>You reduced RD to 50. Why is 50 better than 49? Or 55? I think, after such a value is set, you should stick to it; it is an arbitrary decision anyways.
It's not an arbitrary decision. There is a value that makes the ratings maximally accurate, this appears to be ~30 for lichess. Why it's 50 now and not 30 I don't know, but I guess it's going to be gradually reduced to see the impact.
I support this decision. Your rating is meant to represent your playing strength. If it's not doing that as accurately as it could, then it should be fixed.
> Why it's 50 now and not 30 I don't know, but I guess it's going to be gradually reduced to see the impact.
This makes a lot of sense. If moving it from 60 to 50 generated complaints, then moving it from 60 straight to 30 would probably generate a lot more.
#11 Indeed, GCP's analysis suggests that <= 30 is the best rating floor, in terms of predictive accuracy.
#12 Exactly. Very well said.
I begin to understand. Just watching the graph of my rating development of the lats 3 months is amazing. A very unsteady zig zag line starts to become smooth and continuous some 2 weeks ago.
As a matter of fact, I was never complaining. I just thought any change is bewildering; I still don't get why that change had to be now and not 3, 2 1 years ago ...
>I still don't get why that change had to be now and not 3, 2 1 years ago ...
The original "RD cap" at ~60 was not intentional and not actually a parameter in the lichess code. It was the result of using formulas from the Glicko-2 paper that were meant to be used with fixed-time rating periods (i.e. monthly) after every game. After fixing those to correctly account for the variable passage of time it is possible for the RD to go below 60. It seems the lichess admins will now gradually allow the system to converge to the optimal values by explicitly putting the cap at 60 (as it was implicitly before) and then lowering it in steps to the cap that was determined to be optimal.
You can see the discussion and research related to this in the github issue linked earlier. It wasn't done earlier because the problem wasn't understood and the admins are very careful with changing the rating system. The initial analysis on the problem was done 1.5 years ago!
@Toadofsky How long will it take for notorious rating boosters to get off our leaderboards? Ingrid-Vengeance is a well known account who clearly only plays people around 1500 in rating to stay on the leaderboards (and has done so for ages).
Does the RD floor need to be reduced further?
@letzplaykrazy Provided that they can maintain an RD <= 80, they've earned their position:
lichess.org/@/Ingrid-Vengeance/search?perf=0Currently developers are considering adjusting that threshold based on category; and rejected a proposal to increase by 13% the rate at which RD increases (currently it takes 412 days for a player at the RD floor to become provisional again; before the floor change it used to take 365 days).
The solution in terms of rating systems is to rank the leaderboards by lower confidence bound (also pointed out in the github issue), i.e. rating - 2 * rd, but that's going to be counter-intuitive for a lot of people.
You could also not rate games where the rating discrepancy is too big, say >700 Elo.
It's a bit surprising that playing a provisional rated player who is rated 1000 Elo less has any measurable effect on RD, let alone enough to keep it below 80. Maybe one could derive a suitable lower bound for the leaderboard from that: i.e. search for the point where playing a completely unknown player doesn't reduce the RD enough to compensate for the increase due to time.
Personally I prefer that ranking by lower confidence bound.
I have a bias for action (to continuously tweak parameters) although most developers advocate as you do that good reasons need to be articulated for any change to occur. I do think the RD <= 80 filter improved the leaderboard, although my preference would be for a different system altogether (such as a ladder, league, or something else) to replace it.
#19 Agreed, the RD <= 80 filter was a good improvement to the leaderboard.
#18 I like your idea regarding finding the suitable lower bound for the RD filter.
My main point: If someone can remain on the leaderboard by refusing to play anyone near their own rating, then the leaderboard is not functioning properly.
Until this is fixed, the leaderboards will remain a joke.
Ultrabullet rankings at current posting:
1 GM penguingim1 2900
2 GM Zhigalko_Sergei 2623
3 Ingrid-Vengeance 2504
Only 2 of the top GMs are ahead of someone who only chooses to players below 1600.
What needs to happen to actually change this?
