$90 Houston Match Prediction Contest

The Elo model is similar to yours in that it uses winning margin, the same k values for qual and playoffs, and 20% mean reversion between years. It’s different in that all teams start with 1500 and I additionally weight the Elo update by team Elos going into a match, so higher Elo teams get a larger chunk of the pie. The score scaling is updated throughout the season (I believe your model only uses week 1 data). Also, I don’t believe the exact Elo parameters are too important, since the probability of win is learned by the neural network.

ixOPRs are calculated with 2 iterations during a live event. They get reset at the beginning of the season and are scaled so the neural network works better across seasons.

The network and Elos are updated using all matches since 2007, excluding 2015.