One of our offseason projects at Sports Info Solutions has been to create a weekly projection system for college football like we have for the NFL. In testing our projection of a team’s run/pass ratio in a given week, we found that we were able to do a much better job projecting that for college than for the NFL, which seemed worth digging into.
As it turns out, NFL run/pass ratio is remarkably inconsistent week-to-week, particularly when compared to college (specifically FBS). For consecutive games the last two regular seasons, the correlation in the NFL was 0.15, compared to 0.46 in college.
That makes sense to some extent since college teams are more likely to have rigid offensive systems that won’t be as fluid week-to-week. Triple-option teams are the obvious extreme of that phenomenon; if we remove Army, Navy, Air Force, and Georgia Tech from the above comparison, the college correlation drops to 0.34.
Let’s focus on the NFL for now, since it seems that the college game is more projectable on its face. What’s going on here?
Getting into the Flow
If you play fantasy football, you hear the words “game flow” or “game script” a lot. Team X is a ten-point underdog, so they’re likely to be down in the game and throwing a bit more than usual. The inverse might be said about the team they’re playing (up more = running more). These factors don’t turn weekly projections upside down, but they contribute to the discussion of why players might, in a given week, produce slightly better or worse results than they have in the past.
It makes sense that game script is one of the primary factors making NFL run/pass ratios less stable week-to-week. Teams are generally close in talent level, so it is common for a team who typically runs more than average falling behind early in a couple of games, leading to anomalous passing numbers. As you can see below, you can feel pretty confident that the winning team is going to run more than the losing team, regardless of what you know about the teams ahead of time.
Of course, this plot does not suggest that running the ball leads to winning games. This is a purely correlational result; the winning team tends to have run more than the losing team. And if we can find ways to identify when teams are more likely to run or pass, regardless of the reason, that’s enough for fantasy football purposes.
That said, if we’re trying to assess this before the game happens, even if you’re very confident in who will win the game, the run/pass ratio is pretty elusive.
Predicting the Game Result First
To get at this question, we calculated a proxy for each team’s likelihood of winning using their average total Expected Points Added (EPA) in their previous six games and comparing it to their opponents’. In this case, “average total” means the difference of their EPA on offense and their EPA allowed on defense in each game, averaged over those games. For each game, we took the difference between the totals of the two participants, giving us a sliding scale of how much of a mismatch the game was, and who was favored.
To show that this was a good proxy, we checked how these differences aligned with the probability of a team winning. Just knowing which team had a higher recent average allows you to pick the right team 63 percent of the time. Beyond that, there’s a clear relationship between the size of the EPA difference and the likelihood of winning. Below is a chart of the win percentage of teams with varying degrees of mismatch with their opponent, and it maps well to the line representing a perfect 1-to-1 correspondence, and even better to the shape of a logistic model you might typically use for such a prediction.
It should be mentioned that we also calculated a strength-of-schedule adjusted version of the metric, but it didn’t improve the prediction, so it was scrapped in favor of the simpler version.
Predicting Game Flow
So, if we feel pretty confident that we know who’s going to win, can we predict the teams’ run/pass ratios in that game? To start, let’s look at how this metric predicts run/pass ratio overall.
Feel free to guess the correlation, but suffice it to say there’s not much there to see.
So, it seems there isn’t a meaningful relationship over the course of the entire game. But if run/pass ratio is dictated by game script, it will be most impacted towards the end of the game when the good team is most likely to be salting the game away. How well can we predict second-half pass percentage?
There’s something there in the second half, but nothing to write home about. And for what it’s worth, splitting this into just the fourth quarter looks about the same.
We also considered not using the EPA difference at all but instead using the defense’s recent trend as the basis for the projection. In this case, that meant using the Log5 method — commonly used in baseball to project stats in specific batter-pitcher matchups — to estimate a team’s pass percentage. That proved similarly ineffective in this case.
Where We Go From Here
This is a weird finding given what we generally understand about how the mix of run and pass works in football. Even though we were able to predict with some confidence which team would be winning (and therefore who was running more in the fourth quarter), that didn’t really make a dent in our overall ability to project run/pass distributions. This is a discouraging result for those trying to apply the typical set of heuristics to make decisions about a player’s projected opportunities in a given week.
However, this is also not a definitive rejection of those heuristics. Both team EPA and recent run/pass ratio are blunt instruments to try to predict a game’s run/pass split. There are more refined predictors we might pursue, and we’ll be keeping an eye out for opportunities to advance this discussion in the future. But for now, if there isn’t enough evidence that we can consistently predict game flow, we might be better served to de-emphasize it for fantasy purposes.