What are Expected Points Added (EPA) in the NFL

October 23rd, 2021 / EPA // Expected Points Added // NFL /

If you follow the NFL, and especially if you follow football analytics, Expected Points Added (EPA) is a metric you’ve seen being used more and more. But what is EPA? This post will help explain the intuition behind Expect Points, how Expected Points Added is calculated from Expected Points, and how to think about both in the context of analytics. 

What are Expected Points in the NFL?

The concept of Expected Points (EP) was first introduced in a 1970 research paper by Virgil Carter, who was the Bengals starting QB at the time, and Robert Machol, who was a professor at Northwestern. Their paper quantified a concept intuitively understood by all football fans -- possessing the ball closer to your opponent’s endzone is better than possessing it further away. Carter and Machol did this by adding together the point value and probability of all potential outcomes of a possession. As teams approach their opponent’s endzone, the probability of scoring (Touchdowns and Field Goals) increases.

probability-of-scoring-by-field-position

When a team possesses the ball at the 50 yard line, the probability that their drive ends with a touchdown is 30%, and because the value of a touchdown is 7 points, the expected points from touchdowns on the drive is therefore 2.1 (0.3 * 7). Adding together all the values for all potential outcomes yields the total Expected Points.

Simply put, Expected Points describes how many points, on average, a team is expected to score on a possession given a particular game context.

Application of Expected Points

Today’s Expected Points models have evolved to account for significantly more game context. Instead of simply considering field position, today’s models also account for down and distance, time remaining in the half, and the expected points the opponent gains when the ball is punted or turned over.

This unlocks more useful analysis like understanding the relative values of downs at particular points on the field:

expected-points-added-epa-nfl-by-down-and-field-position

When a team has 1st and 10 from their own 8 yardline, it’s actually more likely that their opponent will score next, resulting in negative Expected Points.

When a team has 1st and 10 from their own 30 yardline, they are expected to score as many points as when they have 3rd and 10 from midfield. From this, we can say that 1st and 10 from the 30 is worth the same as 3rd and 10 from the 50.

Adding distance-to-go provides additional context as to how many yards must be gained to “stay on schedule:”

expected-points-added-epa-nfl-stay-on-schedule

1st and 10 is worth about the same as 2nd and 2 for just about every position on the field, meaning a team needs to gain 8 yards per play to maintain their expected points on a drive (aka “stay on schedule”). Further, there is no point on the field where a 3rd down is worth as much as a 1st and 10, meaning the notion of “3rd and manageable” is actually a myth!

Deriving Expected Points Added (EPA) from Expected Points

On every play, a team has the potential to either increase or decrease their Expected Points. Expected Points Added is the difference between a team’s Expected Points at the end of a play and their Expected Points at the beginning of a play. 

Consider a team that completes a 20 yard pass on 1st and 10 from their own 20:

expected-points-added-nfl-calculation

The team started the play with an Expect Point value of 0.70 and ended it with a new Expected Point Value of 2.06. This team added 1.36 Expected Points with the 20 yard completion.

EPA simply connects the dots between two game states. If a team ended the play with more Expected Points than they started, then EPA will be positive. If a team is left less likely to score at the end of the play, then EPA will be negative.

In this sense, EPA effectively assigns a point value for every individual play. Though points are not scored on every play, EPA provides a framework for attributing points scored to individual plays.

Consider a drive that starts at the 50 and is followed by a run of 10 yards, a pass of 30 yards, and a touchdown run of 10 yards:

expected-points-added-epa-nfl-drive-level

The touchdown is worth 7 points, but because the drive started in advantageous field position with an Expect Point value of 2.8, the drive added 4.2 points of EP. Of that 4.2 points worth of EP, 0.6 was added by the first run, 1.7 was added by the pass, and 1.9 was added by the second run -- each play contributed to the touchdown, and EPA tells us precisely what that contribution was.

Though simple in concept, this is a profoundly useful framework for analyzing the game of football.

Applications of EPA

Every play comes with a set of classifying metadata -- Was it a run or a pass? Who was the quarterback? Was there play action? Combining play level classification with play level point attribution is the basis of nearly all the next-gen analysis being created in the NFL analysis community.

For instance, by grouping a team’s plays based on offense and defense enables a more precise telling of offensive and defensive performance than a simple “Points For vs Points Against” analysis does. You’ll see this approach used in EPA scatter plots that compare offensive efficiency (EPA Gained / Offensive Play) by defensive efficiency (EPA Allowed / Defensive Play):

expected-points-added-epa-nfl-team-tiers

Another point of aggregation is playtype (ie passing vs rushing). Calculating average EPA by passes and rushes reveals passing to be dramatically more efficient:

expected-points-added-epa-nfl-run-vs-pass

EPA can also be aggregated and averaged at the individual player level. For instance, EPA / dropback is a far better measure of a QBs contribution to the offense than passer rating or yards per pass:

expected-points-added-epa-nfl-qbs-quarterback

And a QB’s EPA can be broken down even further based on different aspects of the position -- how many points did a QB lose to sacks and turnovers vs the league average?

expected-points-added-epa-nfl-quarterbacks-qb-components

The ability to mix and match a vast array of meta variables through EPA creates a unique view into the game of football.

Shortcomings of EPA

EPA does have its limitations, most notably in its ability to measure individual performance. EPA analysis works for QBs because that position dictates so much of the game. However, football is a team sport, and attributing EPA to other positions like running-back or wide-receiver typically does not yield much signal. EPA can’t distinguish between a well run route and a well thrown ball.

A further limitation of EPA in its ability to capture non-point benefits or benefits not realized by the play in question. For instance, rushing EPA ignores any positive benefit a strong running game may create in the passing game by slowing down a pass rush or the benefit it may provide to a team in their ability to close out games.

A final limitation of EPA based analysis is the dataset itself. Most EPA analysis is derived from public play-by-play (PBP) data through the nflfastR package. This dataset does not have classifications for formation or scheme, which means EPA based analysis typically doesn’t have context for the degree of difficulty or intention of the play. Does a QB have high EPA because scheme put them in advantageous positions or because they really played at a high level?

Closing

Perhaps the best way to think of EPA is as a means of attribution and as a tool for high level accounting. It can provide deep insight with significant sample size and sufficient metadata classification, but struggles to tell us how valuable any individual player is.

EPA can tell us at a high-level that teams should be passing more, especially in certain situations, but it can’t tell us whether a coach’s decision to pass or run was good based on the defense they faced on that particular play.

In the end, EP and EPA are tools. The proliferation of EPA based analysis has created a step-change in our ability to dissect the game of football, but they leave many questions still unresolved.