Football analytics, explained

The points a team can expect to score from where it has the ball, averaged over everything that usually happens next. First and 10 at the opponent's 20 is worth far more than first and 10 at its own 20.

Before every snap, a model fitted to past seasons estimates it from the down, distance, field position and the time left in the half.

Exactly. nflverse computes it with the nflfastR model, which estimates the probability of each possible next score in the same half: a touchdown (7 points), field goal (3) or safety (2) for the team with the ball, the same three for its opponent (counted negative), or no score. EP adds up each outcome's points times its probability.

Its inputs, as Ben Baldwin documented the model in 2021, are the down, yards to go, yard line, seconds left in the half, both teams' timeouts, whether the team with the ball is at home, the stadium roof and the era of the game.

Ronald Yurko, Samuel Ventura and Maksim Horowitz's nflscrapR estimated those probabilities with a multinomial logistic regression (2018); nflfastR, by Sebastian Carl and Ben Baldwin, replaced it with a tree-based model (Baldwin).

On Downfield. No page prints EP itself. Every figure below is built from its change, EPA.

What one play was worth: the expected points after it minus the expected points before it, for the team with the ball. A 6-yard gain on 3rd and 5 keeps the drive going and adds value; the same 6 yards on 3rd and 10 usually costs some.

A play that moves the offense closer to scoring adds points, and one that stalls it costs them.

Exactly. EPA = EP at the end of the play − EP at the start, for the team that had the ball. If the ball changes hands, the new offense's expected points count against the old one. Downfield takes each play's EPA from the nflverse play-by-play (its epa field) and never computes its own.

Through Nov 1, 2026. Two 6-yard completions on third down: Sam Darnold to Cooper Kupp on 3rd and 5 at the 49ers' 8 (Oct 11) added +1.87, and Trevor Lawrence to Brenton Strange on 3rd and 10 at the Texans' 48 (Oct 18) cost 1.63. Every game page charts each snap's EPA the same way, as on Lions at Panthers.

The average EPA over many plays. Above zero, the plays added expected points on average; below zero, they cost points.

Exactly. The mean EPA over a set of plays. For a team's offense, every run and dropback (pass attempts, sacks and scrambles); for its defense, the same plays by its opponents, where lower is better.

Kneel-downs, spikes, two-point tries and plays wiped out by a penalty are left out, and a play with no EPA value is dropped rather than counted as zero. Team pages also split it into dropback EPA and rush EPA (designed runs).

Through Nov 1, 2026. The Buffalo Bills led NFL offenses at +0.179 EPA per play, peaking at +0.435 in a 41-31 win over the Lions on Sep 17. The Minnesota Vikings had the best defense, allowing −0.167, and held the Saints to −0.351 in a 23-13 win on Oct 11.

The plays a player's per-play figures count: a quarterback's dropbacks, a running back's designed runs, a receiver's targets.

Exactly. Quarterbacks: dropbacks, which are pass attempts, sacks and scrambles (a scramble counts here, not as a run). Running backs: designed runs, which leave out scrambles and kneel-downs. Receivers and tight ends: targets, every pass thrown to them, caught or not, interceptions included; a receiver is credited with the whole play's EPA.

A player's EPA per opportunity and success rate are counted over these plays; the leaders table heads the column EPA/Play. Box-score counts differ: official carries include scrambles and kneel-downs.

Through Nov 1, 2026. Josh Allen averaged +0.303 EPA over 237 dropbacks: 205 pass attempts, 13 sacks and 19 scrambles, among them a 21-yard scramble on 2nd and 17 at the Rams on Oct 12 (+2.73). Jahmyr Gibbs averaged −0.019 over 157 designed runs; Jaxon Smith-Njigba +0.751 over 78 targets.

How much more a team's offense gains per play than its defense gives up.

Exactly. Offense EPA per play minus defense EPA per play allowed.

Through Nov 1, 2026. The Jacksonville Jaguars led at +0.263 and went 5-2; the Philadelphia Eagles trailed the league at −0.230 and went 2-6.

EPA on first and second down only.

Exactly. The same average over first- and second-down plays. Opponent-adjusted and counted per game, it is a component of every offensive position's Downfield Score, from 5% to 12% of it.

Through Nov 1, 2026. Josh Allen averaged +0.241 opponent-adjusted EPA per early-down dropback. The biggest came in overtime at the Rams on Oct 12: a 57-yard completion to Khalil Shakir on 2nd and 16, +4.45 as played.

EPA corrected for the defenses a player faced, so production against a strong defense counts for more and against a weak one for less.

Exactly. Used inside the Downfield Score. For each game, Downfield takes the EPA per play the opposing defense allowed in its other games, on passes or runs to match the play, after first correcting that defense the same way for the offenses it faced.

It pulls the defense's difference from the league average toward zero as if the defense had played 250 more plays at the average, then subtracts that difference from each of the player's plays in the game. The score counts the result per game. The matchups page uses a simpler correction (below).

Through Nov 1, 2026. Jaxon Smith-Njigba's receiving EPA was +0.751 per target as played and +0.750 opponent-adjusted: the defenses Smith-Njigba faced allowed more than the league average in their other games, so the correction takes from the figure.

The share of plays that added expected points: did the play leave the offense better placed than it was before the snap?

Exactly. Plays with EPA above zero, divided by plays with an EPA value, over the same plays EPA per play counts. nflverse's play-by-play carries the same rule as its success flag.

An older, yardage-based rule, as Football Outsiders described it in 2007, counts a carry as a success when it gains 40% of the yards needed on first down, 60% on second, or all of them on third or fourth. Downfield does not use it.

Through Nov 1, 2026. The San Francisco 49ers led NFL offenses at 53.3%. In the running back score it is Positive-EPA run rate: Jahmyr Gibbs turned 49.0% of 157 designed runs into positive EPA. In the drive at the top of this page, 3 of 6 counted plays succeeded.

How much more often a quarterback completes passes than the average passer would on the same throws, in percentage points.

Exactly. For each pass attempt, nflfastR's completion model estimates the chance of a completion; as Baldwin documented it, its inputs are the air yards (and whether they are zero), the distance past the sticks, the down and yards to go, the yard line, whether the throw went to the middle, whether the passer was hit, home field, the roof and the era.

CPOE is the completion (1 or 0) minus that chance, averaged over the throws that have one and multiplied by 100 (nflverse). Sacks and scrambles are not throws, so they are left out.

Through Nov 1, 2026. Downfield prints CPOE inside the quarterback score, as Completion over expected (8%). Josh Allen's CPOE was +4.01: Allen completed that many percentage points more of the throws than the model expected.

Regressed toward the quarterback average by 326 modeled throws, it counts as +1.81, the 81st percentile. A quarterback needs 40 modeled throws for it to count.

How far a pass travels past the line of scrimmage before it is caught or falls incomplete. A screen thrown behind the line has negative air yards.

Exactly. nflverse records air yards as the distance, perpendicular to the line of scrimmage, to where the targeted receiver caught or did not catch the ball (air_yards).

The Downfield Score counts downfield air yards: the air yards of throws that reach the line of scrimmage or pass it (0 to 9 yards short, 10 to 19 intermediate, 20 or more deep), so a throw behind the line counts as zero rather than subtracting.

Per game, that is the regressed air yards per throw times the regressed dropbacks per game for a quarterback, or targets per game for a receiver.

Through Nov 1, 2026. Josh Allen: 9.78 downfield air yards per throw, 313.1 per game after regression, 10% of the quarterback score. The homepage board calls it Air yards / game.

How far downfield, on average, the passes thrown to a receiver travel: air yards per target.

Exactly. Air yards divided by targets for a receiver, or by pass attempts for a quarterback. In the Downfield app's player tables, aDOT is a receiver's box-score air yards divided by targets.

On Downfield. This site does not print aDOT; the app's player tables do.

The yards a receiver gains after making the catch.

Exactly. nflverse measures it from where the receiver made the catch to where the play ended (yards_after_catch). The tight end score counts Yards after catch per game: box-score YAC per target, regressed, times regressed targets per game. This site prints YAC only there; the app's player tables also print YAC per reception.

Through Nov 1, 2026. Brock Bowers: 2.95 yards after the catch per target, 28.0 per game after regression, the 97th percentile among 45 qualified tight ends (10% of the score).

How often a quarterback is sacked on a dropback.

Exactly. Sacks divided by dropbacks. In the quarterback score (7%) it is regressed toward the quarterback average by 310 dropbacks, and a lower rate ranks higher.

Through Nov 1, 2026. Josh Allen was sacked on 5.5% of 237 dropbacks, the 64th percentile. The costliest came on 3rd and 13 against the Chargers on Sep 27, a 6-yard sack with a lost fumble that cost 4.16 expected points.

First-down rate is the share of a receiver's targets that moved the chains. Explosive rate is the share of plays that went for a big gain.

Exactly. First-down rate: targets that produced a first down, divided by targets with a known result (8% to 10% of the receiver and tight end scores).

Explosive rate: targets that gained 20 yards or more, or designed runs that gained 10 or more, divided by those with a known gain (4% to 5% of a score). Both are regressed toward the position average before they are ranked.

Through Nov 1, 2026. 55.1% of Jaxon Smith-Njigba's targets moved the chains and 17.9% gained 20 yards or more, the longest an 82-yard touchdown from Drew Lock at the Cardinals on Sep 20. 17.2% of Jahmyr Gibbs's designed runs gained 10 or more.

Simple averages: rushing yards per carry (YPC), receiving yards per catch (YPR) and passing yards per attempt (YPA).

Exactly. The leaders, player and compare pages divide official box-score totals: YPC is rushing yards ÷ carries, and carries include scrambles and kneel-downs; YPR is receiving yards ÷ catches; the college page's YPA is passing yards ÷ attempts.

The running back score's Yards / run counts designed runs from the play-by-play instead, regressed toward the running back average by 258 runs.

Through Nov 1, 2026. Jahmyr Gibbs: 4.92 yards per carry; Gibbs's yards per designed run count as 4.46 after regression.

The NFL's official passing formula since 1973, built from completion rate, yards per attempt, touchdown rate and interception rate, on a scale from 0 to 158.3.

Exactly. Four parts, each held between 0 and 2.375: (completion % − 30) × 0.05; (yards per attempt − 3) × 0.25; touchdown % × 0.2; and 2.375 − interception % × 0.25. The rating is their sum ÷ 6 × 100 (Pro Football Hall of Fame). Downfield applies it to official box-score totals.

Through Nov 1, 2026. Josh Allen: 135 of 205 for 1,793 yards, 11 touchdowns and 5 interceptions rate 101.1.

The share of a team's targets that went to one player.

Exactly. A player's targets divided by the team's targets, both summed over the games in which the player has a box score: the counts are added first, not weekly percentages averaged.

It is the heaviest component of the receiver and tight end scores and counts in the running back score too (15% to 34% of a score, by position), regressed toward the position average by 3 games.

Through Nov 1, 2026. Jaxon Smith-Njigba drew 36.3% of the Seahawks' targets in 7 games (30.8% after regression), and 11 of their 24 in a 13-10 win over the Patriots on Sep 9.

The share of a team's carries that went to one player.

Exactly. A player's carries divided by the team's carries in the player's box-score games. Box-score carries include quarterback scrambles and kneel-downs, so the denominator is every rushing attempt the team logged. It is 25% of the running back score, regressed by 3 games.

Through Nov 1, 2026. Jahmyr Gibbs carried 77.7% of the Lions' rushing attempts in 7 games, and all 15 in a 32-26 loss at the Panthers on Oct 4.

The share of a team's offensive snaps a player was on the field for.

Exactly. On this site, the average of a player's weekly offensive or defensive snap percentages, as Pro Football Reference reports them through nflverse, over the weeks with a snap report. The score counts snaps per active game instead.

Through Nov 1, 2026. Jaxon Smith-Njigba was on the field for 84% of the Seahawks' offensive snaps in an average week.

How often a defender records sacks, quarterback hits, tackles, tackles for loss, passes defended and interceptions per active game, shrunk toward the sub-position average.

Exactly. Pass-rush disruption per game counts official sacks plus quarterback hits (sacks + QB hits). Coverage disruption per game counts passes defended plus defensive interceptions (PD + INT). Total impact events weight disruptive plays double: combined tackles + 2 × (sacks + QB hits + PD + INT).

Inside the defense score (DL, LB, DB), each rate is shrunk toward its sub-position Empirical Bayes prior (edge vs. interior line for DL, outside vs. inside linebacker for LB, cornerback vs. safety for DB) with multi-game volume shrinkage before percentile ranking alongside defensive snap deployment and opponent-adjusted unit defensive EPA.

Through Nov 1, 2026. Will Anderson Jr. led qualified defensive linemen (96.6 score) with 7.5 sacks and 15 QB hits; Alex Singleton led linebackers (93.8) with 98 tackles and 5 tackles for loss; Quentin Lake led defensive backs (94.1) with 11 passes defended and 3 interceptions.

How many more, or fewer, field goals a kicker made than an average NFL kicker would from the exact same kick distances.

Exactly. For each field-goal attempt between 18 and 80 yards, expected make probability is 1 ÷ (1 + exp(−(6.0983 − 0.1034 × distance))), fitted by maximum likelihood to 10,731 regular-season NFL field-goal attempts from 2016 to 2025.

Expected field goals made (xFGM) sums those make probabilities across a kicker's attempts; xFG% is xFGM ÷ attempts; and FGOE = field goals made − xFGM.

Through Nov 1, 2026. Chase McLaughlin made 19 of 19 field goals (100.0%) against 15.76 expected makes (82.9% xFG%), for +3.24 field goals over expected.

Net yards gained per punt after return yards and touchbacks, the share of punts downed inside the opponent 20-yard line, and net yards per punt above the league average (PNOE).

Exactly. Net punt average = official punt net yards ÷ punts, deducting return yardage and 20 yards per touchback. Inside-20 rate = punts downed inside the opponent's 20-yard line ÷ punts; touchback rate = punt touchbacks ÷ punts.

Punt net yards over expected per punt (PNOE) = a punter's net punt average minus the season-wide NFL net punt average across all punts.

Through Nov 1, 2026. Tommy Doman Jr. averaged 46.1 net yards across 14 punts (+4.3 PNOE per punt), pinning 6 inside the 20 (42.9%).

Downfield's 0 to 100 rating of a player's season so far against the qualified players at the same position, built from recorded play outcomes, box scores and snap counts. It is not a film grade and not a forecast.

Exactly. Downfield Score v7.0 has one model per position. Each takes 7 to 10 components (EPA, yards, disruption, usage and rates), regresses each toward the position average, converts each to a percentile among the qualified players at the position, multiplies each percentile by its weight (the weights add to 100%) and adds them, rounded to one decimal.

A player qualifies with 50 dropbacks (QB), 25 designed runs (RB), 15 targets (WR, TE) or 15 defensive plays (DL, LB, DB), with at least 80% of the plays behind every component known, and a position needs five qualified players. If any component cannot be computed the score is withheld rather than reweighted, and the leaders table and the player's page say why in its place.

Weights and regression constants are fixed, not refit each season. The site's menu calls the scores Grades.

Through Nov 1, 2026. Jaxon Smith-Njigba's 97.0, #1 of 99 qualified receivers, built component by component:

The points add up to 97.03, which rounds to the published 97.0.

Jalon Daniels has 123 dropbacks, past the 50 minimum, and no score: air yards are recorded only on throws, and the score needs them on 80% of a quarterback's dropbacks. 98 of Jalon Daniels's 123 were throws (79.7%); the other 25 were 9 sacks and 16 scrambles.

The measures each position's score combines, and how much each one counts.

Exactly. Downfield Score v7.0, read from each position's top-scored player, heaviest first. Wide receivers and tight ends share the receiving model with separate weights, and defensive linemen, linebackers and defensive backs share the defense model with separate weights.

Before anything is ranked, each component is pulled toward the position average: a lot for a player with few plays, a little for one with many, so a hot start on a small sample does not top the list by chance.

Exactly. For a rate: (the player's total + k × the position average) ÷ (the player's plays + k), as if every player started with k plays at the average. k is the sample at which the rate's signal and noise weigh the same, measured on odd against even weeks of the 2013 to 2025 regular seasons; the average pools every player at the position with plays this season.

Shares and per-game volumes move toward the qualified players' average the same way, by up to three games. A per-game component is the regressed rate per play times the regressed plays per game.

Through Nov 1, 2026. Jahmyr Gibbs turned 49.0% of 157 designed runs into positive EPA; with k = 342, that counts as 42.4%. Jaxon Smith-Njigba's 36.3% target share over 7 games, moved toward the receiver average by 3 games, counts as 30.8%.

Where a player stands on a component among the qualified players at the same position: the 90th percentile is ahead of about nine in ten of them.

Exactly. A blend: three quarters is the rank percentile (the share of qualified peers below the player, ties counting half), and one quarter is where the player's value falls on a normal curve fitted to the peers, capped at 3.5 standard deviations, which keeps some of the gap a runaway leader opens. For quarterback sack rate and team defensive EPA allowed, lower ranks higher.

Through Nov 1, 2026. Jaxon Smith-Njigba's regressed target share sits at the 99.5 percentile of 99 qualified receivers.

A player's place by score among the qualified players at the same position.

Exactly. 1 plus the number of qualified players with a higher score; equal scores share a rank. The leaders table prints it as #rank/qualified.

Through Nov 1, 2026. Jaxon Smith-Njigba: #1/99. No qualified receiver scored higher.

Player pages show each component twice: Raw, the player's own figure, and Shrunk, the figure after regression toward the average, which is the one that is ranked.

Exactly. For rates and shares the two figures are in the same unit. For per-game components they are not: Raw is per play (per dropback, designed run or target) and Shrunk is per game. When a position's average rate is below zero, a positive raw rate can regress to a negative per-game figure and still rank high.

Through Nov 1, 2026. Jahmyr Gibbs's Early-down opponent-adjusted rushing EPA per game: Raw +0.018 per play, Shrunk −0.844 per game, the 52nd percentile.

How closely a measure in one season lines up with the same measure the next season, for the same players: 1 is a perfect straight-line match, 0 is none.

Exactly. The methodology page correlates players' scores in two consecutive complete seasons, for players scored in both (Pearson's r), and sets that beside the same correlation for raw EPA per opportunity.

Rank stability is the same comparison of rank order (Spearman's), and retention is the share of the first season's scored players who were scored again. A high value means the measure holds steady from one season to the next; it does not make the score a forecast.

2024 to 2025. For 111 receivers, the score's year-to-year correlation was 0.746, against 0.254 for raw EPA per target; rank stability 0.750, retention 78.2%.

Points under a common fantasy football scoring system.

Exactly. The fantasy page scores 1 point per 25 passing yards, 4 per passing touchdown, −2 per interception, 1 per 10 rushing or receiving yards, 6 per rushing or receiving touchdown, and per catch 1 (PPR), 0.5 (half-PPR) or 0 (standard).

It does not count two-point conversions, lost fumbles or special-teams touchdowns. The app's Production tables count all three, so the two totals can differ.

Through Nov 1, 2026. Jaxon Smith-Njigba: 196.2 PPR points in 7 games, 42.5 of them in a 31-7 win at the Cardinals on Sep 20 (9 catches, 155 yards, 3 touchdowns).

The fantasy points a player's opportunities would usually be worth, whatever the player did with them.

Exactly. Model-based versions, such as nflverse's ffopportunity, estimate how many points the average player would score from each opportunity in its situation.

The fantasy page uses a flat rate instead: each pass attempt is worth the average points per attempt among the 578 skill-position players this site profiles, each carry the average per carry, and each target the average per target at the player's position, under the scoring chosen.

xFP = attempts × the attempt rate + carries × the carry rate + targets × the target rate. Depth, down and field position do not enter it.

Through Nov 1, 2026. The PPR rates were 0.425 per pass attempt, 0.633 per carry and 1.715 per wide receiver target. Jaxon Smith-Njigba's 78 targets, 3 carries and 1 pass attempt are worth 136.1 xFP, 19.4 per game.

Actual fantasy points minus expected fantasy points: how much more, or less, a player made of the opportunities than they were usually worth.

Exactly. Fantasy points − xFP, for the season and per game played.

Through Nov 1, 2026. Jaxon Smith-Njigba: 196.2 − 136.1 = +60.1 PPR points, +8.6 per game.

Fantasy points per dropback, designed run or target.

Exactly. Fantasy points divided by the player's opportunities (dropbacks, designed runs or targets). A running back's catches add points but not opportunities.

Through Nov 1, 2026. Jaxon Smith-Njigba: 196.2 ÷ 78 = 2.51 PPR points per target.

A per-game blend of a player's expected and actual fantasy points, nudged by the player's Downfield Score, and the season total it would reach over 17 games.

Exactly. Projected per game = (0.62 × xFP per game + 0.38 × actual points per game) × (1 + 0.0035 × (Downfield Score − 72)), with 72 standing in for a player without a score. The 17-game pace is points so far plus projected per game × (17 − games played). It is arithmetic on this season's figures; the site publishes no test of it against later games.

Through Nov 1, 2026. Jaxon Smith-Njigba: (0.62 × 19.4 + 0.38 × 28.0) × 1.087 = 24.7 PPR points per game, a pace of 443.

How often an offense turns a third down into a first down or a touchdown.

Exactly. Third-down runs and dropbacks that gained a first down or scored, divided by third-down runs and dropbacks with a known result, from the play-by-play's conversion flag. Kneel-downs, spikes and penalty-only snaps are not counted.

Through Nov 1, 2026. The Cincinnati Bengals led at 52.9%. Their most valuable conversion: on 3rd and 18 at the Titans' 30 against the Titans on Nov 1, Joe Burrow's 30-yard touchdown pass to Ja'Marr Chase (+4.70).

The wins a team's points scored and allowed would usually produce.

Exactly. Win share = points for^2.37 ÷ (points for^2.37 + points against^2.37); expected wins = win share × games played. The power rankings print expected wins next to the record, and the difference.

Through Nov 1, 2026. Jacksonville Jaguars: 196 points for and 111 against in 7 games give a 0.794 win share and 5.6 expected wins, against a 5-2 record (−0.6).

The average Downfield Score of a club's three highest-scored players.

Exactly. Among the 1,940 players this site profiles, the mean Downfield Score of the club's three highest-scored across all seven graded positions (quarterback, running back, receiver, tight end, defensive lineman, linebacker and defensive back). A club with none scored is set to 68.0.

Through Nov 1, 2026. Detroit Lions, the highest at 92.9: Jahmyr Gibbs 95.7, Aidan Hutchinson 92.8 and Amon-Ra St. Brown 90.3.

One number per club that blends its offense, its defense, its best players' scores and its points-based win expectation.

Exactly. 74 + 10.5 × (0.35 × offense + 0.25 × defense + 0.25 × starters + 0.15 × points).

Each input is first put on a common scale with fixed constants: offense = (EPA per play + 0.02) ÷ 0.14; defense = (−EPA per play allowed − 0.02) ÷ 0.12; starters = (starter grade − 74) ÷ 8.5; points = (Pythagorean win share − 0.5) ÷ 0.25. The result is held between 52 and 97.5.

Through Nov 1, 2026. Jacksonville Jaguars, the highest: offense +0.164, defense −0.099 allowed, starters 84.9, win share 0.794, for a rating of 85.8.

How strong the opponents a team has already played were.

Exactly. Over the games already played: the average of each opponent's season net EPA per play (an opponent met twice counts twice); the opponents' combined win percentage, with a tie as 0.5 of a win; and the average EPA per play the opponents' pass and run defenses allowed, where lower is tougher.

Each opponent's figures include its game against this team, and rank 1 is the toughest. SOS-adjusted EPA = the team's net EPA + 0.5 × its opponents' average net EPA.

Through Nov 1, 2026. The Denver Broncos had faced the toughest schedule, the Chiefs (two times), Jaguars, Rams, 49ers, Chargers, Seahawks and Cardinals: opponents at +0.094 net EPA per play, with a combined win rate of 55.2%. Their own net EPA of −0.006 adjusts to +0.041.

How tough the opposing defenses or offenses a player has faced were, and the player's or unit's EPA with a simple correction for them.

Exactly. On offense, Faced defense rank averages the opponents' pass-defense (quarterbacks, receivers, tight ends) or run-defense (running backs) EPA ranks over the games in the player's log, where 1 is the stingiest. Opp-Adj EPA = the player's EPA per opportunity − 0.5 × the average EPA per play those defenses allowed.

On defense, Faced offense rank averages the opponents' overall offensive EPA rank (defensive linemen, linebackers) or pass-offense EPA rank (defensive backs), where 1 is the highest-scoring offense. Opp-Adj def EPA = unadjusted unit defensive EPA per play in the defender's active games − 0.5 × the average offensive EPA per play those opponents produced. Both are separate from the score's opponent adjustment.

Through Nov 1, 2026. Jaxon Smith-Njigba's 7 opponents allowed +0.060 EPA per dropback on average (average rank 17), so Smith-Njigba's +0.751 becomes +0.751 − 0.5 × 0.060 = +0.721. On defense, Will Anderson Jr.'s 7 opponents averaged +0.050 offensive EPA per play (average rank 13), taking the Texans' +0.033 unit defensive EPA in Anderson Jr.'s games to +0.008.

The same idea as EPA, applied to the chance of winning: how much a play changed a team's win probability.

Exactly. nflverse publishes a win probability for the team with the ball before every play, and WPA as its change over the play (wp and wpa). Downfield does not import win probability, so it prints neither.

How much more often a team drops back to pass than a model expects from its situations.

Exactly. nflfastR estimates the chance that each play will be a dropback from its situation; PROE is a team's actual dropback rate minus that expectation, in percentage points. The Downfield app reports it for teams; this site does not.