NFL Predictor

Predicts NFL game outcomes using machine learning. Uses logistic regression with team stats from 2015-2025, combining long-term averages and recent form to make predictions.

Setup

1. Clone this repo and navigate to the folder

2. Create a virtual environment:

   • Windows: python -m venv venv
   • Mac/Linux: python3 -m venv venv

3. Activate it:

   • Windows (PowerShell): .\venv\Scripts\Activate.ps1
   • Windows (CMD): venv\Scripts\activate.bat
   • Mac/Linux: source venv/bin/activate

   If you get an execution policy error on Windows, run:

   Set-ExecutionPolicy -ExecutionPolicy RemoteSigned -Scope Process

4. Install dependencies:

   pip install -r requirements.txt

Running

Just run:

python main.py

The program loads NFL data (will download on first run), trains a model on historical games, and can make predictions. Some features are commented out in the code - feel free to uncomment Block 8 if you want to see the visualization charts.

The model uses two types of features:

• Anchor: Long-term team averages
• Spark: Recent form (weighted last 5 games)

Team Names

You can use abbreviations or full names (e.g., "kc", "chiefs", or "Kansas City Chiefs"). See team_abbr_map in main.py for all options.

How It Works

The script is organized into blocks that run sequentially:

Block 2: Data Loading - Downloads team stats and schedules from 2015-2025 using nflreadpy

Block 3: Team Mapping - Converts team abbreviations to full names so everything is standardized

Block 4: Data Merging - Combines stats with win/loss records from schedules. Creates rows where each game has data for both teams involved.

Block 5: Feature Engineering - This is where the magic happens. Creates two types of features:

• Anchor features: Expanding mean of all historical stats (lifetime averages). This captures long-term team quality.
• Spark features: Exponentially weighted moving average of last 5 games (recent form). This catches hot/cold streaks.
• Also adds opponent stats (both Anchor and Spark) and home/away indicator
• Each row ends up with ~100+ features combining both teams' stats

Block 6: Walk-Forward Backtesting (currently commented out) - Simulates a season week-by-week:

• For each week in the target season, trains the model on ALL previous data (past seasons + earlier weeks)
• Makes predictions for that week's games
• Calculates accuracy and tracks results
• Uses this walk-forward approach to avoid look-ahead bias

Block 7: Interactive Predictions (currently commented out) - After backtesting:

• Retrains model on full dataset
• Lets you input any matchup (home team vs away team)
• Outputs win probabilities for both teams

Block 8: Visualization (currently commented out) - Creates charts showing:

• Weekly accuracy (volatile, shows "Spark")
• Cumulative season accuracy (stable trend, shows "Anchor")
• Reference lines at 50% (coin flip) and 60% (pro target)

Currently, only Blocks 1-5 run by default. Uncomment the sections you want to use!

Troubleshooting

• Module not found: Make sure venv is activated and run pip install -r requirements.txt again
• Slow first run: Data download takes a few minutes initially, then it's cached