A model to predict the results of the NCAA Division I Men's Basketball Tournament bracket
python bracket.py will train the model for the 2017 tournament and output the predicted teams advancing to each round, moving backwards from the championship game to the first round.
For now this project employs a very simple model. A basic logistic regression or random forest model is used to compute the probability of a team beating another team in a head-to-head matchup using basic stats for each team. The full joint distribution for the entire bracket is then estimated using this head-to-head model and basic conditional probability to estimate the chances of each team advancing to each round.
Potential improvements may include using a more interesting model for head-to-head predictions and acquiring more data to improve the feature set, as tempo-free statistics and information about strength of schedule would probably improve results. I don't think the current features are informative enough on their own to make a model competitive with human predictions.
The following features are used for each team from 2003 to 2016:
- Points per game
- Opponents per game
- 3 point attempts per game
- 3 point %
- Free throws per game
- Free throw %
- Offensive Rebound %
- Defensive Rebound %
- Assists per game
- Turnovers per game
- Steals per game
- Blocks per game
- Fouls per game
These features are computed from historical data obtained from Kaggle's March Madness contest. The training data is included in the repo but not the Kaggle data it's computed from as I don't have the rights to distribute it.