NFL Prediction Model

This repository is my attempt at trying to predict the outcome of an NFL game based on known factors.

Here are the best results so far (just a hair above random guessing):

Decision Tree F1: 0.4976 +/- 0.0305
Decision Tree Acc: 0.5299 +/- 0.0258
SVM F1: 0.4910 +/- 0.0293
SVM Acc: 0.5719 +/- 0.0211
Neural Network F1: 0.4821 +/- 0.0992
Neural Network Acc: 0.5854 +/- 0.0303

Observations:

• Seeing reduced returns on increasing data, indicating an issue with data prior to 2017 either on my end or on its usefulness in general.
• NN is still incredibly low, suggesting more data could improve this metric
• SVM is the best by far, with default parameters. Grid search currently in progress.

Getting Started

The src folder has everything you need:

• Set up a Postgres database, and create the relations using nfl_create.sql. Put the credentials in a .env.
• Use build_db.py [YYYY] to collect data from various APIs, namely the ESPN API, and nflweather.com (huge thanks to the developers of these services).
  • Example Usage: python build_db.py 2023 (accrues all data for NFL games that occurred in 2023 and writes the data to the DB)
• Use generate_csv.py to generate a ML friendly CSV with aggregated game data, with labels of "Home" and "Away" depending on who won the game.
• Finally, train_model.py will train three different models using 5-fold cross-validation (if there's more or less, its because I forgot to change this README) and output their results. At the time of writing, we have a decision tree, SVM, and neural network with default parameters.

data/ directory contains pre-compiled CSV from my local database so you can jump right into model training without doing a bunch of collection yourself.

Contributing

Here are some easy TODOs for anyone interested in contributing to the project:

• documentation for generate_csv.py, since the configuration options for the CSV are almost infinite.
• fine-tuning parameters in train_model.py, and potentially adding more suitable models. (grid searches for SVM and NN)
• update build_db.py to collect game data for prior seasons to 2015 (2023-2015 ~= 2656 samples). As of right now, it'll probably mostly work, but there's anomalies in some seasons that don't occur in others that result in imperfect data. I've only optimized up to 2017.
  • Note that increasing the volume of data is expected to improve results, however at this time, improving the quality of the data in preprocessing is probably more effective.
• The src/dataviewer directory is supposed to be a semi-separate project focused on being able to comprehensively review the data that we collected, maybe make updates to individual lines if necessary -- but mostly just provide visualization options for the data so we can discover some potentially useful insights.
• email me at [email protected] with some cool research papers on this topic