This is a work in progress
current best MAE is 14.75 with RandomForestRegression
Sports betting predictive modelling project
By: PerryGraham & alden-august-wang
The initial purpose of this project is to predict the Over-Under of an NBA game.
Data collection:
Data downloaded from Kaggle from Nathan Lauga.
Data cleaning:
- Added a point total column
- Exculded data before 2018 season
- Removed null values
- Sorted by date
- Combined data from team ranking history
- Encoded team IDs and Season
Data insight & visuals:
- Opened cleaned data file with Tableau to explore the dataset
- Median total points = 221
- Upper hinge = 235, lower hinge = 168
Feature engineering:
- At first I only used past 2 years of data, because I thought since teams change a lot it wouldnt makes sense to take longer periods of data, of teams that dont exist anymore. However, adding the last 10 years of data helped a lot in getting better results
- Then I tried using less features (only ones with high correlation with the target, this actually got worse results across all the model types
- I tried manually scaling all of the data to see if that would help get better result in any of the models
- The random forest resgression model currently yields the most accurate results out of all the regression models that I have tried
Model fitting:
Cross valiation:
Results:
svmihar's approach
problem statement
given 2 distinct teams, what is the score? f(x,y) = P(x|y)
or, to oversimplify, this is basically a recommendation engine, like the movies / books dataset.
data
columns: team_a, team_b, date, score
approach
matrix factorization using apple's turicreate xgbregressor using xgboost lightgbm using lightgbm fastai collab collaborative filtering
future improvements
- add more features
- win streak
- player stats (with a player scoring function)
- ??