This is the first post in the “ML in the Wild” series, where we look at specific machine learning applications in the real world. We’re starting with one football fans might recognize: the tackle probability statistics that appear during NFL broadcasts.
What AWS Built
The NFL partnered with Amazon Web Services (AWS) to build a model that predicts tackle probability in real time. The system uses RFID tags embedded in every player’s shoulder pads and in the ball itself. Receivers mounted in every stadium capture each player’s position, speed, and acceleration ten times per second.
From that data, the model evaluates all 11 defenders on every play and calculates the probability that each one will make the tackle. A single ten-second play can generate 20,000 data points. The model processes all of it in under a second.
The Algorithms Behind Tackle Probability
The underlying algorithms are XGBoost and Random Forest, both ensemble methods that combine many decision trees to produce a prediction. Each tree asks a series of questions about the data: How fast is the defender closing? How many blockers are between him and the ball carrier? What’s the angle of pursuit?
The key step is feature engineering, which is transforming raw numbers into meaningful inputs. “Player is at coordinates X, Y” becomes “player is 3 yards from the ball carrier with no blockers in between.” That transformation is what makes the model useful.
AWS built and trained the model using their SageMaker platform. They tested 15 different model architectures against over a million data points in four months.
75+ ML Models and Counting
The same tracking infrastructure powers more than 75 ML models for the NFL now. Completion probability estimates the likelihood of a catch based on receiver separation and quarterback pressure. A newer model called Coverage Responsibility uses transformer architectures (the same technology behind large language models like GPT) to identify defensive assignments in real time, making previously invisible defensive performance measurable for the first time.
The injury prevention side is significant too. The NFL’s Digital Athlete program runs millions of simulated game scenarios to identify injury risk. That analysis contributed directly to rule changes: the dynamic kickoff, the ban on hip-drop tackles, and helmet redesigns that led to the lowest injury rate the league has seen in decades.
Machine Learning in Consumer Fitness Apps
These aren’t techniques limited to professional sports. Apps like Tempo use computer vision to analyze your exercise form in real time. FitBod builds workout recommendations from your training history. Strava uses your performance data to suggest training adjustments. The scale is different, but the approach is the same: collect movement data, find patterns, make predictions.
The AWS team wrote a detailed breakdown of how they built the tackle probability model. You can read it here.
