LuckyNumber7
Blue Crack Addict
I'll try to dig up some references for you, but I'm pretty sure that Burke's research shows that using probabilities from the current game does a worse job of predicting future results in that game than using league-average results (which is what the model uses). In other words, using current-game probabilities (or making adjustments for such) is another small sample size mistake.
Also, saying that the Patriots had the ability to 'easily' get the ball back seems really optimistic to me.
I'm by no means totally discounting odds from a model, I just really don't think it does it justice. It seems to me as though those odds were good for "any game on any matchup." Meaning, a compiled statistic of all teams with that remaining time and that current score.
On the idea of using current in-game stats... you don't have to run into small sample errors from doing that. Instead, just use the splits, which could be conditionally set from the in-game stats. Playing Denver is a big deal. A limping Denver offense and a solid defense were big deals.