The successor to Footy Tipper. I decided to start over from scratch, because:
- I needed a better name.
- Footy Tipper was more than a little refactoring away from being maintainable or extendible.
From what I had learned the first time around, I was able to write cleaner application code and create a roughly-equivalent model that was much simpler and faster to train. The current iteration is a stacked ensemble with an XGBoost regressor, ELO model, and ARIMA model all feeding into an Extremely-Randomized Trees meta-estimator. As a special bonus, there’s an extra XGBoost regressor for predicting win probabilities in order to participate in the Monash University probabilistic tipping competition. As an extra special bonus bonus, there’s now a mostly-functional website at http://www.tipresias.net that will display tips and track model performance.