In an effort to create objective rankings for REBBL teams and coaches, I've been working on REBBL Elo. Since rankings are worthless unless they can be used to predict results, I'm using this week's predictions as the first big test for the system and you get to see your ratings compared to your divisions peers!
How it Works:
1. Using the great rebbl.net new API-only endpoints, I was able to fetch and parse every result from REL and GMAN since season 1. That's around 15k games!
2. Using the first season as input, the system tries to predict all games in the order that they happened, updating each team's rating according to the results using an adapted version of the Elo Rating Algorithm. Bottom line, you get more points for beating higher rated teams and lose more points for getting beat by lower rated teams.
3. Several parameter sets were compared and I picked the one with the best results to be this season's official version.
- Dealing with draws: although I have some ideas of how a draw chance could be assigned, it doesn't seem to be widely used and I'll need some more work on the future to "prove" that a predictor that tries to pick draws actually performs better than the current version.
- Playoffs and the BIG-O: the current version has a very simple way of storing results taken from the API and what I'll work on first is refactoring it into an actual database, making querying the data much easier in the future. This means dealing with playoffs and results from the BIG-O is not a priority at the moment. Playoffs are important because the teams that do well on those games are usually placed in higher divisions and would be taking rating points into them, increasing the difference in average ranking between the different tiers over time.
- Race matchups are not taking into consideration the past week's games and make no distinction over which phase of development the teams were in. They are displayed just for reference, not being used on the prediction itself.
- Admin teams are treated as normal teams, so they start with an average rating which will decrease as they lose games.
- Coaches are not ranked right now, but the plan is to take that into consideration on future versions.
- This is the first working version and most of the work has been on organizing the data. I suspect the current prediction quality will improve as more features are added.