AWS re:Invent 2025 - NFL Next Gen Stats: Decoding Defensive Coverage Using Transformer Architectures

AWS re:Invent 2025 - NFL Next Gen Stats: Decoding Defensive Coverage Using Transformer Architectures

Introduction to NextGen Stats

• NextGen Stats is the NFL's live tracking data operation, providing over 500 real-time stats per game
• Data is collected via chips in players' shoulder pads, tracking their locations 10 times per second
• This data is used by teams, broadcasters, and the league office for various purposes like player analysis, coaching, and player health and safety

Evolution of NextGen Stats

• Started with simple, logic-based stats in 2016
• Partnered with AWS in 2018 to build new machine learning-powered stats
• Shifted focus to classifying real football concepts to better communicate with broadcast talent

The Challenge: Decoding Defensive Coverage

• Prior to this project, NextGen Stats could only provide basic information like nearest defender to the receiver
• The goal was to develop models to:
  1. Classify each defender's assignment on a given play
  2. Identify the coverage matchups between defenders and receivers
  3. Determine the defender responsible for the targeted receiver

Technical Approach

• Utilized a "factorized attention" transformer architecture, which is more efficient than a standard full attention model
• Trained on 5 years of data (2020-2024) spanning tens of thousands of passing plays
• Key data sources:
  • Trajectory data (player locations, speed, acceleration, etc.)
  • Event data (snap, pass, catch timestamps)
  • Player information
  • Human-annotated coverage responsibility labels
• Performed data preprocessing and augmentation to normalize the data and improve model robustness

Coverage Assignment Prediction

• Model can classify each defender's coverage assignment (e.g., man, zone, blitz) using only pre-snap data
• Achieves high accuracy (>95%) by leveraging the temporal and spatial context in the data
• Able to identify subtle coverage disguises and adjustments made by the defense right at the snap

Defender-Receiver Matchups

• Model can accurately predict the coverage matchups between defenders and receivers, both pre-snap and at pass arrival
• Provides insights into the "stickiness" of coverage by individual defenders
• Identifies the most difficult coverage assignments, such as cornerbacks tasked with shadowing the opposing team's top receiver

Targeted Defender Identification

• Determines the defender responsible for the targeted receiver, going beyond just the nearest defender
• Enables more accurate attribution of coverage performance, accounting for factors like help coverage and route concepts

Business Impact

• 350% increase in year-over-year broadcast integrations of individual coverage performance metrics
• Enables more robust and insightful storytelling for NFL broadcasts
• Provides teams with deeper analytics to evaluate defensive performance and player talent

Conclusion

• The NextGen Stats team, in partnership with AWS, has developed a suite of transformer-based models that can accurately decode defensive coverage schemes in near real-time
• These models leverage the rich tracking data collected by the NFL to provide unprecedented insights into the tactical and strategic aspects of the game
• The impact is seen in enhanced broadcast experiences, more informed team decision-making, and a deeper understanding of the complexities of NFL defenses.