In the NFL -- unlike college football or college basketball -- there are no committees, no "style points" and no subjectivity. If a team wins enough games to earn a divisional title or a wild-card spot, it will make the playoffs and possibly have a chance to compete for the Super Bowl.
For this reason, the ultimate goal when rating teams in the NFL is to measure team strength and project performance going forward.
With this objective in mind, ESPN's Stats & Information Group has created an NFL version of its Football Power Index, or FPI for short. There are a number of other NFL power ratings out there -- FiveThirtyEight (Elo Ratings), Pro Football Reference (SRS), Jeff Sagarin and others have created systems to rate NFL teams -- but FPI has a few additional features (like incorporating quarterback injuries) that sets it apart.
What is NFL FPI?
At its core, NFL FPI is a prediction system for the NFL. Each team's FPI rating is composed of a predicted offensive, defensive and special teams efficiency, as measured by expected points added per play, and that rating is the basis for FPI's game-level and season-level projections.
In the preseason, FPI uses a number of predictive factors to project future team strength. The main component of preseason FPI is Vegas expectations; the expected win totals and money lines for each team are an accurate representation of predicted team strength and provide a strong baseline for teams entering the season.
But relying solely on Vegas has its flaws, and more information is needed to determine what percentage of a team's projected win total can be attributed to its offense, defense and special teams units -- the components that make up FPI.
To gather more information on each unit, ESPN polled a panel of NFL experts regarding the expected offensive, defensive and overall performances of teams for the upcoming season. Also added to the model are previous years' efficiencies for each unit, number of returning starters (on offense and defense), coaching/coordinator/quarterback changes and quarterback injuries.
After combining all of these factors, a preseason FPI rating is determined for each team, which represents the points above or below average a team is expected to be in the coming season. Preseason FPI will serve as the basis of the early-season predictions but will diminish in effect as the season progresses and we learn more about the actual strength of each team.
Although team ratings provide fodder for debate, the ultimate goal of these projections is not to rate teams -- it is to predict performance going forward. The next piece of the puzzle for FPI is its game predictions.
What is accounted for in game predictions?
Like most game predictions, FPI accounts for team strength, opponent strength and home-field advantage. There are a number of unique inputs into each game prediction, however, that are worth highlighting:
On-field performance in previous games: Team performance is measured by expected points added per play, which helps control for the extremely fast- or slow-paced teams. EPA per play is a measure of efficiency that serves as the basis for how FPI evaluates individual units and quarterbacks.
Rest: Teams with a rest advantage of at least six days are 70-65 over the last five regular seasons. An extra week of rest makes a difference, particularly when facing a team coming off short rest. With all else equal, an extra week of rest is worth about 1 point per game, on average.
Altitude: There are only a few teams that experience an altitude advantage, but stadium altitude was found to be predictive. The biggest beneficiary of the altitude impact is Denver, which receives a small, but notable (about 0.3 points per game) increase in its chance of winning at home, compared to a team without an altitude advantage.
Distance traveled: Like with altitude, long travel distances only impact a handful of teams, but in the most extreme cases (say, Seattle to Miami), hosting a team with a significant travel distance is worth about half a point per game, all else equal.
Seasonal effects: Through six weeks last season, offenses were adding about 3.1 points per game to their net scoring margin and teams were averaging a record 1.91 points per drive. In the final six weeks, as the temperature (and offenses) cooled, the league-wide offensive EPA per game dropped to 0.7 points and teams averaged 1.76 points per drive. Whether it is the warm weather or the unfamiliarity with opposing offensive schemes, defenses have historically been at a disadvantage early in the season and have held the advantage later in the year. This trend is accounted for in the game-level projections.
QB injuries/suspensions/absence: A key differentiating factor for FPI's game-level predictions is its ability to account for quarterbacks missing games. On game day, knowing a QB's status is straightforward -- either a player is starting or he is not -- but the model also accounts for the chance that quarterback will miss games throughout the season. For example, looking out from Week 1, there is a higher chance that a quarterback will not play in Week 17 than in Week 5. But with each week that a quarterback remains healthy, the chances that he is available for subsequent weeks rise.
If a starting quarterback is out (or there is a chance he will be out), steps are taken to determine how much better that player is than his backup, and the difference between the two is accounted for in the game-level projection. Each quarterback's efficiency is determined based on past performance (using similar components as what we use to build up QBR), adjusted for an aging curve, and the players without any prior experience are set at replacement level.
All of these factors are combined to make up each single-game projection. Then each team's season is simulated 10,000 times to produce its chance to win its division, make the playoffs, win the Super Bowl and pick first in the NFL draft, among other interesting projections.
Ultimately the Football Power Index gives us a tool to project that future. Because of the level of detail in each simulation and the exhaustive process in building the model (see details on process here) we are confident that it will be one of the most -- if not the most -- accurate systems out for the upcoming season.