What Is EPA in Football? Expected Points Added Explained

Key takeaways

• EPA evaluates a play in context rather than treating every yard as equally valuable.
• Positive offensive EPA means an improvement in expected scoring value. Negative EPA means a decline.
• EPA per play is an average, not the total value accumulated across a game or season.
• Different models and play filters can produce different results.
• EPA describes what happened under a model. It does not, by itself, guarantee future performance or isolate one player's skill.

Expected points comes before expected points added

To understand EPA, first separate it from expected points, or EP.

EP estimates the scoring value of a game situation. Down, distance and field position help describe that situation. Time remaining and other contextual inputs can also matter, depending on the model. A first down near the opponent's goal line is a different opportunity from a third down deep in a team's own territory. [1][2]

Modern EP models can consider possible scoring outcomes for both teams. That is why expected points should not be interpreted simply as the points an offense is guaranteed to score on its next drive. In nflfastR, EP is recorded from the possession team's perspective, while its model documentation includes both possession-team and opponent scoring outcomes. [1][3]

Think of EP as an estimate attached to a state of the game. It can be a decimal even though actual football scores use whole numbers. An average scoring value across possible outcomes does not have to equal a single scoring event.

These estimates come from a model. There is no universal table that assigns the same EP value to every situation regardless of season, provider or model version.

How EPA is calculated

For a non-scoring play where the same team remains in possession, the basic teaching calculation is:

That subtraction asks a straightforward question: did the offense finish the play in a more valuable situation than the one it started in? [2]

A completed pass is not automatically a positive-EPA play. The value of the resulting situation matters. A short gain that leaves a difficult fourth down can be less useful than the same gain on an earlier down.

The simple subtraction above is not a complete recipe for every football play. Scoring plays require accounting for points scored, and a change of possession requires handling the opponent's expected value from a consistent team perspective. Without that adjustment, comparing two EP values recorded for different possession teams can give the wrong interpretation. [1][3]

The principle remains the same: measure how the play changed scoring value. The implementation needs to respect who has the ball and whether points were scored.

A fully worked hypothetical EPA example

The following example is hypothetical. Its EP values are chosen to demonstrate the arithmetic. They are not actual NFL statistics, a lookup from a provider's model, or SportQuantix model output.

Suppose an offense faces third-and-5 at its own 40-yard line. It completes a 10-yard pass, earns a first down and finishes at midfield. The same team still has possession, and the play does not score.

For illustration, suppose the model assigns these values:

The offense has added 0.8 expected points in this teaching example. It has not scored 0.8 actual points. The scoreboard is unchanged, but its modeled scoring position has improved.

A different EP model, or different game context, could assign different values to those situations. The numbers here exist only to make the subtraction clear.

What EPA per play means

Total EPA and EPA per play answer different questions.

Total EPA adds the values of the included plays. EPA per play divides that total by the number of included plays:

Here is a second hypothetical example, independent of any real game:

The total is +0.6 EPA. The average is +0.20 EPA per play. They are not interchangeable.

Rates make comparisons easier when teams have different numbers of plays, but only if the underlying definitions match. Before comparing two figures, check which plays each includes and whether both use the same model.

The same denominator issue appears in NBA offensive rating: an efficiency rate and a total describe different things, even though basketball possessions and football plays are not equivalent units.

Positive EPA, negative EPA and defense

When EPA is expressed from the offense's perspective, positive values mean the play increased its expected scoring value. Negative values mean it reduced that value. A value near zero means little change under that model.

For defense, pay attention to the label. A table showing offensive EPA allowed generally favors a defense that allows less. A table that reverses the sign to describe defensive value can point in the opposite direction.

Do not assume that a larger displayed number is better without checking whose perspective it uses. The nflfastR dictionary defines its EP and play EPA with respect to the possession team; a publisher can then aggregate or relabel those values for its own presentation. [3]

This is also why a negative number should not automatically be called an error. It may correctly describe a play that helped the defense rather than the offense.

EPA per dropback is not necessarily EPA per pass attempt

EPA is a play-value measure. Analysts can average it across different groups of plays, and that choice changes the meaning of the result.

EPA per dropback is not necessarily the same as EPA per pass attempt. In nflfastR, the quarterback-dropback indicator includes a pass attempt, a sack or a scramble. That is a documented definition for that dataset, not a promise that every provider uses identical rules. [3]

A sack can therefore matter in a dropback analysis even though it is not a completed or incomplete pass attempt. Scrambles can also make the dropback population broader than passing attempts alone.

Check the treatment of penalties, spikes, missing values and other excluded plays when a source documents those choices. Two tables can both say EPA per dropback while differing in their filters.

A comparison is most useful when the denominator, included plays and model are aligned.

What is a good EPA in football?

There is no single EPA threshold that should be treated as a permanent definition of good football performance.

Start with the metric's perspective and denominator. Is the figure offensive EPA per play, defensive EPA allowed, or quarterback EPA per dropback? Then check the league, season, model, opponents and sample.

Positive offensive EPA means value was added relative to the starting situations under that model. It does not automatically establish that a team ranks above the league average. A ranking requires a properly matched comparison group.

An average across a full season also means something different from an average across a few plays. A small sample can be dominated by one turnover or explosive gain.

For that reason, a responsible interpretation explains the comparison rather than presenting an undated universal cutoff. This guide does not use current player rankings or claim that a particular number guarantees future success.

Sample size and prediction

EPA helps summarize past play value in context. Using it to forecast a future game is a separate analytical task.

A team can produce a strong EPA average over a short stretch without establishing a stable performance level. Opponents, game situations and the mix of plays can affect what the sample represents.

When combining samples, calculate the rate using the underlying totals and play counts. Simply averaging two EPA-per-play figures gives each sample equal weight, even when one contains many more plays than the other.

A predictive model must be checked on data it did not use for fitting, with the information available at the time of the forecast. EPA is a possible input to analysis, not a substitute for that evaluation.

For the broader distinction between measuring performance and estimating future outcomes, see sports analytics explained.

EPA versus passer rating, QBR and completion percentage

These measures answer different questions. Treating them as interchangeable removes important distinctions.

Traditional NFL passer rating

NFL passer rating combines four passing categories: completion percentage, passing yards per attempt, touchdown rate and interception rate. The Pro Football Hall of Fame describes a ceiling of 2.375 for each category component and the familiar maximum rating displayed as 158.3. [4]

It is a passing measure. Rushing production and sacks are not categories in that traditional formula. A broader evaluation of a quarterback needs additional information.

ESPN Total QBR

ESPN Total QBR is distinct from traditional NFL passer rating. In a 2016 methodology explanation, ESPN described a contextual measure of quarterback contribution that uses EPA and allocates credit, rather than simply applying the public passer-rating formula to passing totals. This is a historical high-level description, not a claim that the methodology is unchanged today. [6]

It is not the same thing as unadjusted EPA, and this guide does not attempt to reconstruct ESPN's proprietary calculations.

Completion percentage and CPOE

Completion percentage is completed passes divided by pass attempts, multiplied by 100. It says how often passes were completed, not how difficult they were.

CPOE means completion percentage over expected. It compares actual completion outcomes with a model's expectations. NFL Next Gen Stats uses tracking information to explain pass difficulty; nflfastR also supplies completion probability and CPOE fields, using its own model. Those models should not be assumed identical. [3][5]

CPOE and EPA are not the same measurement. A pass can exceed its expected completion probability while its resulting game situation still needs a separate scoring-value evaluation.

What EPA cannot tell you on its own

EPA adds context, but it does not remove every limitation of football statistics.

It does not isolate one player's ability. A completed pass can reflect protection, route running, the throw, the catch and yards after the catch. Play EPA cannot cleanly assign those contributions just by reporting the total change in expected value. [2]

It depends on a model. Different inputs, training samples, era adjustments and versions can change the estimates. nflfastR's documentation describes model development and calibration choices rather than a single universal scoring-value formula. [1]

It depends on the data and filters. Missing or differently classified plays, penalty handling and inclusion choices can change an aggregate.

It is not win probability. Expected scoring value and the chance of winning are related ideas, but they are different model outputs. The nflfastR documentation treats EP and win probability as separate models. [1]

It is not a guaranteed forecast. Describing valuable past plays does not prove a team will reproduce that performance in its next matchup.

The broader lesson is similar to expected goals in soccer: an expected measure and an actual result are different things. xG describes shot scoring probability; EPA describes changes in football scoring value. Their units and purposes are not interchangeable.

The useful way to read EPA

Read an EPA figure in this order: perspective, model, included plays, denominator and sample. Then compare it with other figures built on compatible definitions.

That approach makes EPA useful without asking it to answer every question about a team or quarterback. It can explain the value of a play in context while leaving room for film, personnel, opponents and other evidence.

Related reading: What is OPS in baseball? explains why a useful summary metric still needs context and complementary information. It is not a claim that OPS and EPA can be directly compared.

Be first to explore SportQuantix

Pre-register for a free account. It stays pending until launch, and we will email you when the platform goes live.

Pre-register free

Sources and related reading

1. nflfastR model documentation: EP, WP and CP models. Accessible methodology reviewed October 17, 2026. Historical model documentation, not a universal NFL specification.
2. nfelo: What are Expected Points Added in the NFL?. Accessible explanation reviewed October 17, 2026.
3. nflfastR variable dictionary and EP implementation. Substantive raw source reviewed October 17, 2026; current repository definitions can change.
4. Pro Football Hall of Fame: NFL's passer rating. Accessible historical explanation reviewed October 17, 2026. Historical player-record claims are not used.
5. NFL Next Gen Stats: Introduction to Completion Probability. Accessible primary explanation reviewed October 17, 2026.
6. ESPN: How is Total QBR calculated?. Historical 2016 primary-source explanation, “How has Total QBR changed for 2016? A Q&A,” accessed successfully on October 17, 2026. This is not a complete reproducible formula or a claim that the methodology is unchanged today.