Also, for the full rankings:

https://dl.dropbox.com/u/196031/rankings_adroits_2012.pdf

and the current ladder:

https://dl.dropbox.com/u/196031/ladder.pdf

cypher not on top4? the betting community will go nuts

GO PURRI TOP 8

"based on their online performance recently".

So it really only matters who they've been playing. There is no ladder system. Nothing that makes them play at their skill level. For instance:

K1llsen:
win %: 94%
Avg. Win Opponent Elo 1649
Avg. Loss Opponent Elo 2212

Cypher:
win %: 77%
Avg. Win Opponent Elo 1808
Avg. Loss Opponent Elo 2193

Rapha:
win %: 89%
Avg. Win Opponent Elo 1780
Avg. Loss Opponent Elo 2227

Those are averages. Thinking about the numbers for a minute, I think the "avg win opponent elo" is telling. It's not showing the skill level they are able to defeat on average, per se. These are such high level players that their win percentages are extremely high. So a lower "avg win elo" really just means they are playing lower skilled players more often. Hence K1llsen's 94% win percentage, as compared to Cypher's 77%.

Looking at K1llsen's match history on QLR, you see him smashing 1200 elo noobs 37 to 3. If these pristine "victories" are being included in your calculations, then I would question your predictions, yes. Your number crunching is more of a model for the players' behavior than it is a tool for prediction.

It is more intelligent than that. The system is a generalization of ELO, so I am taken into account that k1llsen is playing lots of matches against unskilled players. It doesn't improve his ranking. You are right it doesn't tell you jack shit about how good k1llsen is.

This is also why it has worked well in the past (At DHS 2012 I predicted the 1st to the 4th position correctly, btw). Note that I threw away my optimized configuration I found by simulated annealing before DHS in the above. I have made a re-run in the pdfs, but it doesn't look like there is some large upsets.

It is true that LAN is a different beast than on-line. So I guess I am answering an interesting question: ``To what extent can on-line play be used as a predictor for LAN play''.

Finally, the ladder is a conservative estimate of a players skill. Basically I know what skill range a player falls in with 95% confidence . I take the lowest number of that range times two as the skill of the player for the ladder. The reason this is a good number is that I loose confidence when a player is not playing. So if you don't play, you will fall down the ladder.

I like it man. I'm a software engineer, and I do a fair bit of data analysis and create data models for research from time to time, so I find this very interesting.

LAN is quite often a whole different ball game compared to online romp around.
But that shouldn't really affect the upper echelon of players with vast LAN experience ...