How American football analytics works

American football was slower to adopt analytics than baseball or basketball, for a structural reason: every play involves 22 players in coordinated, position-specific roles, so a single box-score number like "yards" conflates the quarterback's decision, the offensive line's blocking, the receiver's route, and the defense's coverage into one figure that credits or blames the wrong people. Modern football analytics exists largely to pull those roles apart and to answer the in-game decisions — go for it or punt, take the points or the shot at the end zone — that traditional coaching judgment used to settle by feel.

The questions analytics is built to answer

Coaching staffs use it to answer concrete, repeatable decisions: is this fourth down worth going for, given the score, time and field position? Is a quarterback's completion rate reflecting real accuracy or an easy schedule of short, low-risk throws? And which players are actually driving winning football once the play calling and supporting cast are accounted for? Front offices ask a related but longer-horizon question: which statistical profiles predict future performance well enough to bet a contract or a draft pick on.

The data it runs on

Two data layers dominate football analytics, and they've historically been more separated by role than in soccer or basketball.

• Play-by-play data — down, distance, field position, play type, yards gained, and outcome — has existed for decades and underlies most publicly available football analytics, including the two metrics below.
• Player tracking data, sourced from chips in shoulder pads and stadium sensor systems in professional football, adds precise player location, speed and separation on every play. It underlies newer metrics around route running, coverage and pass-rush pressure that play-by-play data alone cannot see, because play-by-play only records the outcome of a play, not what 22 players were doing to produce it.

Advanced route-level and coverage data of this kind remains concentrated at the professional level and among a small number of specialist providers; much college and lower-level analysis still works primarily from play-by-play data.

Measuring the value of a play

[[expected-points-added|Expected points added]] (EPA) is the foundational metric in modern football analytics. It starts from a model of how many points a team is expected to score next, given down, distance and field position, then measures how much a specific play changed that expectation. A short completion on third-and-two that converts the first down adds real expected points even though the yardage gained is modest; a long completion that still leaves a difficult third down adds less than the yardage alone would suggest. EPA lets analysts compare plays, players and play-callers on a single, points-denominated scale instead of yards, which don't translate cleanly into scoring value.

[[success-rate-football|Success rate]] complements EPA with a simpler, more stable idea: a play "succeeds" if it gains enough yardage relative to down and distance to keep the offense on schedule (commonly defined as roughly 50% of yards to go on first down, 70% on second, and 100% on third or fourth). It is less sensitive to a single explosive play skewing the average than raw yards-per-play, and is often reported alongside EPA rather than instead of it.

Measuring a quarterback beyond completion percentage

Raw completion percentage rewards short, low-risk throws and can make a quarterback who avoids difficult passes look more accurate than one who takes on tougher, higher-value attempts. [[completion-percentage-over-expected|Completion percentage over expected]] (CPOE) corrects for this by modeling how difficult each individual throw was — based on factors like distance, receiver separation and throw location — and measuring how much a quarterback outperformed that difficulty-adjusted expectation. A quarterback with a modest raw completion percentage but a strong CPOE is generally completing harder throws than their peers, which the raw number alone would hide.

Measuring in-game decisions

[[win-probability-model|Win probability models]] estimate a team's chance of winning at any point in a game, given score, time remaining, field position and down and distance, using historical outcomes from similar situations. These models are the basis for most public fourth-down and two-point-conversion analysis: a decision is judged not by whether it worked on a given Sunday, but by whether it maximized win probability across the full range of likely outcomes. A coach who goes for it on a analytically favorable fourth down and fails has usually still made the correct decision, in the same way a good poker player can make a correct call and still lose the hand.

How the work is done in practice

Analytics staffs feed win-probability and EPA models into weekly game-planning, informing fourth-down and two-point decision charts coaches consult in real time, and into longer-horizon player evaluation for the draft and free agency. Because football has so few games per season relative to baseball or basketball, analysts lean more heavily on multi-year samples and situational context to avoid overreacting to small-sample outcomes — a single game's EPA total is a much noisier signal here than in a sport with 82 or 162 games.

Common mistakes and misreadings

• Judging a fourth-down decision by its outcome rather than the win-probability math behind it. A correct decision by expected value can still fail on a given play; that's normal variance, not evidence the model was wrong.
• Reading raw completion percentage as accuracy without accounting for throw difficulty, which CPOE is specifically built to correct for.
• Using a small number of games to judge a quarterback or scheme change, when EPA and success rate both need a meaningful sample — several games at minimum — to separate signal from noise.
• Treating yards as a value-neutral currency. A checkdown that picks up eight yards on third-and-fifteen adds far less expected value than eight yards on third-and-six, even though the box score records them identically.
• Ignoring supporting cast when evaluating an individual player from team-level metrics like EPA per play, which reflect the offensive line, play design and receivers as much as the individual being judged.

For how front offices combine these numbers with scouting, salary-cap constraints and injury data when building a roster, see how front offices use analytics. Browse every tool in this category for performance-data vendors.