I believe that you may be uses a sort key in your mapping. You cannot do sim/pred with a sort key.
Take your data set and split it by that sort key and then fit each sub-data set separately. I do not see any other reason why you would not be able to perform predictive checks other that that. Let me know.