100% agree that, regardless of whatever we both found in a small validation sample, there’s still likely a benefit in terms of anti-conventional wisdom arbitrage for the model as is. And yeah, was assuming this was just a first or second step; hope I didn’t come across as meaning otherwise.

In terms of going forward, I definitely concur with moving away from the point spread as your game script predictor, though it’s generally the most logical (and easiest) place to start. I’d imagine the solution would be coming up with your own win probability model that fills in the blanks of where the point spread goes way wrong. Or, before doing that, you could use Burke’s game probabilities (archived on his site via the NYT), since they’ve done pretty well against the spread over time. The other, more important, thing to address (to me, at least) is the last part of my previous comment. Don’t think you want the model to be biased at the extremes, with the residuals being dependent on a team’s P/R Ratio standing.

Good luck with the offseason tweaks. Eager to see what comes next.

Shortening the game is a high variance strategy. (It’s probably not a good one, but c’est la vie.) It may be selective memory, but it does seem like underdog teams have been better about taking risks this year. Jacksonville’s 2-point attempt against Denver was an excellent example.