The rating is deflating for a reason that you did not mention, which is improvement. Let's say the average person starts at 600, but over the course of a year gains 300 rating point's. This will cause the average to go down, because the 300 rating point's came from other players, which decreases the average rating. Since most new players are new to chess they will improve 100s of rating point's. On a large scale, this will decrease the average rating.
That's part of what made this interesting... as far as the site's average is concerned, improvement doesn't matter! If someone improves from 600 to 2600 and it takes them 100 games to do it or 10000 games the average stays the same (if the RD is basically the same for everyone).
I had players play at random, meaning after a while everyone has about the same number of games, same RD. (the maths details of Glicko were more complicated so I used a modified Elo, but same idea).
But yes, your individual rating can be pushed lower because someone else improves. This happens because the points have to come from somewhere. If someone improves from 600 to 2600 after their high RD period then the group as a whole has to pay them (so to speak) 2000 points, but because the group "pays" the average stays the same (same total number of players and same total number of points).
You're trying to represent people's current skill with their past results. There will always be some amount of error.
I had some fun trying to model a chess site's rating distribution with python... it also gave me incentive to learn python (I knew almost nothing before I started this).
Sample graph, 100,000 random users generated in a normal distribution.
---
I tried lots of different settings, but generally kept the pool of players below 10,000... because when that many players play a few 100 games each it takes a while to get your results.
I learned something interesting.
The average rating of a pool heavily depends on how quickly you can get new users to their correct rating... and I mean 0 games (initial rating is correct) is much better than requiring 5 or 10 games (for example).
Why?
Because after the new player burns through their RD, and everyone's RD starts to stabilize, then as far as the average is concerned, the player's real strength doesn't matter anymore. They can be a GM or a noob, but after the RD stabilizes the number of points they gain or lose in every game is roughly equal to the number of points their opponents gain and lose...
... which means at that point the average rating stays the same (with surprisingly little movement).
---
This means the average moves up when your high RD players are staying too high for too long.
And the average moves down when your high RD players are staying too low for too long.
I mean, it's obvious after you hear it, but I didn't realize it at first, and I didn't realize how important it is to quickly move a player to their correct rating... for example bad days, good days, widening or narrowing the rating range of opponents... it simply doesn't matter after players are gaining and losing roughly the same amount. The pool's average gets stuck in place.
I thought that was pretty interesting.
---
Sample output
At the bottom you see one of the new players was 2100 strength.
But his first rating on the site was 1950.
His highest rating was 2100, so that's good, he got to the correct rating.
His current rating is lower, but that's ok, I have it so people under and over preform.
He played 40 games, winning 26.
(Since the rating math doesn't distinguish between 2 draws and 1 win + 1 loss, I don't deal with them at all).
2 randomly selected players. One from the initial group and one from the joining group, and then:
2 random players selected, and then:
First is first rating when joining the site
Base is the player's true strength. Their results are based on this +/- an amount
Curr is current rating
High is highest rating
Lowest is lowest rating
Games is total number of games played
W is total number of wins for that player
L is total number of losses for that player
The rating is deflating for a reason that you did not mention, which is improvement. Let's say the average person starts at 600, but over the course of a year gains 300 rating point's. This will cause the average to go down, because the 300 rating point's came from other players, which decreases the average rating. Since most new players are new to chess they will improve 100s of rating point's. On a large scale, this will decrease the average rating.