Originally posted by Crowley
I've been moaning about this for a long time now.
Sounds good, but I'm no mathematician, can you give a better explanation or maybe an example?
I can indeed but to save on typing lists of 300 values, I'll simplify a bit!
A practical example:
Last 30 games, ratings are 1500, 1490, 1480 … 1210 (decreasing 10 pts a game)
The 270 games before that: 1800 (50 times); 1850 (100 times); 1900 (70 times); 1900 to 1500 (a slide over 50 games).
Step 1) The upper decile of last 30: 10% of 30 is 3, so were looking for the third top value.
In order, we have 1210, 1220 ... 1470, 1480, 1490, 1500. The third top value is thus 1480.
Step 2) The upper decile of last 300: 10% of 300 is 30, so were looking for the value 30th from the top. As we have 71 instances of 1900, then the value 30 from the top is 1900.
The biggest from these two is the player’s tourney rating: i.e. 1900.
Under the current scheme, it would be the max of the last 30 which was 1500, a value well below this player's true strength.
On the other side of the problem (over-inflated rating), a solid 1200 player who gets + 40 and +40 on two games by T/O when playing a big rating-slide player ends up at 1280 under the current system. Using the above, there tourny rating would still be 1200 as you need 3 extreme values in step 1 before the tourney rating shifts. Thus players are not penalised from lower-banded tournys when a couple of 'lucky' wins happen.