Ah ok I think I figured it out, it seems to be because I set the “y” values to NaN in my out-of-sample dataset. Somehow that triggers pymc to drop the whole dataset so nothing happens, or some such? But then it works ok if I specify var_names for vars that don’t get impacted by those NaN’s?
Anyway if I set the dummy “y” values to zero instead of NaN then something happens at least. I haven’t checked that the output is sensible yet, but it made sense dimensionally at least.
Is that the expected behaviour? NaNs should not be used to indicate absent dependant variable values? A warning message or something might be helpful here.