So this off-season I've been tinkering again with NFL modeling (), and after creating a ranking system I'm happy with that finally incorporates strength of schedule (hundreds of iterations per week), I've gotten very close to bettering the opening lines. The average error of my model over three seasons of out of sample data (in points) is 11.08 from the actual game outcome. Depending on the season, this represents a comparable line to the opener.
Albeit a fairly simple ranking process for now, I can't help but feel like this has value in some regard...and that carrot is once again dangling in front of my face: "How do I shave that extra half-point or so to ACTUALLY better the openers?"
For what it's worth, I am not using play-by-play in my model as of right now. So would it be reasonable to assume that without p-b-p data, this is about as low of an avg. error I will get?
Any brains to pick on this? Maybe someone who modeled NFL fairly extensively before play-by-play data started being used?
Thanks.