EPA — short for expected points added — is the metric behind almost every modern NFL efficiency stat, and it rests on a simple idea: not all yards, downs, or field positions are worth the same amount. A three-yard gain that gets a first down is worth more than a three-yard gain that does not, and an NFL expected points model is the machinery that turns that intuition into a number.
Where Expected Points Added Comes From: Carter and Machol’s 1971 Model
The expected points idea did not start with a modern statistics department or a football analytics blog. It started with Virgil Carter, a backup quarterback for the Cincinnati Bengals who was also pursuing a graduate degree in operations research, and Robert Machol, a professor at Northwestern University. In 1971 the two published a three-page paper, “Technical Note: Operations Research on Football,” in the academic journal Operations Research.
Carter and his wife Judy hand-coded 8,373 plays from the first 56 games of the 1969 NFL season, tracking down, distance, field position, and the eventual outcome of each drive. From that dataset they estimated, for every yard line, the expected point value of having first-and-10 at that spot on the field — the first documented attempt to assign a point value to field position, not to treat every yard as interchangeable. The paper made two arguments that still define the field today: that different yard lines and down-and-distance situations carry meaningfully different expected point values, and that NFL teams of the era were too conservative on fourth down relative to what maximized those expected points.
The concept sat mostly dormant in public analysis for decades. Brian Burke, writing at Advanced Football Analytics (formerly Advanced NFL Stats) in the 2000s, rebuilt and popularized expected points using far more historical play-by-play data, formalizing expected points added (EPA) as the play-to-play change in expected points, not just a static table of field-position values. What Carter and Machol demonstrated on one half-season of hand-coded 1969 data, modern EPA models now compute automatically across every play tracked since.
From Hand-Tallied Plays to a Modern Expected Points Model
A modern EPA model asks a narrower, more answerable question than “how many points will this drive produce.” It asks: given the current situation, what is the probability of each possible next scoring event, whoever scores it?
The Inputs: Down, Distance, Yard Line, and More
The public model distributed with nflfastR — the open-source R package built on nflverse’s play-by-play data — is fit using down, yards to go, yard line, seconds remaining in the half, each team’s timeouts remaining, whether the possession team is playing at home, the stadium’s roof type, and an “era” grouping variable that accounts for rule changes across seasons. Down, distance, and yard line do most of the work; the remaining variables sharpen the model’s accuracy in edge cases, particularly end-of-half and end-of-game situations where time pressure changes what a given field position is worth.
The Seven Possible Outcomes
The model does not predict a point total directly. It predicts the probability of seven mutually exclusive outcomes for whichever team scores next in the half: a touchdown for the possession team, a field goal for the possession team, a safety for the possession team, no more scoring before the half ends, or the mirror image of each of those three outcomes credited to the opponent instead. This multinomial structure — modeling what happens next, not how many points result — is what nflfastR’s own calculate_expected_points() documentation reflects, returning a probability column for each of the seven outcomes (td_prob, fg_prob, safety_prob, no_score_prob, and their opp_ equivalents) alongside the derived ep value.
Turning Probabilities Into a Single Expected Points Number
Once the seven probabilities are estimated, expected points is a weighted sum, with point values signed from the possession team’s perspective:
where is +7 for a possession-team touchdown, +3 for a field goal, +2 for a safety, 0 for no more scoring, and −2, −3, −7 for the mirror outcomes credited to the opponent. cfbfastR’s documentation for the equivalent college-football model gives a clean worked illustration of the same math: a play with a 40% touchdown probability, 30% field goal probability, 20% no-score probability, and 10% opponent-touchdown probability is worth
expected points. That number — 2.9, not 7 or 0 or −7 — is what “expected” means here: a probability-weighted average across every way the situation could resolve, not a guess at the single most likely outcome.
Under the hood, the current generation of public models (nflfastR’s included) estimates those seven probabilities with a gradient-boosted decision tree model (XGBoost), not logistic regression — a change made specifically to better capture nonlinear interactions — end-of-half situations in particular, where a few seconds remaining can swing the probabilities sharply in a way a simpler linear model struggles to represent.
What Expected Points Added (EPA) Measures
Expected points on its own describes a situation. Expected points added (EPA) describes a play — specifically, how much a single play changed that situation’s value:
A Worked EPA Example
Suppose a hypothetical offense faces first-and-10 from its own 25-yard line, a situation worth roughly 0.4 expected points. A 12-yard completion moves the ball to the offense’s own 37 and resets the count to first-and-10 again, a situation now worth roughly 1.1 EP. The play’s EPA is 1.1 − 0.4 = +0.7 — a solidly positive result, reflecting real progress toward scoring, even though the play did not come close to the end zone.
Contrast that with a 2-yard run on third-and-5 from the same 25-yard line. Before the play, third-and-5 from the 25 might be worth roughly 0.3 expected points; after the play, facing fourth-and-3, that value drops to roughly −0.2 EP (a fourth down deep in one’s own territory usually means punting, which hands the opponent the ball and therefore has negative value for the original possession team). That 2-yard gain produces EPA of −0.2 − 0.3 = −0.5, despite technically moving the ball forward — because it did not accomplish the down’s purpose.
This is the core reason EPA replaced simple yardage tallies in serious NFL analytics: it prices a play by what it did to the expected scoring outcome, not by the yardage total alone.
EPA Per Play for NFL Offenses
Summed and averaged across a season, EPA becomes a team-efficiency stat. League-average EPA per play is 0.0 by construction, since the underlying model is fit on league-wide data and every play’s EPA is measured relative to that baseline. Offenses that sustain roughly +0.10 to +0.20 EPA per play across a full season are performing well above average; a season-long average near 0.0 is roughly league-average, and a negative full-season average marks a well-below-average offense.
Defensive EPA per play runs in the opposite direction of intuition at first glance: it is typically reported as the EPA a defense allows, so a strong defense wants a negative number — it is suppressing the opponent’s expected scoring value play after play. Some public leaderboards flip the sign so that a larger number always means “better” regardless of side of the ball; the convention is not universal across sources, so checking which convention a given table uses matters before comparing offensive and defensive EPA figures side by side.
EPA Per Dropback for NFL Quarterbacks
For individual quarterback evaluation, the standard unit is EPA per dropback — the average EPA across every pass attempt, sack, and scramble that originates from a passing play, which isolates a quarterback’s contribution from the team’s running game. Because passing is generally a more efficient way to gain EP than rushing on a per-play basis, the league-average dropback tends to run positive, not sit at exactly 0.0, and the gap between an average and an elite quarterback season shows up clearly in this per-dropback figure — MVP-caliber seasons post EPA-per-dropback figures well clear of the league mean.
EPA per dropback is not a complete quarterback evaluation by itself. It says nothing about how much of a given play’s outcome should be credited to the quarterback versus the receiver, offensive line, or play design, and it inherits every limitation of the underlying model — including, as the next section covers, its treatment of garbage time.
How EPA Handles Turnovers and Penalties
Turnovers are where the expected points model earns its keep, because a raw yardage stat has no natural way to express how costly a turnover is relative to field position. An interception thrown from a team’s own 20-yard line hands the opponent a first-and-10 near midfield-adjacent territory — not deep in their own end — so the offense’s EP swings from a small positive value to a meaningfully negative one in a single play — often a larger single-play swing than any non-scoring gain could produce. A fumble lost at the goal line on a would-be touchdown is even more extreme: the offense goes from an EP value near +6 (a near-certain score) to a negative value (the opponent now has the ball deep in what was the offense’s scoring territory), producing an EPA swing that can exceed 7 points on a single play — larger than the touchdown itself would have been worth.
Penalties are handled the same way the model handles any other down-distance-yard-line change: a defensive holding call that turns second-and-10 into second-and-5 registers as a positive-EPA event for the offense even though no player touched the ball, because the situation genuinely improved. A false start that turns first-and-10 into first-and-15 registers as a negative-EPA event for the same reason. This is a deliberate feature, not an artifact — the EP model does not distinguish “value gained by a peculiarity of the rulebook” from “value gained by a completed pass,” because both change the same underlying quantity: how many points the current situation is worth.
Limitations of the EPA Model
Garbage Time and Score Effects
A raw EPA calculation includes every play in the database regardless of score margin or time remaining, so a touchdown scored with a comfortable fourth-quarter lead against a defense that has stopped contesting the game still counts at full value — even though the game’s outcome is no longer meaningfully in question. There is no single standardized garbage-time cutoff across public EPA sources: some analysts filter plays after win probability crosses a threshold, others weight EPA by win probability directly (producing win-probability-added, a related but distinct metric), and others report raw EPA without adjustment and note the caveat separately. Comparing EPA figures across two different public sources without checking their garbage-time handling can produce misleading conclusions.
Model Instability at the Edges
Down-distance-yard-line combinations near the extremes of the field — fourth-and-goal from inside the 1-yard line, or first-and-10 pinned deep in a team’s own end zone — occur far less often in the training data than a routine first-and-10 near midfield. A 2024 paper by Brill, Yee, Deshpande, and Wyner (published on arXiv, September 2024) found that all combinations of just yardline, down, and seconds remaining produce more than 7 million possible game states, creating severe sparsity relative to the roughly half-million plays available in a typical multi-season training window — meaning some rare situations are estimated from very little direct data, even in a model trained on decades of play-by-play.
Era Changes
An NFL EP model trained on one era of rules can misprice a current situation if the rules governing scoring have shifted meaningfully since. That is why nflfastR’s public model explicitly groups seasons into eras — 1999–2001, 2002–2005, 2006–2013, 2014–2017, and 2018 onward, roughly tracking major rule changes like expansion to 32 teams and shifts in defensive contact enforcement. It does not pool every season from 1999 to the present as statistically interchangeable.
Not a Complete Player-Evaluation Metric
EPA figures are attributed to whichever player the play-by-play data credits with the play — the passer on a completion, the rusher on a run — but a single play’s outcome is rarely the product of one player alone. A perfectly thrown pass dropped by a receiver, or a well-blocked run stuffed by an unblocked defender, still produces one EPA value assigned to the primary ball-handler, with no accounting for which participants on the play actually drove the result. Efforts to decompose credit further — separating a quarterback’s decision-making from a receiver’s yards after catch, for instance — exist in the public analytics literature (nflWAR is one published example) but sit on top of the base EPA model — they do not replace it — and none of them fully resolve the attribution question.
Building an EPA Calculation From Play-by-Play Data
nflverse-data’s play-by-play release already includes pre-computed ep and epa columns for every play back to 1999, distributed under a CC-BY-4.0 license, so most practical work with EPA data starts by loading that release directly, not by re-fitting a model from scratch:
import pandas as pd
pbp = pd.read_csv(
"https://github.com/nflverse/nflverse-data/releases/download/pbp/play_by_play_2026.csv.gz",
compression="gzip", low_memory=False
)
# Offensive EPA per play, regular season, excluding no-play penalties
plays = pbp[pbp["epa"].notna() & (pbp["play_type"].isin(["pass", "run"]))]
offense_epa = (
plays.groupby("posteam")["epa"]
.mean()
.sort_values(ascending=False)
)
# Quarterback EPA per dropback
dropbacks = pbp[pbp["qb_dropback"] == 1]
qb_epa = (
dropbacks.groupby(["passer_player_name", "posteam"])["epa"]
.agg(["mean", "count"])
.query("count >= 100")
.sort_values("mean", ascending=False)
)
The groupby(...).mean() step is doing exactly the arithmetic described above — averaging the per-play change in EP across every play a team or player was involved in — because the heavy lifting (estimating the seven outcome probabilities and converting them to a point value) has already been done upstream in the ep/epa columns nflverse publishes with the release. Fitting the underlying model from scratch is rarely necessary for most EPA analysis; nflfastR’s own model weights are published alongside the package specifically so that downstream analysis can start from the ep/epa columns directly without repeating the estimation step.
How EPA Fits Alongside Other Rating Approaches
EPA analysis answers a different question than a season-long rating system. An Elo rating or a continuous Markov chain rating produces a single number meant to summarize a team’s overall strength, updated game by game or solved across a full schedule at once. EPA per play instead scores every individual play against a fixed, external baseline — the league-wide EP model — which makes it useful for diagnosing why a team is strong or weak (its EPA per play splits cleanly by down, by play type, by quarter) in a way a single overall rating number cannot.
A win probability model is the closest relative to an expected points model: it uses many of the same inputs (down, distance, yard line, score, time) but predicts the probability of winning the game, not the probability of scoring next, which is why EPA and win-probability-added can tell different stories about the same play, particularly late in a lopsided game. A team trailing by four scores with two minutes left can still put up a large-EPA touchdown drive that moves its win probability by almost nothing, because the expected points model has no concept of “the clock has run out of meaningful plays” the way a win probability model does.
None of this makes EPA analysis a replacement for a season-long rating — it answers a narrower, play-level question well, and that narrowness is exactly what makes it composable with other tools. A rating system can be built by aggregating EPA per play across a schedule, similar to how k-means clustering groups players by a set of per-play statistics, not season totals, or a possession-based ranking algorithm can use EPA per play as one of its inputs alongside pure win-loss record. The weekly ratings page applies EPA per play, among other inputs, to the current season’s play-by-play as it accumulates, and the companion piece on uncertainty in expected points estimates covers how much confidence a single EPA figure — especially over a small sample of plays — actually deserves.
Frequently asked questions
What does EPA mean in football?
EPA stands for expected points added. It is the change in a team's expected points value from before a play to after it: EPA = EP(after) − EP(before). A play that improves field position, down, or distance in a meaningful way has positive EPA; a play that worsens the situation has negative EPA, regardless of how many yards were gained.
What is expected points (EP) in football?
Expected points (EP) is the number of net points a team is expected to score before its opponent scores again, given the current down, distance, and yard line. It is a weighted average over every possible next scoring outcome — touchdown, field goal, safety, or the opponent scoring first — each multiplied by its point value and summed.
Who invented the expected points model in football?
Virgil Carter, a Cincinnati Bengals quarterback with a graduate background in operations research, and Robert Machol, a Northwestern University professor, published the first expected points model in 1971 in the journal Operations Research. They hand-coded 8,373 plays from the first 56 games of the 1969 NFL season to estimate a point value for first-and-10 at each yard line.
What is nflfastR's expected points model?
nflfastR's EP model is an open-source, gradient-boosted model (built with XGBoost) that predicts the probability of seven next-score outcomes — touchdown, opponent touchdown, field goal, opponent field goal, safety, opponent safety, or no more scoring in the half — from inputs including down, distance, yard line, seconds remaining, timeouts, and era. It has been fit on every play tracked since 1999.
How is expected points calculated from probabilities?
Expected points is a weighted sum: EP = Σ (probability of outcome i × point value of outcome i), where point values are signed from the offense's perspective — touchdown +7, field goal +3, safety +2, no score 0, and the negative equivalents for the opponent scoring first (−2, −3, −7). A play with a 40% touchdown probability, 30% field goal probability, 20% no-score probability, and 10% opponent-touchdown probability is worth 0.4(7) + 0.3(3) + 0.2(0) + 0.1(−7) = 2.9 expected points.
What counts as a good EPA per play for an NFL offense?
League-average EPA per play is 0.0 by construction, since the model is fit on league data. Offenses sustaining roughly +0.10 to +0.20 EPA per play across a season are performing well above average; anything consistently negative is a below-average offense. Defensive EPA runs in the opposite direction — a defense wants to hold opponents to negative EPA per play.
What is EPA per dropback for a quarterback?
EPA per dropback averages a quarterback's expected-points-added value across every pass attempt, sack, and scramble on a passing play, isolating quarterback play from the running game. Passing is generally more efficient than rushing on a per-play basis, so the league-average dropback tends to run positive, and elite quarterback seasons are associated with EPA per dropback well above that average.
Does EPA account for garbage time?
Not by default. A raw EPA per play calculation includes every play in a database regardless of score or time remaining, so a touchdown scored with a 30-point lead late in the fourth quarter still counts at full value even though the game's outcome is no longer in question. Analysts commonly filter or weight EPA by win probability, or exclude plays after a score gets lopsided, to reduce this distortion — there is no single standardized garbage-time cutoff across public EPA sources.
What data does an expected points model need as input?
At minimum: down, yards to go, and yard line. nflfastR's public model adds seconds remaining in the half, each team's timeouts remaining, whether the possession team is home, stadium roof type, and an 'era' variable that groups seasons together to account for rule changes — for example, the expansion to 32 teams, changes to defensive contact rules, and other league-wide shifts that move the baseline value of field position over time.
Why does an expected points model use a multinomial outcome instead of predicting points directly?
Predicting the next score type (touchdown, field goal, safety, opponent score, or no score) as separate probabilities, then converting to points afterward, lets the model be evaluated and calibrated on each outcome type separately and lets the point-value conversion be changed without re-fitting the model. A model trained to output a single point value directly would blend these outcome types together and be harder to diagnose when it misbehaves in a specific situation, like fourth-and-goal from the 1-yard line.
How is EPA different from yards gained?
Yards gained treats every yard as equal; EPA does not. Three yards on third-and-2 wins the down and keeps a drive alive, so it produces sharply positive EPA. Three yards on third-and-9 leaves the offense facing fourth down, so the same three yards produces a small or even negative EPA. EPA prices a play by what it did to the expected scoring outcome, not by the yardage total alone.
Is EPA the same as win probability added (WPA)?
No. EPA is scored against a team's expected points independent of the current score and time remaining, while win probability added (WPA) is scored against the probability of winning the game, which depends heavily on score margin and time left. A touchdown late in a blowout can carry a large EPA value but a tiny WPA value, since it barely changes who is likely to win.
How far back does public NFL expected points data go?
nflverse-data's play-by-play release, which nflfastR's public EP model is fit on, covers every NFL season from 1999 onward, distributed under a CC-BY-4.0 license via GitHub. Carter and Machol's original 1971 model, by contrast, was built from a single half-season of hand-coded 1969 data — 8,373 plays from 56 games — a dataset several orders of magnitude smaller than what a modern model trains on.
Can expected points be applied to fourth-down decisions?
Yes — this was one of the two central applications in Carter and Machol's original 1971 paper, which used expected point values by field position to argue that NFL teams of that era were too conservative on fourth down relative to what maximized expected points. Modern fourth-down decision tools built on public EP models (comparing the expected points of going for it, punting, or attempting a field goal) descend directly from that argument.
What is the difference between EPA for an offense and EPA for a defense?
Offensive EPA per play is calculated from the possession team's perspective — positive means the offense gained value. Defensive EPA per play is typically reported as the EPA the defense allowed, so a good defense has a negative number: it is preventing the opponent's expected points from growing. Some public leaderboards report defensive EPA with the sign flipped so that a higher number always means better performance; the convention varies by source.
Does an NFL expected points model need to be re-fit as rules change?
In practice, yes, which is why nflfastR's public model includes an explicit 'era' variable: it does not treat 1999 and the present as statistically identical. Rule changes that shift scoring rates — kickoff and extra-point rule changes, defensive holding and contact enforcement, and rosters expanding — change the underlying relationship between field position and expected points over time, so a model trained only on old eras can misprice current situations.
Sources
- Carter, V. & Machol, R.E. (1971) — Technical Note: Operations Research on Football, Operations Research 19(2): 541–544
- nflfastR — calculate_expected_points() reference documentation
- Open Source Football — nflfastR EP, WP, and CP models (Ben Baldwin, 2020)
- cfbfastR — College Football Expected Points Model Fundamentals, Part I
- Advanced Football Analytics — Expected Points (EP) and Expected Points Added (EPA) Explained (Brian Burke)
- ESPN — Virgil Carter's legacy of EPA and NFL analytics
- nflverse-data releases (play-by-play, CC-BY-4.0)