, and @zlegge for the great question!
FWIW I wanted an example that combined everything, and added explicit drift, so created a gist here: Example model for holdout (future) set PPC using a GaussianRandomWalk in pymc3 · GitHub - possibly it helps the next reader 
I’m not sure if the poor performanace is due to the nature of the data, or using the wrong likelihood (tried Poisson and Lognormal but they were dreadful), or the model is missing e.g. @junpenglao’s smoothener / identifier trick …