Models favor stats that are indicators of success over a large number of snaps. Success rate, for example, is a stat that is relevant for every snap of every game. Compare that to something like 3rd and long % allowed on defense or 4th down % allowed on defense - those might only be relevant for a couple of plays in any given game so the predictive value is seen as much lower.
So these models take in a bunch of different stats and give more weight to the ones that are seen as being more important over lots of snaps, then they simulate a game like 1000 times to generate a prediction. So while we freak out because allowing Dekkers to convert a bunch of third and longs is horrible and almost cost us the game, the computers might have only had that happening in 80 out of 1000 simulations, and we still won. So as long as we keep having good results on most plays, they're going to keep liking us.
We're #13 in offensive success rate and #18 in defensive success rate (not sure if those numbers include ISU). I think tOSU, Michigan, UGA, Syracuse, and Minnesota are the only teams above us in both stats.