As in because it’s a bunch of dead rating points which have made some impact to the ratings of others?
It is what it is I suppose.
If quitting was a problem for the ratings system as much as letting players join at any (potentially incorrect rank) then us correspondence players must be destroying the ratings by playing slowly – in some cases effectively “quitting” for a few months waiting for games to end 
Yeah, that’s a good question. Wouldn’t players joining in, improving a little then quitting deflate the rating system a little. Like if we all simultaneously got 2 stones stronger, then ranks wouldn’t change (since relative strength is the same) but each rank would be harder. Aren’t we doing the same but in a less organized way with everyone getting stronger little by little.
If it just comes down to people not joining at their true rank, not some fundamental implosion of the whole system, then that’s actually an argument for letting beginners join as beginners.
As long as on average players take away or leave behind 0 rating points between joining and quitting, I suppose the overall rating point pool should be unaffected. That may be the case with the current default initial rating, which seems to be fairly close to the average OGS player rating.
If the default initial rating were much lower, I expect the rating system to become deflationary over time, because most players would take rating points away from the other players while quickly gravitating to their true (usually higher) rating, or just slowly improving to a higher rating.
In that case you might have to add some mechanisms to counter deflation.
The EGF rating system has 2 mechanism to counter deflation:
I think with the current Glicko system there isn’t a “conservation of rating” rule as well right?
As in if I beat someone I get x rating points, but they don’t necessarily lose x rating points?
| ended | game_id | played_black | handicap | rating | deviation | volatility | opponent_id | opponent_rating | opponent_deviation | outcome | extra | annulled | result |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 1628083641 | 35765583 |
1 | 0 | 1859.29 |
65.45 | 0.060019 | 717368 | 1380.80 |
60.88 | 1 | null | 0 | Resignation |
| 1628035747 | 35765584 | 1 | 0 | 1857.76 |
64.88 | 0.060020 | 856379 | 1825.63 | 61.33 | 1 | null | 0 | Resignation |
Here’s a recent game with id 35765583 and the rating history.
The history of the opponent is
| ended | game_id | played_black | handicap | rating | deviation | volatility | opponent_id | opponent_rating | opponent_deviation | outcome | extra | annulled | result |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 1628143334 | 34909949 | 1 | 0 | 1371.61 | 61.54 | 0.059887 | 380817 | 1476.11 | 64.86 | 0 | null | 0 | Resignation |
| 1628083641 | 35765583 |
0 | 0 | 1379.44 | 61.54 | 0.059889 | 427361 | 1857.76 |
64.88 | 0 | null | 0 | Resignation |
| 1628055336 | 35314297 | 0 | 0 | 1380.80 |
60.88 | 0.059889 | 345855 | 1377.95 | 61.53 | 0 | null | 0 | Resignation |
So I gained 1859.29-1857.76=1.53 points for a win while they lost 1380.80-1379.44=1.36 .
The more drastic example should be a new player vs an established one.
| ended | game_id | played_black | handicap | rating | deviation | volatility | opponent_id | opponent_rating | opponent_deviation | outcome | extra | annulled | result |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 1627293093 | 35620327 |
1 | 1 | 1163.25 | 136.66 | 0.059997 | 906618 | 1308.30 |
60.76 | 0 | null | 0 | 47.5 points |
| 1627292070 | 35620088 | 1 | 0 | 1210.01 | 147.62 | 0.059998 | 798997 | 1772.64 | 66.91 | 0 | null | 0 | Resignation |
| ended | game_id | played_black | handicap | rating | deviation | volatility | opponent_id | opponent_rating | opponent_deviation | outcome | extra | annulled | result |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 1627293093 | 35620327 |
0 | 1 | 1316.84 | 60.88 | 0.060008 | 1017686 | 1210.01 |
147.62 | 1 | null | 0 | 47.5 points |
| 1627292396 | 35620280 | 1 | 0 | 1308.30 | 60.76 | 0.060009 | 1013740 | 1346.75 | 60.61 | 1 | null | 0 | Resignation |
In the above example the first player, fairly new (with high deviation ~130-140) lost to a player with low deviation (~established) of ~60. The first player lost 1210.01-1163.25=46.76 points while the second player gained 1316.84-1308.30=8.54.
So the ratings aren’t zero sum in any sense.
I think there’s the potential for there to be some rating leakage or inflation in such a system right? I kind of expected it anyway if one wants to have deviation dependent rating changes and asymmetric changes whereby an established player losing to a new player won’t lose as much rating, since one is uncertain their real strength. Also one wants to adjust new players more drastically than their opponents.
I don’t know the long terms effects of such a system of course, especially if there’s any random injections of new players at various ratings, or people abusing the system, sandbagging, airbagging, lot of new players joining and quitting etc.
You are misinterpreting my posts. As I stated, the reason Glicko was instituted, according to what I was told, was to get away from having constant requests to adjust people’s ranks. My comments about sandbagging are entirely based on things that I personally noticed and have nothing to do with the adoption of Glicko. The effect on sandbagging is merely an additional value of Glicko, not the main rationale.
OK, then it is like this?
Too many rank adjustment requests => adopt Glicko => no more rank setting (fixing the original problem), added advantage of reducing sandbagging, added disadvantage for ability of new players to find suitable opponents
I don’t think it’s explicitly the rating algorithms fault that beginners and new players can’t find suitable opponents. It’s instead the pairing algorithm as @S_Alexander shows.
https://forums.online-go.com/t/new-player-ogs-user-experience/38482/36?u=shinuito
In theory if the rank restrictions on games weren’t present then one could be a new player and challenge anyone they want in a custom game. Similarly with automatch the problem is that there is a limitation on getting a rated game with more than 9 ranks difference (which of course doesn’t make sense given the ladder and tournament systems).
Ofcourse it’s better to at least make it possible for new beginner to filter for beginner opponents.
But even if the limits were widened, how is a new beginner supposed to know how to modify that range filter to find other beginners (at a much lower rating than they have themselves)?
And it would not even work well when many of those other beginners are to be found at that (much too high) default rating, like themselves.
Is it something calculable? If so, we can try to calculate it.
I addressed this in an earlier post that no one commented on. Restating: Leave humble rank in place to preserve Glicko, but automatically match all provisional players with 25k players until they achieve a real rank. Or, if the game pool is too small, match them with 20-25k players (or some other reasonable range). Provisional players would be given no other options until achieving a real rank, so they wouldn’t need to know anything—just click a button that says “play game,” with a choice of sizes.
You can see the problem with this though right? Imagine there’s another initiative like the Western Dan Challenge, which aimed to get more Dan players playing regularly on OGS (and other servers). A bunch of new dan players sign up and have to thrash established 20-25kyus or other potential beginners to get a rating until their account is no longer provisional 
And I bet their established rank would still be far from their true rank. Beating ten 25k players is probably not enough (even for Glicko) to establish a (near) dan rank.
I have a cynical response to that, but I will leave it alone. Thanks for the comment, however.
I don’t think it’s neccessary to have users specify an exact rank when registering, but I think it would be better to at least let new players make a rough selection between for example “beginner”, “casual”, “intermediate” and “advanced”, initialize their rating to a value that seems to roughly match that label (25k, 18k, 11k, 4k?) and let Glicko take it from there.
2 choices are enough. Normal will work like now.
Beginners will be paired with as weak rank as available in automatch and they shouldn’t be able to play ranked with SDK in custom for a while.
And if you records their selection we could measure what each level correspond to. And how much dunning kruger is there.
You’d have to tally the opponent’s rating changes of all their games, I guess.
Maybe the dunning kruger balances the sandbagging 
I apologise if I come across badly, I know the phrasing could’ve been nicer.
I guess I just meant that a solution involving forcing all new or provisional players to do X probably will run into issues, because new and provisional players could be dan/sdk/ddk players who don’t necessarily want to play beginners any more than new beginners want to play dan/sdk/ddk players 