New Elo Ranking System for Tennis Ladders

I think it still needs a little work.

Trevor,
I just started 4 new ladders on my network. 16 players per Ladder divided in Cat. A, B, C and D.
One thing that I have some doubt is the penalties for decline. I beleive that should be possible to set this in % of players points. Maybe 100 points of penalty is a lot for a 700 points player, but not too much for a 2000 points player.
At this time I am not sure what is a good penalty for my ladders. I started with 150 , to avoid declines. But I see when the points get better calculated this will be a lot.
Any comments?
Tks,
Daniel

Trevor,

After a large discussion with some players and after one month that we adopted the new ELO Ranking we need urgently some corrections. Analyzing the formula, the K factor of 800 being divided by the number of the games played is causing a dificulty for a player for example that already played 10 games to jump over a player that is in front of him and have only 4 or 5 games played. Our suggestion is to have the possibility to manipulate this K factor and also to have a fixed K factor. In our case we beleive that starting with 1500 points and having a fixed K of 200 would work much better than the actual model.
Thanks
Daniel

Just FYI, there are variants of ELO ratings which take into account margin of victory. Perhaps there might be some interest in experimenting with these to see whether they perform better.

For instance, Nate Silver's group at FiveThirtyEight uses the following formula for basketball games (see fivethirtyeight.com/features/how-we-calculate-nba-elo-ratings/ ):

K * (MOV+3)^(4/5) / (7.5 + 0.006*elo_diff)

where MOV is the margin of victory -- in our case, presumably this would be something like max( 1, (total games won by winner) - (total games won by loser) )
so that the winner of the match always has a positive MOV. Also elo_diff is
(Elo ranking of winner) - (Elo ranking of loser),
which will be positive if the higher-ranked team wins but negative if the lower-ranked team wins. The values 3, 4/5, 7.5, 0.006 don't seem to be set in stone, instead it sounds like they were obtained by trying a few values until the ranking changes "felt right".

The Elo system was originally designed for chess, where there was no way to quantify a lopsided win versus a close match. When it is applied to games which have a score (like football or basketball), it seems that people usually incorporate margin of victory to make it work better. For instance, if two people have similar Elo ratings going into the match then a 6-0 6-0 score suggests that their ratings should change a good deal, while a 7-6 6-7 6-7 score suggests that they should maintain similar ratings. The formula used at FiveThirtyEight is based on the principle that Elo ratings provide an "expected score" for a given match, based on the difference between the initial ratings of the two players. If the actual match score is similar to the expected score, then there will be little change in the players' Elo ratings. But if the match score differs significantly from the expected score, then this gets incorporated into the computation of the post-match ratings. For one more example: if a 1200 player plays an 1800 player, and the match comes down to a 3rd-set tiebreak, then it seems strange that the outcome of the final few points should have a dramatic impact on the two players' ratings -- it seems that the fact the match was close tells us more about the players' relative levels than merely knowing who won.

How about using the Calculated Level of a player multiplied by 1000 as the default rating?
This way, the ladder starts with the already established ability of a player instead of the top and lowest ranked players having the same default ratings point.