This is project I've done for the past few years to predict the outcomes of the NCAA bowl games. This year we heavily leveraged Hex and a very helpful python package cfbd to analyze the data and aggregate game data respectively.

I would like to aknowledge two of my friends who helped me with this project.

• ezipe
• RyanD893

The whole process anchors on getting the last 10 years of CFB games and those teams stats for that year, then training a model on that data and using that model to pick the winner of future games, in this case the bowl games.

To get team performance data I wrote a scraper using pandas and selenium to scrape the NCAA website for the last 10 years of team performance data.

To get game data I used the cfbd python package to get the last 10 years of game data. This was a lot easier than scraping the NCAA website. Huge shoutout to the creator of that package.

The hardest part of pre-processing was getting normalization on team names to work, many teams have a shorthand or abbreviation that is used in the game data but not in the performance data. I ended up making some facilities in Hex to make this process easier but it was still very manual

Note that some games in the last 10 years are excluded because the performance data is not available for those teams. This is a small number of games and should not affect the model too much. Mostly D2 and D3 teams.

For training I made some tools in Hex to visualize model performance and confidence, and then pretty much tried a variety of pipelines and models until I found one that worked well.

Part of using sklearn is that it has a built in predict_proba function that returns the probability of each class. I used this to rank the teams in each bowl game based on how confident my model was in each prediction.

• My model was either very confident or very unconfident in its predictions. Meaning that it was either very sure a team would win or very sure a team would lose. This is a problem because it is very hard to rank teams when there is a large gap between the most confident and least confident prediction. I am not sure how to best remedy this other than getting more performance data to be able to train on more years of bowl games.
• I was interested in also making a general game predictor given how a team has performed the last 4 years (assuming some of those playesrs would still be on the team/skill carryover) but this proved beyond the scope of the project.

I should probably have a confussion matrix here but I was too tired to make one. Check out this bar graph.

This image shows the confidence curves of all the models I evaluated :