AWS for M&E Blog
NFL Next Gen Stats and AWS decode defense coverage patterns
In football, the quickest stats to see are the ones on the scoreboard. Touchdowns, passing yards, completion percentages—they leap off the box score. However, for decades, one of the game’s most consequential elements has been all but invisible: defensive coverage.
Ask any coach and they’ll tell you about how great coverage changes the complexion of an entire game. The best cornerbacks can erase an elite receiver without ever making a tackle by forcing the quarterback to look elsewhere. A disciplined safety rotation can bait a pass into the teeth of a defense. Still, there’s a problem—until now, coverage has been difficult to quantify.
“On a box score, coverage shows up as interceptions, passes defensed, or maybe a tackle,” says Mike Band, NFL Senior Manager of Next Gen Stats. “But those tell such a small part of the story. The most valuable coverage snaps—the ones where the quarterback doesn’t even throw your way—never appear in the stats.” The new Coverage Responsibility models of the NFL, built in partnership with Amazon Web Service (AWS), aim to change this.
For the first time, the league can identify which defender was targeted on a given pass, as well as their actual assignment on the play. They can also tell who the defenders were matched up against across the field—even when they weren’t targeted. This statistical framework, with new AI models, now transforms a subjective eye test into a data-driven measurement.
From guesswork to ground truth
Until now, the league’s targeted defender stat—who was in coverage on a completed pass—relied on identifying the nearest defender to the receiver when the ball arrived. That worked about three-quarters of the time, but there are plenty of situations where this framework failed to accurately capture the dynamics of coverage responsibilities.
“Sometimes the nearest defender at the catch point wasn’t responsible at all,” says Keegan Abdoo, Research Analytics Manager of the NFL Next Gen Stats team. “Maybe the corner passed off the route, or the safety was rotating down. We were missing the actual coverage responsibility and nuances of the play.”
The new AI models powering Coverage Responsibility replace those approximations with a probabilistic, frame-by-frame view of coverage. They use the player tracking data of the NFL—the X, Y coordinates, speed, and acceleration of every player on the field, recorded 10 times every second. The new targeted defender stat calculates, for each frame of a play, the likelihood that a given defender was matched to a given receiver. This probability can shift mid-play as defenders read the distribution of routes of the other team.
“Think about a stacked receiver alignment,” Abdoo explains. “Before the snap, a defender might be matched to the guy in front of him. But if those two receivers cross, his matchup changes. Now, we can see that shift in real time—not just guess at it after the fact.”
Beyond the broadcast view
Part of the magic is getting past the TV camera. Broadcast angles rarely show the entire secondary, so most coverage happens off screen. The all-22 coaches’ film does show it, but only to those with access and the time to break it down. “This is about quantifying the game within the game,” Band says. “The rotations, the disguises, the assignments—all the things that happen before the ball is even thrown.”
Since the models capture zone assignments, it can reveal the structure of a defense in Madden-like detail. They can tell you which defender had the deep-half, the hook-curl, or the flat, and how those responsibilities changed their matchups mid-play. They can even highlight unconventional rotations, like a corner bailing to a deep-half while a safety drops into the flat.
Three AI models, one story
Under the hood, the Coverage Responsibility models are actually three tightly linked machine learning models:
- Targeted defender identification – Pinpoints the defender responsible for a pass attempt.
- Matchup identification – Determines which receiver each defender was covering at any given moment.
- Coverage assignment identification – Identifies the specific area or zone a defender was responsible for, especially in zone coverages.
Each model was trained separately, but built on shared infrastructure. They use similar feature sets, share intermediate outputs, and even swap learnings from one to another. The models incorporate team-level coverage classification—the defensive scheme—as an input, enabling richer context. “We experimented with a single, all-in-one model,” says Band, “but found that three specialized models talking to each other actually worked better. The outputs from one informed the next.”
This alignment of model objectives is a non-traditional concept for a Next Gen Stats project. The AWS team, which helps build new ML-powered stats with the NFL each offseason, leaned on a spatial-temporal deep learning architecture. The architecture accounts for both where players are and how they move over time. It enables frame-level predictions, instead of only one label each play, which is key for understanding how the coverage picture evolves from the snap to the moment the ball is thrown.
This was also one of the first NFL Next Gen Stats projects where the AWS data science team leaned on generative AI coding services to accelerate development. “It was about speeding up the parallel processes—letting them try more approaches in less time. That productivity gain meant we could meet all our primary goals and still have time to experiment,” says Abdoo.
Driving inspiration from the NFL Big Data Bowl
The project drew from innovations born during the annual NFL Big Data Bowl, a competition where data scientists build models using anonymized tracking data. A 2021 winning entry inspired the first NFL team-level coverage classification model. More recently, a 2024 submission demonstrated the power of incorporating pre-snap tracking data into coverage predictions.
“The Big Data Bowl showed us that predicting coverage before the snap, and then comparing to post-snap coverage, was a great way to detect disguise,” says Band. “We productionized that idea at the team level, and now we’ve taken it down to the player level.”
One novel training technique from the AWS and NFL build: random truncation. During model training, instead of always feeding the full play from snap to pass, the team randomly cut plays at different points. They sometimes used three seconds before the snap, or sometimes mid-route, to help the model generalize to varied situations. The approach improved accuracy and may be used in future models.
What this unlocks for football
The immediate application of the Coverage Responsibility models are cleaner, more accurate stats. However, the longer-term effects are wide-ranging.
For broadcasters, it sharpens one of the sport’s favorite storylines: the star receiver compared with the star cornerback matchup. “It’s one of the purest one-on-one battles in football,” says Abdoo. “Now we can show exactly how often they actually matched up, how often the QB [quarterback] looked their way, and what happened when he did.”
The models also enable long-requested insights, such as double coverage rates. “Broadcasters have been asking for it for years,” Band notes. “Now we can define it precisely and, interestingly, show that it’s rarer than fans might think.”
For teams, it offers a scouting and self-scouting solution. With frame-level assignment probabilities, coaches can study tendencies in coverage disguise, identify blown assignments more accurately, and even spot potential tells in alignment or stance.
Seeing the whole picture
The first release of the Coverage Responsibility models will focus on the three core outputs—targeted defender, matchup, and assignment. However, the infrastructure is designed for expansion.
Potential derivative stats include:
- Coverage disguise rates at the player level.
- Versatility scores based on the variety of zones or routes defended.
- Space ownership metrics showing how defenders control areas of the field.
- Matchup exploitation rates, quantifying how often offensive coordinators scheme a receiver onto a less favorable defender.
Building toward the future
Football has always been a game of matchups. But until now, coverage matchups were defined mostly by what happened when the ball arrived. The rest—the vast majority of the snaps—remained in the dark. Now, with the Coverage Responsibility models, the hidden battles, the quiet victories that never made the highlight reel, and the tactical disguises that unfolded in the space between frames can all be measured, compared, and understood.
“For ten years, Next Gen Stats have brought fans deeper into the game—showing speed, separation, and pressure on live TV. The new Coverage Responsibility models take it a step further. For the first time, we will be able to show with confidence how often a defender and receiver face off, beyond just when the receiver was targeted,” says Abdoo. “It gives fans a whole new way to understand the battles happening downfield and off-screen in the traditional broadcast angle. This is just the latest step in the development of the NGS toolbox, and we can’t wait to see what the next decade of stats will unlock.”
Contact an AWS Representative to know how we can help accelerate your business.