Ratings Explained

By Michael Gertz
Thursday, July 5, 2012 (Updated Thursday, February 23, 2017)

2015 Update

For 2015 and later seasons, the team ratings model has been changed from being based on our EPA stats to being based on our player ratings. A team's rating is the sum of all it's players' estimated point impacts per game. The player impacts are based on the player's rating, snap percentage, and position. Each position has a different mutliplier based on it's estimated influence on team ratings from an optimization of prior years. A full explaination of the current ratings model can be found here.

Pre-2015

The ratings at ProFootballLogic are formed by a computer model to best estimate the true quality of teams in the NFL. This means that rather than simply attempting to allocate a description for past results like other rankings, the ratings are geared toward the prediction of future results. The model essentially estimates how a team will perform in a number of play types based on how they have performed so far, using historical league results as a guideline. League averages are also used to factor in home field advantage, and the process of forming the ratings automatically adjusts for strength of opponents. Ratings for each play type are then summed for total offense, defense, and overall ratings. They represent the number of net points expected to be generated relative to a league average in each category, with the difference in overall ratings between two teams representing a point spread expected should they play at a neutral site.

The ratings take into account the play type and basic results of every NFL play from the current season. They are completely objective team ratings based purely on statistical information, and therefore do not specifically account for player changes or injuries apart from how such things have already impacted past performance. Because the ratings do not currently use information from previous seasons, they function like BCS ratings, considering all teams equal to start, and being much more reliable later in the season.

Taking only the current season of data has the negative impact of making ratings less reliable very early in the season, but the positive effect of adjusting more quickly to teams that experience drastic changes in quality from one year to the next, especially in cases of teams changing quarterbacks. During the unreliable ratings early in the season, users can lean heavily on a team's ratings from the prior season, as most teams' true quality does not change much in just one season barring significant quarterback changes. Knowledgeable users will be able to spot impactful roster changes and weigh both ratings to compliment their own conclusions.

The theory behind the model is based on the notion that like most naturally occuring events, variation in performance of teams is governed by a normal distribution, or bell curve. The distribution governs how often teams perform at a level close to or far from their average performance, with half of results better than average and half worse. Variation in football comes from players physically not being able to perform the exact same every play, the combination of offensive and defensive play calls affecting each other, and referree imperfection.

Because the addition of normal distributions creates another normal distribution, these unpredictable variables can be clumped together and treated as random variation in the model. Likewise, separate distributions can be estimated for each play type which together combine to form an overall distribution of team quality.

The model is based on averages from multiple seasons of our EPA stats data that establish values for the amount of random variation in each play type for offense and defense. These values represent exactly how likely results are to be repeated, as play types which are less reliable are naturally assigned higher levels of random variation.

Less reliable play types are instances of the game where no difference in quality can produce a large difference in score. This can either be due to the nature of football, or instances where either the offense or defense has a larger impact, making the results associated with the other more unreliable.

At the same time, values for the average team quality and amount of variation among team quality for each play type are estimated. These also follow a normal distribution, as the number of teams of each quality in the league follows a normal curve.

With the model having established baseline curves specific to the average variations shown among teams and play types in the NFL, calculating ratings for a given season is simply a matter of optimizing to find the most likely ratings for all teams at once.

Home field advantage is factored into the game results at this stage. Because results for each play are defined as the combination of offense and defense, strength of opponent is automatically accounted for in the optimization process. And because all teams are judged by the odds that they will perform at a certain level of quality in the future, the model automatically accounts for the small sample size early in the season.