How EPA Values a Play—and Why Soccer Differs

EPA means Expected Points Added in American football: it estimates how much a play improves or worsens the offense’s expected scoring position. If you saw EPA in soccer data, do not assume it has the same definition; EPA is not a provider-independent soccer statistic.

Choose where you saw EPA to get the definition and checks that apply.

How Expected Points Added Works

An expected-points model assigns a value to the game situation before the snap. That value is based on factors such as:

• down
• distance to a first down
• field position
• time remaining
• other game-state variables used by the model

The model values the situation again after the play. In its simplest form, EPA equals expected points after the play minus expected points before the play.

The implementation must also account for points scored. If possession changes, the post-play value must be expressed from the original offense’s perspective.

Positive EPA means the play improved the offense’s expected scoring outlook. Negative EPA means it made that outlook worse. An EPA of zero means the estimated scoring position did not change.

EPA therefore measures more than yards gained. The value of a gain depends on what the offense needed, where the ball ended up and what situation the play created.

The Same Eight-Yard Gain Can Have Opposite Values

An ESPN explanation of its expected-points model compares two plays that begin at the offense’s own 20-yard line:

Situation	Result	ESPN EPA	Interpretation
Third-and-10	Gain of 8 yards	−0.2	No first down
Third-and-7	Gain of 8 yards	+1.4	First down converted

The yardage is identical, but the new game situations are not. The third-and-10 play leaves the offense short of the first down and facing fourth down. The third-and-7 play converts and keeps the drive alive.

EPA captures that difference. A short conversion can be more valuable than a longer gain that fails to move the chains.

The precise values are specific to ESPN’s model and data. Another expected-points model may assign different numbers while reaching the same broad conclusion.

EPA Per Play Measures Average Efficiency

A single play’s EPA is the change attributed to that play. Analysts commonly aggregate the measure in two ways:

• Total EPA: the sum of EPA across all included plays.
• EPA per play: total EPA divided by the number of included plays.

Total EPA combines efficiency with volume. An offense that runs more plays has more opportunities to accumulate it. EPA per play instead describes average efficiency across the selected set of plays.

For offense, higher EPA per play is generally better. For defense, lower EPA allowed per play is better. Some defensive tables reverse the sign so that stronger performance appears as a positive number, making the column label and methodology necessary context.

Play Filters Can Change the Number

The denominator is not always identical across sites. Analysts or providers must decide which plays belong in a passing, rushing, team or player sample.

In one nflfastR team-comparison example, sacks and quarterback scrambles are classified as pass plays. Plays involving penalties can also be included, according to the nflfastR guide.

Two sites can consequently report different EPA per play figures without either calculation being a simple arithmetic error. They may use different expected-points models, include different plays or classify the same play differently.

Before comparing EPA figures, check:

• whether the value is total EPA or EPA per play
• which play types are included
• how sacks, scrambles and penalties are classified
• whether defensive EPA is shown as EPA allowed or with its sign reversed

EPA Is Not Win Probability Added

EPA measures the change in expected scoring value. Win probability added measures the change in a team’s estimated chance of winning.

The measures often move in the same direction, but they answer different questions. Late in a game, an offense protecting a lead may benefit substantially from keeping the clock running even when the play adds little expected scoring value.

Expected points should not be read as a literal prediction of how the current drive will end. The nflfastR model publishes probabilities for the next scoring event within the half. Those possible outcomes include an opponent score or no further score, not only the chance that the current possession produces points, as described in the nflfastR expected-points documentation.

EPA in Soccer Requires a Provider Definition

Soccer action-value models can use a related idea: compare the attacking value before and after an action. Their names, units and definitions vary, however, so an EPA label on a soccer platform cannot safely be interpreted from the initials alone.

Stats Perform, for example, described a Possession Value model that estimates the probability of a team scoring during its possession. Its Possession Value Added (PV+) is the change associated with an action. In one published sequence, the estimated scoring probability rose from 3.3% to 33.9% after a carry and through-ball, according to Stats Perform’s framework.

That approach resembles EPA conceptually because it values the change created by an action. It is not the same measure. American football EPA is expressed in expected points and incorporates down, distance and field position. The cited soccer model is expressed as a change in goal probability during a possession.

When EPA appears in soccer data, check the provider’s glossary or methodology for:

• the metric’s full name
• whether its unit is expected goals, goal probability, points or another value
• whether it evaluates an action, possession, match or larger sample
• the denominator used for any rate statistic
• whether values from different competitions or providers are comparable

The practical distinction is straightforward: in NFL or college football analysis, EPA normally means Expected Points Added. In soccer analysis, its meaning remains provider-specific until the platform defines its unit and method.

The same caution applies to possession statistics: a familiar label does not guarantee that two providers counted or valued the same thing.