@ricardoV94 thanks for the pointer! I watched the tutorial and found it very interesting.
However I believe that there should be something simpler than keeping around the original data and then adding a sample with the target y and missing x, then running a full .sample() again.
I am struggling to understand how to do something similar using .sample_posterior_predictive() instead.
I added a gist with a notebook and a link to the blogpost “Out of model predictions with PyMC” above. Would be great to know if I am on the right track there.