Friday, May 20, 2011
I just discovered a significant error in the code underlying the ISR, Wilson and Colley ratings. It's a variant of a problem I've had before -- information from later games "leaking" back into the ratings for earlier games. In this case, it was just operator error -- I was using a function that returned all the games for the season where I should have used one that returned all the games so far. I'm re-running the assessments of the affected rating systems and will post results when I have them, but the preliminary results indicate that the performance of these rating systems will take a significant hit, which may return TrueSkill to the top of the leaderboard.