NHL Meter

Predictive win probability for National Hockey League (NHL) games using neural networks trained on game state and play-by-play data.

About The Project

This project was originally developed alongside Alex Hagood for CPT_S 475: Data Science at Washington State University taught by Assefaw Gebremedhin. It was inspired by visualizations like ESPN's Win Probability graph for basketball and football. Similarly, this model provides live estimates of NHL game outcomes.

The play-by-play data is sourced directly from the NHL API using hockey-scraper and represents every in-game event between the 2007-2008 and 2023-2024 seasons.

Additional libraries include:

• Keras and TensorFlow for neural networks implementation.
• Dash and Plotly for web dashboard and visualization.

Implementation

The three regulation periods of each game are divided into 30 second time slices representing the game state at the end of each interval, including each team's Elo, total score, shots on goal, hits, strength, and penalty minutes. Regulation and overtime periods are split into sliding windows of 3 plays to determine the plays leading to scoring opportunities for golden goal overtime. Slices and plays are fed into their own LSTM neural network to separately calculate regulation and overtime win probabilities, which are concatenated for the final graph.

The Dash app allows users to generate graphs by filtering by Home and Away Teams, then selecting a Game from the dropdown.

Usage

Built and tested with Python 3.11.9 but may support other versions.

Installation

To just run the dashboard with the included data and models, install packages from requirements.txt

pip install -r requirements.txt

To scrape and clean the NHL data and retrain the models, install both requirements.txt and requirements-dev.txt

pip install -r requirements.txt requirements-dev.txt

Running the Dash App

python app.py

The web app will be available at http://localhost:8050/

Docker

The container uses the python:3.11.9-slim image and installs packages required for prediction and visualization.

docker build -t nhl-meter .
docker run -h localhost -p 8050:8050 -d nhl-meter

Scraping, Cleaning, and Training

1. Scrape play-by-play and shift data from NHL API. Note that this may take multiple days to complete

   python ./dev/data/scrape.py

2. Handle raw data

   python ./dev/data/raw_to_reduced.py

3. Train the regulation model using lstm.ipynb and the overtime model using lstm_ot.ipynb. The new models are ready to be used with the dashboard.

Testing

To run the test files using pytest:

pytest tests/test_callbacks.py
pytest tests/test_prediction.py

test_callbacks.py will test the Dash app functionality. test_prediction.py will run the appropriate win prediction models on all games (~22,000) in the dataset.