As you probably know, TL has its own rating system separate from FICS standard ratings. The separate system, known as "fixed ratings", has a couple of benefits. First, it discourages sandbagging (intentionally losing rating points) before Teamleague begins, because losing FICS standard rating points will have only an indirect effect on the fixed rating. Second, for some players, past TL performance may be a better predictor of future TL performance than FICS standard rating; one reason is that FICS standard rating applies not only to 45+45 games, but to significantly faster time controls as well. Hence by taking TL performance into account, it may be possible to determine a more accurate rating for TL games.

Fixed ratings are generally set before a TL edition begins. The exact timing is not announced, to further discourage sandbagging.

Due to an apparent miscommunication between programmers, the fixed ratings set by jaberwock when setting up T50 were not ideal. The algorithm that set the initial ratings took an average of the FICS standard rating of each player over time, and did not take into account TL performance at all. Some players noticed their new ratings did not make sense.

Therefore the admins decided to recalculate the fixed ratings using the algorithm used in past editions. As far as I know, this algorithm, written by seberg, hasn't been publicly documented, but it wasn't a secret, either. It uses a weighted average of the user's current FICS standard rating and the user's TL performance rating in the past 4 editions. TL performance was given at most a 30% weight (less if the user has fewer than 5 games in the past 4 editions).

It would be easy to recalculate the ratings using that method, but I thought this seemed as good an opportunity as any to improve the algorithm. So I made these changes:
* Increase the weight of the TL performance rating to a maximum of 50%. The weighting is less if the player has fewer than 10 games in recent editions.
* Use the TL performance rating from the past 8 editions, instead of the past 4 editions.
* Give recent TL editions a higher weight in the performance rating calculations than editions farther in the past.

I based these changes very loosely on Jeff Sonas's Chessmetrics method. The major difference between the TL performance rating and Chessmetrics is that I don't pad the results like Chessmetrics, because the specific padding numbers Jeff uses are designed for top-level players and wouldn't work well for TL. My choice to use the past 8 TL editions roughly corresponds to Jeff's recommendation to use a 24-month time period.

I've recalculated the fixed ratings with my changes to the algorithm. I realize the ratings can be a controversial issue. If people don't like the changes, it's always possible to go back to the old algorithm that was 30% TL performance.

Thank you for sharing this informative discussion with the community, wmahan. I, for one, support all the changes that were made, especially elongating the time frame and giving greater weight to more recent performance.

One small question: the TL rating is computed based on the games of the last x TLs, rather than the overall performance in each TL, right? In other words, if during one TL a player completed 3 games and in another 6 games, the latter would carry twice the weight of the first (ignoring the "recent-ness" weighting for the moment), correct?

I like the changes. And this is not the past 8 TLs for a given player (which could go far into the past if the player hasn't been active), right, but instead the past 8 TLs hosted?

tseltzer, that's correct; the calculations are based on games, so playing 6 games in one edition is weighted higher than playing 3 games in another edition, ignoring the recentness.

JoshuaR, right, it is based on the 8 globally most recent editions. For example, if a player has not played since T40 and joins T50, then the fixed rating is simply set to the player's current standard rating (as is done for new players who have never played TL).

Thanks for this! Very clarifying.