I backtracked from
Baxter's thread to the old forum and found
Pook's post containing the ELO calculating code (hoping it's still up to date). I'm posting this here, since it seemed too technical for the thread over in the general forum.
Code: Select all
-- BEGIN PROCESSING GAME
INSERT INTO @TeamELO(TeamNumber, TeamPlayerCount, ELORating, Winner, Playtime, MaxPlayertime)
SELECT gt.GameTeamNumber, Count(p.Member_ID), SUM(((p.ELORating - 1500) * p.Modifier) + 1500) / Count(*), gt.GameTeamWinner, SUM(p.Playtime), MAX(p.Playtime)
FROM GameTeamMember gtm, GameTeam gt, @Players p
WHERE gt.GameTeamIdentID = gtm.GameTeamID
AND p.Member_ID = gtm.GameTeamMemberMemberID
AND p.TeamNumber = gt.GameTeamNumber
AND gt.GameID = @GameID
AND p.Playtime > 0
GROUP BY gt.GameTeamNumber, gt.GameTeamWinner, p.TeamNumber
So the team ELO is SUM( ((p.ELORating - 1500) * p.Modifier) + 1500 ) / Count(*)
If I understood correctly, p.Modifier is a float between 0.0 and 1.0 that expresses for each player how long time he or she played in the game as a fraction of the total game time. I assume Count(*) is the number of team players. Imagine a team where new people continually joined and only stayed for a short while. For all these players, their p.Modifier will be close to 0, so the team ELO can approximately be simplified to
SUM( 1500 ) / Count(*)
What I'm trying to say is that people who only join the game for a short period push the team ELO towards 1500, regardless of their ELO. I think something like this would be a more correct assessment of the teams' skills and efforts:
SUM(p.ELORating * p.Modifier)
or in pseudo-code
TeamELO = sum( [p.ELOrating * p.playTime for p in TeamPlayers] )
So if a very skilled player plays for you for a short time, he contributes approximately as much as a voob who plays for you the entire game. The number of players doesn't matter when aggregating the team's ELO.
Hmmm... It would be interesting to copy the existing database and re-calculate all elo-ratings using this algorithm... I wonder if that would be more accurate. I've tried to think of a simple way to measure how "accurate" the ELO ratings are... I haven't come up with anything conclusive, but...
At the heart of the ELO system are the predictions of how large chance a team has of winning a particular game, the "Expected Outcome". If the ELO ratings are accurate, then these estimates should be fairly accurate.
Now... Consider another situation: If you want to see if a particular dice is weighted, you can roll the dice a hundred times and see how often each face comes up. So one way to test it would be to take two teams and let them play against each other 100 times. Then compare the win-percentages to what ELO predicted. But that's not feasible. Instead of throwing the same dice many times, we're throwing a new dice each time, since every pickup game is unique.
But what if we can group similar dice... I mean games... together? For example, we could check all the games where the estimated outcome was between 75-25 and 65-35. Then calculate how many games were won by the underdogs. That would be one way to measure the accuracy of ELO. If it deviates significantly from 30, then we should become suspicious.