So today I’m releasing a new version which introduces three new features:
1. The model will factor in the distance each team has traveled as a fourth predictor (along with a form Elo formula, a rating of each player in each selected team, and how often the team plays home games at the venue).
2. It centres and scales each of these four predictors from -100 to +100, meaning that I will be able to give an easy-to-interpret assessment of the reasons for each prediction (see below).
3. Each week it selects a few of the empirically best of potentially 68 possible machine learning models in a weighted stack rather than just one. These include random forest models, gradient boosted trees and generalised linear models with elastic net regularisation.
I’m hoping these changes will lead to more accurate predictions and a better display of what the model is looking at before making each decision.
