NFL Player Tracking Data: How Next Gen Stats Works and How to Use It

The NFL's operations site describes the setup. Each venue has 20 to 30 ultra-wide band receivers. Players carry two to three radio-frequency identification (RFID) tags installed in their shoulder pads, and tags are also placed on officials, pylons, sticks, chains, and in the ball. The league says about 250 devices are in a venue for a given game and three operators are required at every game to confirm the systems are working.

The system was developed with Zebra Technologies and Wilson Sporting Goods and runs on Amazon Web Services, according to the same page. The league lists its uses as broadcast and in-stadium presentation, club game planning and performance analysis, and player health and safety work.

Raw readings. The NFL says the system captures each player's location, speed, distance travelled, and acceleration at a rate of ten times per second. Because the tags sit in shoulder pads, the data describes where a player's torso is, not where his hands or feet are.

Derived measures. Next Gen Stats turns positions into football metrics. The nflverse data dictionary for NGS tables lists, among others, average time to throw (snap to throw, excluding sacks), aggressiveness (share of attempts into tight coverage), average cushion (yards between receiver and defender at the snap), average separation (yards to the nearest defender at the catch or incompletion), expected and actual completion percentage, expected yards after catch, and expected rushing yards with rush yards over expected.

Model outputs, not observations. Several of those are predictions from NGS models. Completion percentage above expectation is only as good as the completion probability model behind it.

Full league-wide tracking feeds go to clubs and partners. The public has three practical routes.

Next Gen Stats leaderboards

nflverse tables

Big Data Bowl releases

Big Data Bowl tracking typically arrives as one row per player per frame, keyed by game, play, player, and frame, with coordinates, speed, acceleration, orientation, and direction, plus separate play and player tables. Formats change between editions, so read that year's data description first.

Normalise direction. Plays run both ways down the field. Flip coordinates so the offence always moves the same way, or every spatial feature is mirrored half the time.

Align to events. Frames carry event tags such as the snap or a pass release in many editions. Anchor features to those moments rather than frame numbers, because pre-snap durations differ.

Smooth before differencing. Speeds and accelerations computed from raw positions amplify measurement noise. Use the provided fields where they exist, or smooth positions first.

Mind the scale. A season of frames for 22 players is very large. Store it in a columnar format partitioned by game and filter to the frames you need before joining.

Receiver separation and route value. Distance to the nearest defender at the catch point, and how it evolves through a route.

Pass rush and pressure. Time from snap to the quarterback being within a set distance, by rusher.

Run blocking and ball carrier vision. Expected rushing yards compares a carry with what the blocking and defensive positions offered.

Coverage identification. Classifying man and zone from defender movement relative to receivers.

Projection inputs. For player prop models, tracking aggregates such as targets with separation or time to throw can add signal to projections, but test each feature against closing prices before trusting it.

Big Data Bowl data is released under competition terms; read them before using it outside the competition. Releases are narrow, covering selected weeks or play types, so models trained on them may not generalise. Ten readings a second is coarse for very fast events such as a catch. Shoulder-pad tags do not record limb position or body orientation in the way optical skeletal tracking can. And NGS model definitions change, so record the definition date alongside every stored metric.

This page describes data sources and method and is not betting advice.

Begin with the aggregated tables: pull a season of NGS receiving data through nflverse, join it to play-by-play, and see how separation relates to target share. Then move to a Big Data Bowl release, normalise direction, and rebuild one published metric from frames. Rebuilding something known is the fastest way to find out whether a pipeline handles coordinates and events correctly.