Expected Points Added (EPA) is a football statistic that seeks to measure the value of individual plays in terms of points. This is done by calculating the Expected Points (EP) of the down, distance, and field position situation at the start of a play and contrasting it with the situation at the end of the play. A three-yard gain on first-and-10 is pretty different than a three-yard gain on third-and-two, something not usually captured in conventional statistics. The Expected Points framework helps translate raw gains into value.
The chart below, pulled from Pro Football Reference, illustrates the Expected Points Added for a drive late in the first half in Super Bowl XLIII, culminating in one of the most dramatic plays in Super Bowl history: James Harrison’s 100-yard interception return.
|Time||Down||ToGo||YL||Detail||Sc||EP Bef||EP Aft||EPA|
After an interception, the Arizona Cardinals take over on the 34 of the Pittsburgh Steelers. They start the drive with a 3.31 EPA, which makes sense, as they are already in field goal range. Two incompletions lower that to 2.08, before a 3rd-and-10 conversion to Tim Hightower raises the EP to 3.97, an EPA of 1.89, the drive’s best play. The second-best play by EPA is just a four-yard pass to Anquan Boldin, but it converts a 2nd-and-three and sets up the Cardinals at the one.
With first and goal at the one-yard-line, the Cardinals began the play with an Expected Points total of 6.97—a virtual guarantee of a touchdown (note that this is not 100% accurate given only 18 seconds remaining, but EPA does not consider clock). That’s when fate intervened, in the person of Harrison. Since the play resulted in a defensive score for Pittsburgh, the ending EP was -7.0. That’s a swing of 13.97 points—essentially two full touchdowns! Thus, the play scores as -13.97 EPA for Arizona and +13.97 EPA for Pittsburgh.
History and Variations
Former quarterback Virgil Carter holds a place in NFL history as the signal-caller around whom Bill Walsh’s designed his famous “West Coast Offense.” Walsh served as an offensive coach for the Cincinnati Bengals, who had the luxury of one of the most promising quarterback talents the game has ever seen: Greg Cook, the 1969 AFL Rookie of the Year and a prototypical big-armed passer who still owns rookie records for yards per attempt (9.4) and yards per completion (17.5). But Cook suffered a serious rotator cuff injury, and would only throw three passes in his career after that rookie campaign. In stepped Carter, who possessed smarts, athleticism, and short accuracy but nothing like Cook’s cannon arm. Walsh and legendary head coach Paul Brown developed an offensive attack centered around the short passing game to take advantage of Carter’s strengths and minimize his weaknesses, and an offensive system was born, one that would win the San Francisco 49ers five championships in the 1980’s and 1990’s.
What does this have to do with Expected Points? Well, when I said Carter was smart, I meant it: he went to BYU on an academic scholarship, earning a degree in statistics, and followed that up with earning a master’s degree from Northwestern while playing. And in 1971, as he and Walsh were making history on the field, Carter and Northwestern professor Robert Machol published a piece in Operations Research studying more than 8,000 plays from the 1969 season and calculating the expected point values of various field positions. The guinea pig for the Walsh offense was also the progenitor of the Expected Points framework.
The game has obviously changed a lot since 1969 and Carter and Machol’s figures are out-of-date, but Bob Carroll, Pete Palmer, and John Thorn dug up their research and developed an updated Expected Points framework for their 1988 book The Hidden Game of Football, a seminal work in pigskin analytics. They also developed Win Probability Added (WPA), a similar framework that calculated a team’s chance of winning based on the Expected Points framework, the score, and time remaining.
Brian Burke of Advanced Football Analytics brought Expected Points into the internet era in the mid-2000’s, taking advantage of more widely-available play-by-play to create a more robust model, including an EPA Calculator (sadly, no longer active). In 2015, Burke moved to ESPN, who already had a version of the EPA framework, both on its own and built into its Total QBR model for quarterbacks. Pro Football Reference (as shown above) has been reporting their own version of Expected Points and Expected Points added since the 2012 season. Football Outsiders has yet another variant called Total Points. Note that there is no one definitive way to calculate EP or EPA and various sites might have slightly different values for different situations.
Analysts have used Expected Points in a variety of ways. Berkeley economist David Romer used EP to study fourth-down decision-making. The New York Times later turned this framework into the Fourth Down Bot, which studies decisions to punt, kick the field goal, or try to convert in an automated fashion. Benjamin Morris of FiveThirtyEight used EPA to quantify how changes in field goal kicking influence such decision-making.
EPA also gives us a framework for understanding the contributions of individual plays, and thus, individual players. Average yardage metrics like yards per attempt or ANY/A are heavily influenced by outlier plays, overrates deep passing, and arguably doesn’t weigh sacks and other negative plays enough. EPA captures the value in third down conversions, red zone production, and plays that set up easy touchdowns even if they don’t produce scores. And since all game states carry an Expected Points, we can use EPA to compare between running and passing and between offense and defense and special teams.
That doesn’t mean EPA is perfect, however. Trey Causey observed that EP has increased over time, which tracks with offenses becoming more efficient. But that means calculations of EP either have to use a small data set from the current year, increasing the statistical noise, or go back to prior years that might undersell EP. Josh Hermsmeyer illustrates many of the challenges calculating EP in a recent post, summarizing EPA as “difficult to calculate, hard to explain, inherently noisy,” and not particularly stable.
EPA is also subject to the same criticisms that plague nearly all football statistics: small samples, inability to distinguish between a player and his teammates, blurred lines between what is descriptive and what is predictive, blurring situations that impact wins and losses and “garbage time” where the outcome is no longer in doubt.
There are no be-all and end-all statistics in football. But used appropriately, Expected Points Added is a powerful tool that can help solve a variety of problems, whether at the macro level or looking at individual players or decisions.