<
>

An inside look at College FPI

Rich Barnes-USA TODAY Sports

ESPN’s College Football Power Index (FPI) was developed in 2013 as a way to measure team strength and predict game and season outcomes going forward. Over the years, fans have clamored for ESPN to open the “black box” of FPI and share more information.

In the interest of transparency, below is everything you need to know about college FPI.

What is college FPI?

FPI is a predictive rating system designed to measure team strength and project performance going forward. The ultimate goal of FPI is not to rank teams 1 through 128; rather, it is to correctly predict games and season outcomes. If Vegas ever published the power rankings it uses to set its lines, they would likely look quite a lot like FPI.

Correctly predicting game outcomes can’t be done by evaluating teams’ records because some teams are stronger than their records indicate (lots of close losses), and others have favorable schedules. Both of these situations are reflected in the game- and season-level projections.

It is important to note what FPI is not -- FPI is not a playoff predictor, and it is not designed to identify the four teams most deserving of making the College Football Playoff. ESPN has other metrics, including Strength of Record, that can be used to identify the most deserving teams.

What goes into FPI?

Each team’s FPI rating is composed of a predicted offensive, defensive and special teams component. These ratings represent the number of points each unit is expected to contribute to the team's net scoring margin on a neutral field against an average FBS opponent.

In the preseason, these components are made up entirely of data from previous seasons, such as returning starters, past performance, recruiting rankings and coaching tenure (more on the preseason component below). That information allows FPI to make predictions (and make determinations on the strength of a team’s opponents) beginning in Week 1, and then it declines in weight as the season progresses. It is important to note that prior seasons’ information never completely disappears, because it has been proved to help with prediction accuracy even at the end of a season. Vegas similarly includes priors when setting its lines.

Once the season is underway, the main piece of information powering these offensive, defensive and special teams predictions is past performance from that season’s games, in terms of expected points added per game. Expected points added, or EPA, is a measure of success/failure that takes into account yards, turnovers, red zone efficiency and more to determine how many points each unit is contributing to the team's scoring margin. For example, if a team wins by an average of 10 points per game, it could be that plus-seven of that is offense, plus-four is defense and minus-one is special teams. Because expected points added is built on play-by-play data, it’s fair to say that FPI looks at every play of every game in the season.

Expected points added on offense, defense and special teams are individually adjusted for each game based on the strength of the opposing unit faced and where the game is played. Additionally, FPI applies a capping of sorts to each of these components to minimize effects of blowout games and improve prediction accuracy. As we learn more about the true ability of each team, FPI retroactively readjusts each game within the season using the team's latest predicted components.

In conjunction with the opponent adjustment, FPI uses a Bayesian regression to update each team’s offense, defense and special-team components, which combine to produce the rating. This is an iterative process that is constantly updating and improving itself after every game of the season.

What goes into the preseason ratings?

Preseason ratings historically have their flaws, but ultimately they allow for an opponent adjustment after Week 1 and are a great tool to preview the season. As noted, there are four components to the preseason rating: prior performance, returning starters, recruiting rankings and coaching tenure.

-- Prior performance is built off the framework of expected points added. The most recent year’s performance is by far the most important piece of information powering preseason FPI, but three more years are added to measure consistency and account for outliers in performance. The most recent year counts almost twice as much as the three years before it.

-- Returning starters on offense and defense, with special consideration given to starting quarterbacks or transfer quarterbacks with starting experience, is the second piece of information powering preseason FPI. Because starters interact with other inputs, it’s not as simple as saying an extra returning starter is worth one point. Nonetheless, a starting quarterback is worth about 3.3 points per game to a team returning an average offense (all else equal), and a transfer quarterback is given half the weight of a starter.

-- FPI uses four recruiting services -- ESPN, Rivals, Scouts and Phil Steele -- to measure the talent on a team’s roster and add an additional piece of information about which teams are on the rise. The addition of recruiting has been a controversial piece of FPI, but it’s worth noting that it is a very minor component that helps with prediction accuracy.

-- Coaching tenure is primarily a way to capture the addition of a new head coach. With all else equal, a team’s predictive offensive, defensive and special teams ratings will regress slightly to the mean with the addition of a new coach.

Preseason FPI debuted in 2014, and you can read more about how it performed in these recaps of the 2014 and 2015 seasons. In short, if preseason FPI, which was run retroactively to 2005, had been used with no update to predict every game over the last 10 seasons, the FPI favorite would have won 72 percent of FBS-versus-FBS games (Vegas closing line was 75 percent accurate).

FPI’s game and season predictions

FPI’s 1-through-128 rankings are fun to debate, but the ultimate goal is to correctly handicap games. FPI’s game predictions begin with each team’s FPI and then add information on game site, number of days of rest, distance traveled and game type (bowl game, conference championship game, regular season or non-FBS).

For example, an additional 5 1/2 days of rest more than your opponent is worth one point per game (all else being equal), and every additional 1,000 miles traveled more than your opponent costs you a point.

Each team’s schedule is simulated 10,000 times to produce season-level outcomes such as each team’s chance to win its conference, enter bowls undefeated and make a bowl game.

In the season projections, the importance of a team’s schedule and path to a conference championship cannot be stressed enough; two teams with the same FPI can have drastically different projections, given their schedules. Even when those teams are in the same conference, their chances to win that conference can differ significantly given their divisions and competition within those divisions.

FPI’s accuracy

Over the last 10 seasons, the FPI favorite has won 75 percent of FBS-versus-FBS games, which is comparable to the Vegas closing line. Looking at the last four seasons, that percentage has risen to 77 percent, and in games that FPI and Vegas differed, the FPI favorite won 55 percent of the time. If you want to follow along with how FPI performs throughout the season, feel free to go to the prediction tracker website.

It’s worth noting that the results of analytics such as FPI are not black-and-white -- they give us likelihoods of outcomes, not certainties. Therefore, when FPI gives a team a 75 percent chance to win and that team loses, FPI is not necessarily “wrong.” A team with a 75 percent chance to win should lose one out of every four times, and if every team with a 75 percent chance to win does in fact win, the system is broken.

Therefore, we track how well a system is calibrated, and as can be seen in the chart above, when FPI gave a team between a 70 percent and an 80 percent chance to win, those teams actually won 73 percent of the time.

Of course, no system will be 100 percent accurate, and every year there are teams that FPI is wrong on. Michigan State is one team that has consistently outperformed FPI’s expectations over the years. As college football fans, we do not agree with every prediction or rating, but in total, FPI has proven to be accurate.

Full FPI rankings are available at ESPN.com/fpi, and each team’s game projections are available by clicking on that team from the FPI page.