PDA

View Full Version : Generating ratings



Aaron Guthrie
23-01-2012, 11:49 AM
I want to generate ratings from: games between 20-40 players; who have no previous ratings; who play between 20-100 games.

Can anyone help me by point me in the direction of the easiest way to do this, for ELO, Glicko, or any other rating system?

Also, how many games does it take to get a reliable set of ratings? And are there problems with using a closed pool from a small number of players?

Rincewind
23-01-2012, 12:05 PM
The easiest way is probably to give then a median starting rating and just run the games through in a single period using either Elo or Glicko. After doing so you could then compare the performance rating with the final rating if it is close then you probably have a reasonable set of ratings. If not then you could try repeating the process but instead of starting everyone with the median rating, use the rating you calculated from step 1 and at the end of this process again compare performance ratings with final ratings.

Not for players with exceptional performances (close to 0% or 100%) this process might never converge and so you would be best to ignore them in the comparison.

This all assumes that the period over which all games all took place was short enough that the individuals did not substantially improve. If this assumption is false you might want to break the set of games up into a number of periods and run each as rating periods, with the output rating of one period feeding in as the starting rating of the next.

Patrick Byrom
25-01-2012, 02:50 AM
I've written a program to calculate ratings from Swiss Perfect files where there is a large percentage of unrated players (it's used for the Qld Junior Rating List), which you can download at the bottom of this page: http://www.southsidejuniorchessclub.org/Downloads.htm.

It uses the Elo system, but that probably won't be too significant initially. Once most players have ratings you can switch to a more sophisticated method.

You should have more than enough games to produce reliable ratings, but they may not be very accurate. I would try and estimate the rating of one (or more) players in the group(s). If you know one rating with reasonable accuracy, then the rest should be okay.