Well, for starters I would use the actual stats, not just the rankings. And I wouldn't arbitrarily give weightings to the categories unless I had proof that certain ones should be worth double.
Basically, I would run a linear regression to find the appropriate weights for the different stats. So, instead of looking at team ranking in point differential, yardage differential, TO differential, etc., I would take the actual numbers in those categories and see which ones correlate most highly to actual winning percentage. If you can find historical data (like this), you could dump a few years' worth into Excel, regress on wins, and apply your formula to this year's data to find how many games each team "should" have won based on their stats. And from there, you could even make an SOS adjustment, or what have you.
Anyway, when constructing a statistic, I think people should always make sure their weights are derived from some sort of regression (which will be the best hope of reflecting reality), instead of just arbitrarily picking weights. If you get a model that's a pretty good predictor of past wins from game stats, it will be able to give you a good picture of who the best teams are this year, based on their stats.
