2015 NFL Team Ratings Model

By Michael Gertz
Tuesday, October 20, 2015

Prior to the 2015 season, our NFL team ratings were derived solely from our expected points added stats for each team. The stats were adjusted for strength of opponent at the same time as they were regressed toward the mean a different amount for each individual play type to account for how predictive each play type was. While helpful and relatively predictive, the ratings had two main flaws.

The first is that we treated each play type independent of the others. In reality, some play types are slightly correlated with others, especially on defense. The result was ratings that were slightly too regressed toward the mean. This aspect could have been somewhat improved if we adjusted the model to include more seasons of data in order to account for the correlated play types, but even then accurately estimating the exact levels of correlation is very difficult to do.

The second problem was that because the ratings didn't account for individual players, they were not very useful at the beginning of each season, and became inaccurate when significant injuries occurred. Even if the first problem was fixed relatively well, this second problem would still plague the ratings, so a significant overhaul of the ratings model was in order.

The New Team Ratings

Our new Team Ratings are formed completely from our player ratings. They still have the same meaning and scale of expected points added, and are able to produce expected game spreads and odds. But now they are based on our player ratings rather than our expected points added stats.

In order to form the model, a significance value for each position was optimized such that past player ratings produced accurate team ratings, while accounting for what percentage of snaps each player played. Because there are no perfect team ratings to optimize the positional values from, our previous team ratings and team stats were used as a baseline, with some adjustments made by hand afterward as a sort of reality check to make sure the values weren't overfitted to the specific data set. Even though the optimization included all of the 5 most recent NFL seasons, like the correlations of team ratings above, even more seasons than that would be needed for a perfect model.

The result is a model that can give offensive and defense team ratings for any specific team's combination of players and expected snap counts, which we adjust each week to account for player injuries. The model largely follows our expectations for how important each position is, but does provide some new insights that were unexpected. The exact values will continue to be tinkered with slightly in order to produce the most accurate team ratings possible.

What Does it Tell Us?

We won't reveal our exact values for each position at this time, but the values are consistent with the fact that the modern NFL is dominated by the passing game. Obviously, quarterbacks are the most important position. But the level of value is staggering; they are about 5 times as important as any other position. While this does not agree with the contracts that quarterbacks often sign, it does align very closely to how betting lines move when quarterbacks are injured. The matter of QB importance and these topics specifically is something that we will address in more detail in future articles.

Other than QBs, the rest of the positions on the football field are relatively similar, but some positions are up to about twice as important as others. In general, the positions that are more congested and require heavier players are generally less important than positions that require more athleticism in open space and are usually more imvolved in the pass game. This mirrors the current trend in the NFL to spread formations out and pass more.

One of the biggest surprises from the model is the strong importance of tight ends, and to a lesser extent, coverage linebackers. These positions do not typically garner large paychecks, but upon further reflection, an argument for their importance could be made. They are perhaps in the sweet spot such that they still occupy large spaces to have significance in the pass game, but are also significantly involved in the run game. While linemen also factor into both the run and pass games, their importance in both is probably overstated, especially in the pass game.

The most overrated positions appear to be linemen, and in particular, pass rushers. While pass rushers typically command contracts as large as about any other non-QB position, they don't appear to have the significance to justify such contracts. While sacks feel very devastating, and it can occasionally feel like a team's pass rush is dominating a game, the reality is that even the best pass rushers contribute only a sack and a few pressures per game. Compare that to the fact that even a low level pass rusher will generate about a half a sack and a pressure per game, and the difference simply isn't very large.

Going Forward...

While the new model factors only player ratings and not team stats into the equation, our expected points added team stats are constantly used to inform decisions about manually updating the player ratings used. In the future that process may follow a more strict formula if one can be established, but for now it follows a more subjective nature.

While the basis of the new model is providing accurate team ratings to project games, it opens up many new lines of analysis for us to explore that weren't possible previously. In the future, look for analytical articles to explore topics such as individual player value, player salary models, and draft pick valuation.