So what I'm seeing in this graph is that Carbonite has a much longer range, and even at about 1000 points out from the start of the League, Carbonite players are still getting pulled in more than Kyber. You would expect this larger correction when that tail is longer, but I still have to wonder if the asymmetry of the distribution is a huge reason why the quantiles are not being preserved during the squish.
So, either the formula needs to be changed to account for this (which I believe is essentially what your suggestions about separate calculations boil down to), or they need to do something about the flooring effect of people sinking through the Carbonite ranks.
So what I'm seeing in this graph is that Carbonite has a much longer range, and even at about 1000 points out from the start of the League, Carbonite players are still getting pulled in more than Kyber. You would expect this larger correction when that tail is longer, but I still have to wonder if the asymmetry of the distribution is a huge reason why the quantiles are not being preserved during the squish.
So, either the formula needs to be changed to account for this (which I believe is essentially what your suggestions about separate calculations boil down to), or they need to do something about the flooring effect of people sinking through the Carbonite ranks.
The biggest problem IMO is that linear projections don’t work on skewed distributions with long tails…… and that they need to monitor the effects of their changes more carefully so they can do midstream corrections.
Tbh no one on discord cares about carbonite - in my crowdsourced survey, 90% of responses were kyber 1 and 2 players. And the correction in carbonite benefits players rather than hurting them. Thus the focus on kyber.
Short term we need to correct kyber divisions back to target numbers. That will be a one time adjustment.
Long term I think they need a much flatter adjustment through k2 and bottom of k1 with possibly an exponential squish rather than linear once you get past ~3850 or so ratings. Squish the top of kyber1 down hard to help with matchmaking, but shelter the boundary with k2 to preserve the percentages of players.
We're down to around 2700-2800 players in K1 after the first week of 5v5. Since CG is going on vacation next week the January squish should take us close to 2000 players for K1. Perhaps this is intentional, but it certainly will adjust player behavior as well.
Replies
You're still being pushed to the middle of the player base though - that's chromium league. It looks like all of Aurodium is pushed downwards.....
December squish
My prediction is we will be down to ~3000 players maximum in kyber 1 in week 2, after the squishes impact brackets.
So what I'm seeing in this graph is that Carbonite has a much longer range, and even at about 1000 points out from the start of the League, Carbonite players are still getting pulled in more than Kyber. You would expect this larger correction when that tail is longer, but I still have to wonder if the asymmetry of the distribution is a huge reason why the quantiles are not being preserved during the squish.
So, either the formula needs to be changed to account for this (which I believe is essentially what your suggestions about separate calculations boil down to), or they need to do something about the flooring effect of people sinking through the Carbonite ranks.
The biggest problem IMO is that linear projections don’t work on skewed distributions with long tails…… and that they need to monitor the effects of their changes more carefully so they can do midstream corrections.
Tbh no one on discord cares about carbonite - in my crowdsourced survey, 90% of responses were kyber 1 and 2 players. And the correction in carbonite benefits players rather than hurting them. Thus the focus on kyber.
Short term we need to correct kyber divisions back to target numbers. That will be a one time adjustment.
Long term I think they need a much flatter adjustment through k2 and bottom of k1 with possibly an exponential squish rather than linear once you get past ~3850 or so ratings. Squish the top of kyber1 down hard to help with matchmaking, but shelter the boundary with k2 to preserve the percentages of players.