(:38) (Shotgun) L.Jackson right guard for 79 yards, TOUCHDOWN. Jackson sets record for 100-yd, rushing games by QB.
Everyone remembers who carried the ball. But a 79-yard touchdown run needs a lot more than one player. Someone opened the hole. Someone took a bad angle. Someone hustled downfield. Press play and watch each player’s credit build as the play unfolds.
Almost all of the ball carrier’s credit arrives in the first few seconds, when he hits the hole and breaks free. After that his bar barely moves: once he’s in the open field, an average player would probably score from there too. The decisive moment is the cut, not the sprint.
Individual Points Added. The expected points a player added to a play, compared with what an average player would have done in exactly his spot. Positive means he helped his team; negative means he hurt it. It’s measured in real points, the same units as EPA.
The rating. Per-snap iPA, compared with players at the same position and adjusted for sample size. 100 is average, and every 10 points is one standard deviation of talent. A 120 guard and a 120 receiver are equally far above their peers.
Every player gets an iPA on every snap: the quarterback, the receivers who ran decoy routes, the guard who pulled, the safety who took a bad angle. The iPAs on a play, plus what an average lineup would have produced, add up exactly to how the play turned out.
A neural network watches all 22 players and the ball, ten times a second, and estimates how many points the play is heading for. At the snap it knows little. By the end it knows a lot.
Whoever has the ball “owns” its movement: the carrier, or the passer while a throw is in the air. That’s how a QB gets credit for a throw and a runner for a cut.
A second network predicts what an average player would do next in each player’s exact situation. That prediction is his ghost, the bar he’s measured against.
Every 0.3 seconds, real players are swapped in among their ghosts in many random orders. Each player’s iPA is how much he changed the outcome compared with his ghost.
So what exactly is a ghost?
The ghost model learned from
about 380,000 real moments. For each one, it saw the whole field and
then where every player actually was 0.3 seconds later. Crucially, it
never knows who anyone is. It only sees “a cornerback here, trailing
a receiver moving this fast, with the ball over there.” Thousands of
different cornerbacks have been in roughly that spot, so the model
learns what they do on average. That’s the ghost: the
league-average player, dropped into this exact moment.
torch. Each of the
22 players and the ball is a token with 11 features (position, velocity,
acceleration, orientation, distance to the ball, possession). A separate
token carries down, distance, field position, time since the snap and
game state. It’s trained to predict both final EPA and yards
gained.iPA says Ben Powers added +1.62 points on this run. What does that mean in plain terms? To find out, we replayed the play with one change: Powers was swapped out for an average guard, who reacts to the play the way a typical guard would. Everyone else did exactly what they really did. At every moment, we asked the model one question: how many yards is this run heading for?
What actually happened was a 79-yard touchdown, with Lamar turning a good situation into a great one. These numbers are the model’s expectation at that moment, the average of all the ways the play could have gone from there, not a guarantee of how it would have ended.
How to read it: the two lines match at the start, split apart right when Powers makes his block (the yellow gap), then come back together once Lamar is past the line. They rejoin because Lamar’s real run is locked in: the replay can’t make him run differently, only measure how much better the play looked at that moment because of the block. Replaying the rest of the play with everyone reacting to the change is a natural next step.
Coming next: swap in a specific player. Today the stand-in is an average guard. Teaching the ghost model individual players’ tendencies would let us drop in anyone: what if a different guard had been there? What if that route had been run by Julio Jones instead of DeVonta Smith?
(14:09) (Shotgun) J.Allen pass deep middle to G.Davis for 98 yards, TOUCHDOWN [M.Fitzpatrick].
The quarterback’s iPA climbs while the ball is in the air. The throw is his. The receiver’s jumps at the catch and keeps growing through the run after it, and the defenders he beat sink below zero.
(10:43) (No Huddle, Shotgun) J.Herbert pass short right intended for G.Everett INTERCEPTED by Ja.Watson at KC 1. Ja.Watson for 99 yards, TOUCHDOWN.
Turnovers get shared out too. The defender who jumped the route earns a big positive iPA, and the quarterback whose ball it was takes his share of the blame.
2022, weeks 1–9, minimum 100 snaps. Top five at every position by iPA+ (100 = average).
Search or sort every player rated in 2022 (minimum 100 snaps).
Because iPA+ is standardised (100 is always the position average), a player can be compared with himself from one season to the next. The Progression Index is simple:
Newest season’s iPA+ ÷ previous season’s iPA+. The newest season always goes on top. Above 1 means he’s improving relative to his position; below 1 means he’s declining. 1.10 means he’s 10% better against his peers than the year before.
The 2021 tracking stops when the ball is thrown, so 2022 is rated on pass plays only for this comparison, and only for positions whose main job happens before the throw: QB, OL, DL, LB and DB. Running backs and pass catchers earn too much after the catch to compare yet.
With only two seasons, this is a first look rather than a career arc. Each extra season adds another link to the chain, and a run of Progression Indexes above 1 is what a player on the rise would look like.
Two seasons are just the start. Because iPA is calculated on every single snap, the same numbers can be added up by game, by season or across an entire career, and every view is on the same standardised scale. That makes it possible to tell a player’s whole story in one chart: when he got better, when he got hurt, when he peaked and when he started to slow down.
Chain the Progression Index year after year. Three straight seasons above 1 is a player on the rise, and multiplying the links gives his total growth: 1.10 × 1.08 × 1.05 means he’s about 25% better against his peers than when he started.
Missed games show up as gaps, since a player who isn’t on the field has no snaps. Comparing his iPA+ before an injury with his first games back shows whether he returned at full strength or needed time to recover.
His best season by iPA+. Shrinkage matters here: a great half-season after an injury gets pulled toward average, so a short hot streak doesn’t get mistaken for a peak.
Two or more seasons in a row with a Progression Index below 1 marks the start of a decline. Averaged across many players, the same idea gives each position its own aging curve: when guards, receivers or cornerbacks typically peak and fade.
The finest version of this already works on today’s data. Here’s Ben Powers through the first nine weeks of 2022: iPA per 100 snaps in each game, against the average offensive lineman.
With more seasons of tracking data, the same idea stretches across a whole career. The chart below is an illustration with made-up numbers, not a real player. It shows the story iPA+ could tell for someone over ten years.
Read left to right, it’s a whole career at a glance: steady growth to a peak in year 4, a lost season to injury, a short comeback year that’s shrunk toward average because of the small sample, a strong return, and then three straight seasons below 1 as he starts to decline. With every NFL season tracked this way, the same chart could be drawn for any player in the league.
A stat only matters if it says something about the player rather than luck. The test: rate every player separately on odd weeks and even weeks. If iPA captures real skill, the two halves should agree.
Zero would mean pure noise and one would mean perfectly repeatable. Each half is only four or five weeks of games, so nobody should expect values near one.
Where iPA shines: the offensive line. Linemen are the players traditional stats can’t see at all, and they’re where iPA is most repeatable, at 0.82. A lineman who rates well one half of the season tends to rate well the other half too. That’s exactly the “missing metric” this project set out to build.
Tackles, guards and centres do different jobs, though, so part of that consistency could come from role rather than skill. The real test is whether iPA separates players within each position:
| Position | Players | Split-half correlation |
|---|---|---|
| Centres | 39 | 0.62 |
| Guards | 79 | 0.57 |
| Tackles | 85 | 0.34 |
If these stay well above zero, iPA is telling linemen apart, not just positions.
Where it’s still learning: quarterbacks and receivers. QB (0.05) and WR (-0.03) barely repeat yet. There are only a few dozen starting quarterbacks, each with four or five games per half, and receiver iPA hinges on a handful of big catches, so a few weeks is too little to separate skill from luck. Both also depend heavily on teammates. More seasons of tracking data should help these positions most.
Swap the ghost for a real player. Today every player is measured against the average. The same machinery can measure him against anyone: train the ghost on one player’s tendencies and drop him into another player’s snaps. How many more points would that route have produced with Julio Jones running it instead of DeVonta Smith?
Everything here is built from public data. Player tracking records the position, speed, acceleration, direction and orientation of all 22 players and the ball, ten times a second.
The main source for 2022. Full-play tracking from the moment players line up, plus play-level and player-level detail. It supplies the large majority of the 16,096 plays from 2022 (6,783 runs and 9,313 passes).
Covers the same 2022 games with a focus on tackling. Used to fill in a small number of plays missing from the 2025 release, after removing duplicates.
Built around pass rush and protection. Tracking runs from the snap until the ball is thrown (or a sack or scramble), which adds 8,045 plays of protection, pressure and quarterback decision-making, but nothing after the throw. That’s why 2021 and 2022 are always rated separately.
Expected points added (EPA) for every play, which is the value iPA splits up, plus yards gained, down and distance, field position, game clock and score. Joined to the tracking data play by play.
Player names and positions come from the player files included with each Big Data Bowl release. Plays are standardised so the offence always moves in the same direction, and kneels, spikes and special teams are excluded.
Everything runs in R on a home computer. Run times below are from an AMD Ryzen 7 7700X with an NVIDIA RTX 3060 Ti (8 GB). An NVIDIA graphics card is strongly recommended: without one, the training steps and the credit run would take days rather than hours.
⬇ Download all scripts (.zip)
Create a free Kaggle account, open each competition page (NFL Big
Data Bowl 2023, 2024 and 2025), accept the competition
rules, and download the data. Unzip each into its own folder inside one
project folder: bdb2023/, bdb2024/ and
bdb2025/. Or, with the Kaggle command-line tool:
kaggle competitions download -c nfl-big-data-bowl-2025 (and
likewise for 2023 and 2024). The nflverse play-by-play data downloads
automatically in step 1.
⚠ Several gigabytes per competition. Leave plenty of disk space.
Install the packages: install.packages(c(“data.table”, “arrow”,
“nflreadr”, “ggplot2”, “torch”, “plotly”, “DT”, “zip”,
“rmarkdown”)). Then switch torch to its GPU version (this needs
an up-to-date NVIDIA driver): Sys.setenv(CUDA = “12.8”);
library(torch); install_torch(reinstall = TRUE, timeout = 600).
Restart R and check that cuda_is_available() returns
TRUE.
⚠ About a 3 GB download. If you’re on a very new version of R and the GPU install fails, the previous R release has ready-made GPU builds.
01_load_and_clean.R
Combines the three Big Data Bowls, removes duplicate plays, joins EPA from nflverse, flips every play so the offence moves the same way, and saves the tracking by week.
⚠ Reads tens of millions of tracking rows. 16 GB of RAM or more recommended.
02_prepare_model_data.R
Turns every moment after the snap into model input (23 tokens × 11 features), works out who has the ball with the sticky possession rule, and adds the game situation.
03_train_model.R (uses
model_def.R)
Trains the value model, the transformer that predicts a play’s EPA and yards gained from any moment, and prints its accuracy on held-out games.
⚠ GPU required. Stops automatically once it stops improving.
06_train_ghost_model.R (uses
ghost_def.R)
Trains the ghost model, which predicts what an average player does over the next 0.3 seconds, and compares it with simple baselines.
⚠ GPU required.
08_credit_all.R (uses
credit_engine.R)
The main event: Shapley credit for every player in every 0.3-second stretch of every play, about 300 million model evaluations. Saves after every 250 plays, so it can be stopped and restarted without losing work.
⚠ The long one, best left running overnight. To stop it safely, wait for
a “Chunk … done” line first. Setting M = 16 roughly halves
the time, with slightly noisier single plays.
09_player_ratings.R
Turns per-play credit into the standardised ratings (iPA per 100 snaps, iPA above average and iPA+), runs the split-half check and prints leaderboards.
10_showcase_data.R and 11_counterfactual.R
Save the featured play breakdowns, the ghost accuracy check and the “what if” replay used in this report.
⚠ GPU required.
nfl_credit_report.Rmd
This page. Open it in RStudio and click Knit, then Publish to share it on RPubs.
Each script starts with a setwd() line pointing at the
project folder (and this report has a matching root.dir
line at the top). Change those to your own folder before running. In
total, expect about five to six hours, almost all of it
step 5.