This looks related to Predicting out-of-sample for autoregressive models. The short version is that you are writing down a model for Y using X1 and X2, so you might use the parameters you learn with PyMC3 to generate samples going forward, or pad X1 and X2 with NaN's.