College Football Score Predictions

Machine learning powered predictions for college football games using historical data, team metrics, and game conditions.

Project Overview

This project emerged from my passion for college football and curiosity to try and build a machine learning model. I wanted to create a prediction engine that could compete with professional handicappers while providing transparent insights into how predictions are made. The challenge was to develop a system that could process complex historical data and generate accurate predictions for future games, no matter if those teams consistently played each other or not.

Approach

The model uses gradient-boosted trees (XGBoost) inside a scikit-learn pipeline to predict a game's margin (home minus away) and total points, from which the two team scores are recovered. Crucially, it is trained only on information knowable before kickoff — pre-game Elo ratings, rest days, home field, schedule, and weather — and validated on a true temporal holdout (train on 2008–2010, test on the unseen 2011 season) so the reported accuracy isn't inflated by data leakage. On unseen games it identifies the straight-up winner roughly 70% of the time with a margin error of about 14 points, beating both naive-average and Elo-only baselines. (Betting lines were unavailable in this dataset, so the model is benchmarked against statistical baselines rather than the Vegas spread.)

Directions

Simply navigate below and select the two conferences from the drop down list. You will then select the teams you wish to choose, which will then predict the scores and provide the spread and model weights.

Key Learnings

  • Data quality and cleanliness predicate any quality model
  • Look-ahead leakage is the silent killer: features like rolling averages must be lagged (shifted) so a game never "sees" its own outcome, and the train/test split must respect time — otherwise validation scores become fiction
  • Honest baselines matter: a model is only worth shipping if it beats naive and Elo-only benchmarks on a held-out future season
  • ML models require subject matter experts to review results and help to train those models. Many times the predictions revealed results that did not make sense, such as the away team's venue id being the most important feature

Challenges Overcome

  • Initial models continued to weight away features more heavily
  • Managing data quality across different seasons and rule changes made predictions more challenging
  • The lack of player data resulted in less quality predictions, especially for early season matchups

Future Applications

  • Implementing automated model retraining. For example, the outputs of one function can be read by the AI to adjust the code to the next input
  • Creating similar models to other sports

Game Prediction

VS

Prediction Results

--

--

--

--
Favorite: --
Spread: --
Over/Under: --

Model Weights