I've got a trained horse racing model and can predict from this as the prediction variables are known at runtime :- Odds, weight etc. This works very well and I then turned my eyes to tennis. I can quite easily train a model on various stats from the past and the prediction looks OK. The problem is, a lot of the past results are not known before hand of course like number of sets won, server speed etc. How would I predict a player for values not known or am I just trying to do the impossible? I thought about averaging the values perhaps?