As a data scientist, this makes me tingle in places we don't talk about at parties. I love it.
One big question for me: did you remove any players that have gone down with season-ending injuries and recompute?
Hurd, Tuttle, McKinsie, JRM, etc are on the roster... but they shouldn't be used when calculating the one-game prediction. Our talent level that we'll put on the field will be different than the roster talent level.
Suggestion for version 2 of this model: apply a multiplier for each year of experience in the program. A true freshman would receive a 0.75X multiplier, while a senior might get a 1.5X multiplier. If you stacked the team FULL of 4* and 5* players, but they were all true freshmen, that wouldn't be a very good team. A bunch of talented seniors is a different ballgame.
You are correct, there are ways to increase the predictability of this model. But, at some point the time invested goes near vertical with very little return. As a free analysis, posted on a message board, used to illustrate likely outcomes and seasonal expectations 70/30 is about as good as you're going to get. Understand that there are more advanced analysis that I simply cannot, nor will I share, as I view those as being the property of my employer. I don't want to ever entangle my work with my employer's.
I have been doing this since approximately 2011, and I have tested almost every variable that one could imagine. In fact, I spent the better part of a year inputting and analyzing all sorts of variables and data for my employer. In years past, I have adjusted teams for attrition. I have also tried to weight teams based on experience. Both are popular ''what if'' scenarios. Neither really tweaked the outcome in a way that was worth the effort. In fact, of the samples that I tested, while the evaluation of a team might have adjusted in relation to itself, the relative evaluation didn't change much at all. In the rare event that there was a change, one way or another, it was rather insignificant. Are there cases where attrition and experience make huge differences? I am sure there are. They are the anomaly as far as I can tell.
My hypothesis is that talent is effected more by experience when the relative talent is low. For example, Pinkel at Missouri had an interesting trend. He recruited consistently in the mid thirties (quite mediocre). His coach effect trend line was 3 years or so of performing to talent expectations, then a two year bump of exceeding expectations. The irony is that he entered the SEC right as that explosion occurred, right on time. His recruiting was consistent, only the performance in relation to that relative talent changed. I believe that is a verifiable case of experience modifying talent. I don't tend to see that much with teams that recruit in the top 3rd. That isn't to say those situations don't exist, only that the change in attrition and experience are within ranges across the majority of teams to not greatly effect their relative position to each other.
I hope that answers your questions.
EDIT: As an aside, here is a graphic I used to gauge the coach effect of a few "hot name" coaches back to the 2014 season. Note Pinkel's performance in relation to talent to help illustrate the point I was making. If you were to add last year to it, the performance returned to expectations.
