NFL EPA Model

A custom Expected Points Added (EPA) model for NFL play-by-play analysis.

Overview

This project builds an EPA model from scratch using machine learning to predict expected points for any game situation, then calculates EPA as the difference between expected points after and before each play.

Installation

pip install -r requirements.txt

Usage

Train the Model

python main.py train --seasons 2020 2021 2022 2023

Calculate EPA for a Season

python main.py calculate --season 2023 --output epa_results.csv

Project Structure

epa/
├── requirements.txt
├── README.md
├── src/
│   ├── __init__.py
│   ├── data_loader.py      # Load and cache NFL play-by-play data
│   ├── feature_engineering.py  # Create model features from raw data
│   ├── model.py            # Train expected points model
│   └── epa_calculator.py   # Calculate EPA for plays
├── models/                 # Saved trained models
└── main.py                 # Entry point for training and analysis

How EPA Works

1. Expected Points (EP): For any game situation (down, distance, yard line), the model predicts the expected points that will be scored next (relative to the team with possession).

2. EPA Calculation: EPA = EP_after_play - EP_before_play

   • A positive EPA means the play improved the team's scoring expectation
   • A negative EPA means the play hurt the team's scoring expectation

Features Used

• yardline_100: Distance from opponent's end zone (0-100)
• down: Current down (1-4)
• ydstogo: Yards to first down
• half_seconds_remaining: Time remaining in half
• score_differential: Current score difference
• posteam_timeouts_remaining: Timeouts remaining

Future Extensions

• Player evaluation (QB EPA, receiver EPA, etc.)
• Team performance analysis
• Game predictions and win probability