NFL Big Data Bowl 2025 - Player Movement Prediction

This project was a submission to the NFL's Big Data Bowl 2025 - Prediction Competition. The goal was to predict player movements after the football is thrown, using tracking data from the 2023 NFL season.

📊 Project Overview

The project uses machine learning (XGBoost) to predict the next position (dx, dy) of players on the field during passing plays. The models are trained on pre-throw player tracking data and predict post-throw movement patterns.

🎯 Key Features

• Player Movement Prediction: XGBoost models trained to predict x and y displacement for all players
• Catch Probability Model: Logistic regression model to estimate catch probability for targeted receivers
• Play Visualization: Animated and static visualizations of player routes and predicted movements
• Feature Engineering: Comprehensive set of tracking, temporal, and contextual features

📁 Repository Structure

NFLBDB/
├── data/                          # Data directory
│   ├── pre_ball. csv              # Pre-throw tracking data
│   ├── post_ball.csv             # Post-throw tracking data
│   ├── supplementary_data.csv    # Game and play context
│   ├── catch_probability.csv     # Catch probability predictions
│   └── full_predictions.csv      # All player predictions
├── train_data/                    # Training data by week
│   ├── input_2023_w*. csv         # Weekly pre-throw data
│   └── output_2023_w*.csv        # Weekly post-throw data
├── models/                        # Trained models
│   ├── model_dx.json             # X-displacement model
│   └── model_dy.json             # Y-displacement model
├── animations/                    # Generated visualizations
├── training. ipynb                 # Model training notebook
├── getting_data.R                # Data loading and preparation
├── functions.R                    # Visualization helper functions
├── running_functions.R           # Script to generate visualizations
├── completion. R                  # Catch probability modeling
└── Slideshow#1.pptx              # Project presentation

🛠️ Technology Stack

Python (88. 8%)

• pandas - Data manipulation
• numpy - Numerical computing
• xgboost - Gradient boosting models
• scikit-learn - Train/test splitting
• matplotlib - Static visualizations

R (11.2%)

• dplyr - Data wrangling
• ggplot2 - Static plotting
• gganimate - Animated visualizations
• nflreadr - NFL roster data

🚀 Getting Started

Prerequisites

Python:

pip install pandas numpy xgboost scikit-learn matplotlib jupyter

R:

install.packages(c("dplyr", "ggplot2", "gganimate", "nflreadr"))

Running the Pipeline

1. Data Preparation (R):

source("getting_data.R")

This combines weekly tracking data and adds NFL roster information.

2. Model Training (Python):

jupyter notebook training.ipynb

Trains XGBoost models for player movement prediction.

3. Catch Probability (R):

source("completion.R")

Fits logistic regression model for catch probability.

4. Generate Visualizations (R):

source("running_functions.R")

Creates animated plays and freeze-frame comparisons.

📈 Model Architecture

Movement Prediction Models

• Algorithm: XGBoost (Gradient Boosting)
• Objective: Regression (squared error)
• Target Variables:
  • dx: X-coordinate displacement to next frame
  • dy: Y-coordinate displacement to next frame
• Features (32 total):
  • Position & motion: x, y, speed, acceleration, direction, orientation
  • Temporal: frame_id (normalized), game_clock
  • Physical: player height, weight, position
  • Contextual: ball landing location, distance to ball, down, yards to go
  • Strategic: offense formation, coverage type, route type

Catch Probability Model

• Algorithm: Logistic Regression
• Target: Binary pass result (Complete/Incomplete)
• Features: Tracking-based metrics for targeted receivers
  • Nearest defender distance and speed
  • Relative speed and direction
  • Ball landing location
  • Distance to sideline

📊 Model Performance

Movement Prediction (RMSE):

• dx model: 0.01523 (validation)
• dy model: 0.01608 (validation)

These low RMSE values indicate high accuracy in predicting player positions frame-to-frame.

🎨 Visualizations

The project includes two types of visualizations:

1. Animated Plays (animations/play_test.gif):

   • Shows player movement throughout the play
   • Color-coded by offense/defense
   • Displays jersey numbers and field markings

2. Freeze-Frame Analysis (animations/play_*. png):

   • Compares actual vs predicted next positions
   • Highlights targeted receiver routes
   • Shows first-down markers and end zones

Example usage:

animate_play(
  game_id_val = 2023091100,
  play_id_val = 3214,
  df = pre_ball,
  supp = supplementary,
  save_path = "animations/play_test.gif"
)

📝 Key Insights

• Movement patterns are highly predictable from tracking data
• Contextual features (formation, coverage) significantly improve predictions
• Targeted receivers show distinct movement patterns compared to other players
• Catch probability is influenced by defender proximity and ball trajectory

🏆 Competition Context

This project was submitted to the NFL Big Data Bowl 2025, an annual analytics competition where participants analyze NFL's Next Gen Stats data to generate actionable insights.

📧 Contact

Jake Blumengarten

• GitHub: @JakeBlumengarten

📄 License

This project uses publicly available NFL Big Data Bowl data. Please refer to the NFL's data usage policies for terms and conditions.

---

Last Updated: December 2025