What Is EPA in College Football? Expected Points Added Explained (Part IV)

EPA — expected points added — is the change in expected points across a single play. If a team’s situation was worth 0.66 expected points before the snap and 1.92 after it, the play was worth +1.26 EPA. It is the standard way to put a points value on one football play, and nearly every modern team and player metric is built by summing or averaging it.

$\mathrm{EPA} = \mathrm{EP}_{\text{after}} - \mathrm{EP}_{\text{before}}$

Part I covers what expected points is and Part II how the model estimates it. This article is about turning that into per-play value.

Why EPA instead of yards

Yards treat every situation as identical. EPA does not, and the difference is the entire point.

A 3-yard gain on 3rd and 2 continues a drive. A 3-yard gain on 3rd and 8 ends one. Both are “3 yards”. Their EPA differs by more than two points, because the model knows what happens next in each case.

Start from a routine situation — 1st and 10 at your own 25, tied, thirty minutes left in the half — worth +0.66 expected points. Here is what various outcomes of that single play are worth:

Play outcome	Resulting situation	EP after	EPA
Gain of 40	1st & 10, own 65	+3.57	+2.90
Gain of 25	1st & 10, midfield	+2.61	+1.95
Gain of 15	1st & 10, own 40	+1.92	+1.25
Gain of 8	2nd & 2, own 33	+1.62	+0.95
Gain of 10	1st & 10, own 35	+1.55	+0.89
Gain of 5	2nd & 5, own 30	+0.96	+0.30
Gain of 3	2nd & 7, own 28	+0.30	−0.36
Incompletion	2nd & 10, own 25	−0.02	−0.69
Sack for −7	2nd & 17, own 18	−1.13	−1.79
Interception at the spot	Opponent 1st & 10	−0.66	−1.33
Pick six risk: INT returned to your 10	Opponent 1st & 10, your 10	−4.96	−5.62

Look at rows four and five. A gain of 8 yards is worth more EPA than a gain of 10. Eight yards leaves 2nd and 2; ten yards leaves 1st and 10 five yards further downfield. The model prefers 2nd and 2, and it is right to — a 2nd and 2 converts at a far higher rate than a fresh set of downs needing ten.

That is not a quirk. It is the clearest possible demonstration that down and distance carry more value than raw field position in the middle of the field, and no yards-based statistic can express it.

Notice also that a 3-yard gain on 1st and 10 is negative. Gaining yards is not automatically good. Gaining fewer than the situation demands is a loss.

The third-down case

The same idea, sharper. 3rd and 3 at the opponent’s 35 is worth +2.75:

Outcome	Resulting situation	EP after	EPA
Convert, gain of 5	1st & 10, opponent 30	+3.98	+1.23
Stuffed for no gain	4th & 3, opponent 35	+0.88	−1.87
Turnover on downs	Opponent ball, their 35	−1.55	−4.30

A three-yard gain here is worth +1.23. A no-gain is worth −1.87. The same three yards that were worth −0.36 on 1st and 10 are worth +1.23 on 3rd and 3.

Sign conventions

This is where most confusion lives, so it is worth being explicit.

EPA is always signed from the perspective of the team with the ball at the snap. Positive EPA is good for that offense, negative is bad.

Consequences that catch people out:

Defensive EPA is the same number negated. A defense that allows −0.15 EPA/play is good. When building defensive metrics you either flip the sign or say “EPA allowed” and remember lower is better. cfbfastR reports the offense-signed value; the flip is yours to make.
On a turnover, expected points for the new offense is negated before differencing. That is why an interception at the spot costs −1.33 rather than −0.66: you lose your own +0.66 and hand the opponent an equivalent situation.
Return men differ by play type. On a kickoff the returning team is the possession team, so returner EPA is positive-good. On a punt return the punting team is the possession team at the snap, so the same raw column needs negating for the returner. This is the single most common error in player-level EPA aggregation.
Penalties are folded into the play they attach to, so a defensive holding on an incompletion produces one net EPA, not two.

EPA, EPA per play, and total EPA

Three different statistics that get called “EPA”:

Statistic	What it measures	Use it for
EPA (one play)	value of a single play	play-level analysis
Total EPA	sum over plays	volume — season value, workhorse backs
EPA per play	mean over plays	efficiency — comparing teams and players fairly

A quarterback with 1,200 total EPA over 500 dropbacks (2.4 per play) is not comparable to one with 1,200 over 900 dropbacks. Per-play is almost always the right comparison; total EPA is the right way to credit workload.

What is a good EPA per play?

For teams, from the 2025 FBS season as ranked on Game on Paper (adjusted offensive EPA per play, 135 teams):

Rank	Team	Adj. EPA/play
1	Vanderbilt	+0.35
4	Ohio State	+0.26
11	Miami	+0.22
21	Georgia	+0.15
51	Pittsburgh	+0.04
81	Kennesaw State	−0.01
135	Ball State	−0.31

So, as a rule of thumb for team offenses:

above +0.25 — elite, top-five
+0.10 to +0.25 — very good
around 0.00 — average; the median FBS offense sits near zero by construction
below −0.20 — among the worst in the country

The full range across 135 FBS teams is roughly −0.31 to +0.35. That is a narrower spread than newcomers expect: the gap between the best and worst offense in the country is about two thirds of a point per snap.

Note these are adjusted — corrected for opponent strength, home field and garbage time. Raw EPA per play runs wider and flatters teams with easy schedules, which is exactly why the adjustment exists.

Computing it yourself

cfbfastR ships EPA on every play, so normally you just read it:

library(cfbfastR)
library(dplyr)

pbp <- cfbfastR::load_cfb_pbp(2025)

pbp |>
  dplyr::filter(!is.na(EPA)) |>
  dplyr::select(pos_team, def_pos_team, down, distance, yards_to_goal,
                play_text, ep_before, ep_after, EPA) |>
  dplyr::slice_head(n = 10)

Team efficiency, offense and defense, from raw play data:

offense <- pbp |>
  dplyr::filter(!is.na(EPA)) |>
  dplyr::group_by(team = pos_team) |>
  dplyr::summarise(plays = dplyr::n(),
                   epa_per_play = mean(EPA),
                   total_epa = sum(EPA), .groups = "drop")

defense <- pbp |>
  dplyr::filter(!is.na(EPA)) |>
  dplyr::group_by(team = def_pos_team) |>
  dplyr::summarise(def_epa_per_play = mean(EPA), .groups = "drop")
  # lower is better -- these are offense-signed

offense |>
  dplyr::inner_join(defense, by = "team") |>
  dplyr::mutate(net = epa_per_play - def_epa_per_play) |>
  dplyr::arrange(dplyr::desc(net)) |>
  dplyr::filter(plays > 300)

Note the plays > 300 filter. Per-play averages over small samples are noisy; a receiver with four targets and +2.0 EPA per target has told you nothing.

And reproducing the worked table above, straight from the published model:

# Self-contained: Part I defines ep_at() inside its own eval=FALSE chunk, so a
# reader pasting this into a clean session would hit
# "could not find function 'ep_at'". Both setup steps are repeated here.
library(xgboost)
base <- paste0(
  "https://github.com/sportsdataverse/sportsdataverse-data/",
  "releases/download/cfb_model_artifacts/"
)
f <- tempfile(fileext = ".ubj")
download.file(paste0(base, "ep_model.ubj"), f, mode = "wb")
ep_model <- xgboost::xgb.load(f)

ep_at <- function(ytg, down = 1, dist = 10, secs = 1800, score_diff = 0) {
  X <- cbind(
    TimeSecsRem = secs, yards_to_goal = ytg, distance = pmin(dist, ytg),
    down_1 = as.integer(down == 1), down_2 = as.integer(down == 2),
    down_3 = as.integer(down == 3), down_4 = as.integer(down == 4),
    pos_score_diff_start = score_diff
  )
  pred <- predict(ep_model, X)
  # normalise: may be an n x 7 matrix or a flat row-major vector
  p <- if (is.matrix(pred) && ncol(pred) == 7L) pred else matrix(pred, ncol = 7, byrow = TRUE)
  as.vector(p %*% c(7, -7, 3, -3, 2, -2, 0))
}

before <- ep_at(75, down = 1, dist = 10)          # 1st & 10, own 25

data.frame(
  outcome = c("gain 8", "gain 10", "gain 15"),
  ep_after = c(ep_at(67, down = 2, dist = 2),
               ep_at(65, down = 1, dist = 10),
               ep_at(60, down = 1, dist = 10))
) |> transform(EPA = ep_after - before)

What EPA does not do

Stated plainly, because a metric is only useful if you know its edges:

It is not opponent-adjusted. Raw EPA credits a touchdown against a bad defence exactly like one against a good defence.
It has no team-quality input, so score margin partly proxies for it — see the caveat in Part II.
It does not assign credit within a play. A 40-yard completion produces one EPA figure; how much belongs to the quarterback, the receiver, the line or the coverage bust is a separate modelling problem.
It is descriptive, not predictive. High EPA per play describes what happened. Predicting next week needs stability adjustments, because explosiveness regresses harder than efficiency.
Garbage time distorts it unless you filter or weight for win probability.

Seeing it applied

Game on Paper computes all of this for every FBS game. Any game page shows EPA per play, success rate and a win probability chart, with a play-by-play listing ep_before, ep_after and EPA for every snap — the table at the top of this article, for a real game. Season leaderboards rank every team by offensive and defensive EPA per play, and players by passing, rushing and receiving EPA per play.

Frequently asked

Is a high EPA good? For an offense, yes — positive EPA means the play improved its scoring expectation. For a defense the sign flips: you want the offense’s EPA against you to be negative.

What is a good EPA per play? For an FBS offense, above +0.25 is elite and around 0.00 is average. The full 2025 range was roughly −0.31 to +0.35.

What is the difference between EPA and EPA per play? EPA is one play’s value; total EPA sums it (a volume measure); EPA per play averages it (an efficiency measure).

Can EPA be negative on a play that gained yards? Yes, and often. A 3-yard gain on 1st and 10 is worth −0.36 EPA, because it leaves a worse down and distance than the offense started with.

Does EPA account for the opponent? Not on its own. Adjusted versions correct for opponent, home field and garbage time on top of raw EPA.

Part V covers win probability — the same modelling idea applied to the outcome that actually matters, why vegas_wp exists alongside wp, and how WPA differs from EPA in late-game situations.

Data and artifacts

Models — cfb_model_artifacts
Season play-by-play — cfbfastR-data
Source — cfbfastR · sportsdataverse-py
Applied, live — Game on Paper
Raw API — CollegeFootballData.com, courtesy of @CFB_data

Citation

Gilani, S., Easwaran, A., Lee, J., and Hess, E. (2026). cfbfastR: Access College
Football Play by Play Data. R package version 3.0.0.9000.
https://cfbfastr.sportsdataverse.org

Authors, contributors and related SportsDataverse packages are listed on the package home page.

Saiem Gilani, Akshay Easwaran, Jared Lee

2026-09-03