Feels like competition is harder at a given rating now.

Sort:
Paleobotanical
Duckfest wrote:

I’ve been looking at the back and forth of the whole rating change. I’ve done a simple calculation to get some feeling for the effect of some changes in the players pool.

 

I love that you've done this!  One thing to take into account, though, is that new players gain or drop points a LOT more easily than long-time opponents, due to how Glicko handles a parameter called "rating deviation."

More info here: https://en.wikipedia.org/wiki/Glicko_rating_system

So, points aren't conserved in matches that include a player whose rating is not-well-known and an influx of overrated new players does not tend to inflate the system directly due to the overrating.  The activity of active but bad players, on the other hand, does tend to add rating points into the system as they bounce against the rating floor.

Points behave a lot more like a conserved resource among players whose ratings are generally well-known by the system (thus they have low RD values.)

Duckfest
Paleobotanical wrote:
Duckfest wrote:

I’ve been looking at the back and forth of the whole rating change. I’ve done a simple calculation to get some feeling for the effect of some changes in the players pool.

 

I love that you've done this!  One thing to take into account, though, is that new players gain or drop points a LOT more easily than long-time opponents, due to how Glicko handles a parameter called "rating deviation."

More info here: https://en.wikipedia.org/wiki/Glicko_rating_system

So, points aren't conserved in matches that include a player whose rating is not-well-known and an influx of overrated new players does not tend to inflate the system directly due to the overrating.  The activity of active but bad players, on the other hand, does tend to add rating points into the system as they bounce against the rating floor.

Points behave a lot more like a conserved resource among players whose ratings are generally well-known by the system (thus they have low RD values.)

Is the simplest interpretation of this mechanism some thing like :

ELO gained by old_player is less than ELO lost by new_player_with_inflated_rating?

 

Paleobotanical
Duckfest wrote:

Is the simplest interpretation of this mechanism some thing like :

ELO gained by old_player is less than ELO lost by new_player_with_inflated_rating?

 

 

Yes, the amount gained or lost by the old player is reduced while the amount gained or lost by the new player is greatly increased.

Duckfest

There is also the possibility that it’s not macro factors at work here. Of course, it’s safe to assume a direct correlation between accuracy and ELO rating, especially in the long run. 

But with smaller sample sizes, it might be more nuanced. I find it difficult to say how many games are needed to determine whether the pattern you observed is happening. It may be true that your accuracy was higher from September till now compared to April to August. But only by 1%. Your average opponent was 949 from 1 April to 1 Aug (or 951 from April 1 to 31 Aug), compared to 903 since 1 September (only checked for white though). 

If your rating is 925, your win rate against  a 950 is 46% and against a 900 is 57%, suggesting you would win 11 more games out of a 100 games. That's a big difference.  But if you are an 800 or a 1000 against these opponents, the difference is only 6 or 7 games out of 100. Let’s assume it’s not random fluctuation, and the difference is significant.

 

Changing your play might also change your accuracy and rating in other ways. From a theoretical standpoint there are multiple ways in which higher accuracy would not translate into higher rating (i.e. more winning), or just very little. Two examples I can think of:

  • Defending lost positions. Fighting for every game, no matter how lost you are. If you challenge yourself to find the best move in a position, even when all moves are completely lost, is a good way to boost your accuracy, while you might only be able to save 1 in every 10 dead lost positions. If you get a dead lost position once every 10 games, your win rate will only increase by 1% (10 losing games in every 100, of which you can save only 1). The inverse is also true. Losing focus because your game is lost anyway, tanks your accuracy without affecting your win rate.
  • Playing sub-optimal moves that are terrible against engines, usually also terrible against good players, can be very effective against human opponents. For example:
    • I’ve won many games in very much losing positions, especially when behind in material a lot, by switching to risky play. Instead of slowly grinding to a lost endgame, I focus on the only mate threat I see. Miracles can happen. 
    • Also structurally, the decision to play gambits or traps vs solid move will impact the relation between your accuracy and your rating. With black I play the Caro-Kann vs e4 and used to play the Englund vs d4. My Caro-Kann is much more accurate (citation needed), but my Englund is much more winning.
    • Playing totally random, non-book, games tend to be much less accurate then games where both players follow generally accepted opening lines. My experience when I started getting into chess more was that I stopped making stupid opening mistakes and played more main lines. Which meant I also played more into lines my opponents were familiar with.

What are your thoughts on this?

Paleobotanical
Duckfest wrote:

What are your thoughts on this?

 

Thank you for taking the time to share your thoughts!  I do agree that there are plenty of confounding factors (and all the things you list are important considerations when contemplating my study of the game.)

Particularly your comments about switching to risky and suboptimal play when things get bad.  I do that quite a bit compared to my peers, according to that other meta-analysis site I use.  It has certainly pulled wins or draws from the jaws of defeat (like in this game, where I deliberately offered up a juicy capture to achieve a stalemate.)

 

 

 

Paleobotanical
@CooloutAC: “I think you are totally ignoring FIDE and their avg rating.”

I said this earlier, but you don’t seem to get it: FIDE’s average being 1400 literally tells you nothing. The numbers are relative.

Wikipedia’s article on the Elo system specifically discusses the kind of rating inflation and deflation we are discussing here and uses FIDE as an example.

https://en.wikipedia.org/wiki/Elo_rating_system#Ratings_inflation_and_deflation
Paleobotanical
Wiki discusses the "rating floor"  which I was saying will be much higher when more players are rated higher,  Not lower as you are implying.  

And what about the fact the average chess.com rating use to be much higher,  at 1200 at one point and 1000 after that,  and now at 800.  Do you think its because less new people are playing now according to your logic?   Because apparently its quite the opposite.

 

I'm not sure what you're saying with the first one.  The rating floor on Chess.com is fixed at 100 and hasn't changed.  When that article talks (or I talk) about inflation or deflation, we're talking about numerical ratings changing WITHOUT underlying play quality changing.  The article talks about a high rating floor generally causing rating inflation, which happens because players who are misrated above the rating floor but whose skill is below it tend to feed extra points into the rating system.  Chess.com's floor of 100 is low enough that this is a small effect over time, but in the 70s, FIDE's rating floor (currently 1000) was 2200!!  That's definitely high enough to cause problems.

Also, you keep bringing up FIDE's average rating of 1400, but remember that everyone under 1000 is dropped from the rating list.  It's literally impossible for FIDE to have an average rating of 800 even when the patzers invade, because a FIDE rating of 800 simply doesn't exist according to the rules.  All those hypothetical beginners playing in opens would be simply "unrated" and not counted in the statistics.

Regarding chess.com's distribution:  First, keep in mind that (as I've been saying) the absolute number doesn't mean anything.  Chess.com actively manages their rating system to try to keep it somewhere in the ballpark of other similar chess rating systems, and to try to keep it somewhat comparable across game types (which is why, in September of 2020, they just straight-up added 150 points to everyone's Bullet rating.)

Also, though, the average rating doesn't tell you much about what's going on with the population.  The population's actuall chess skill distribution can stay right where it is and average rating can drop due to rating deflation that's an artifact of the rating system or how it's managed.  Or, average rating can drop when there's no deflation or inflation if there's a wave of incoming players at the bottom of the skill range who aren't matched at the top.  However, and this is important, that event may cause deflation or inflation down the line even if it doesn't cause it on the day it happens.

I personally believe that those numbers suspiciously mirror the site's changes in starting rating.  What if the average active player logs on to chess.com and plays five games then never returns?  Starting them at 1200 rating vs. letting them pick from a range where they call themselves a "beginner" and start at 800 may start to make a big difference to the average "active player" number, especially if "active players" are dominated by people who are just passing through, so to speak.

My original speculation was about whether there's been modest ranking deflation in the last half-year to a year.  As that Elo article points out, one source of systematic deflation is when new players come in and start improving fast enough to play above their rating, which is very possibly an after-effect of the Queen's Gambit wave.  This is less a problem with Glicko (used on chess.com) but it's not gone.

The deflationary mechanism I was suggesting above is a little different:  basically there are people out there who are sitting at the rating floor of 100 and still playing.  Because more modern rating systems like Glicko self-balance so that a certain rating difference corresponds to a win likelihood, changes in the play quality at that rating of 100 would, if the system were working optimally, result in changes of all the ratings' play qualities.

As an example:  take a player who's at an actual skill level corresponding to a rating of 100 (by this I mean the floor isn't artificially inflating their rating.)  In the Elo system (used for convenience, but Glicko works similarly), their chance of taking a win off a player who's 100 points higher, at 200, is about 36%.  That 200-rated player's chance of taking a win off a 500-rated player is about 15%.

But, let's say that we lose 3/4 of our players across all ratings.  There will be fewer people at the long tails of the distribution:  the very low and very high end.  And, that 200 rated player now just can't find an even match as often.  They'll be matched more than half the time against higher-rated players and their rating will drop because they lose more than they win.  If they hit the floor of 100, they'll stop there, but since their skill hasn't changed, that 500-rated player from before, whom they can beat about 15% of the time, will now drop to 400, because it's that difference (not the actual number) that's predictive of win rates.

Anyway, in my original comment I did two things:  I put forth some evidence (average accuracy numbers across large numbers of games for two periods with different ratings) that there's been modest rating deflation over the last six months.  I also speculated where I thought such deflation might have come from.  There are lots of moving parts and reasons deflation can occur, so my speculation is just an idea, but it's at least rooted in how the rating system works

Paleobotanical
CooloutAC wrote:

Again you are saying the rating goes down with influx of new players which I agree was your original comment.  But yet you now disagree the average rating will go up when new or low rated players then leave or stop coming?  I'm still not understanding the logic there bud.

 

An influx of new, bad players will lower the overall distribution at the start, while those players get Glicko's special treatment for being new and their ratings are dropping rapidly.  But, the distribution changing does not mean that an individual player's rating moves the same direction.  In fact, the distribution as a whole can move downward while existing individual players' ratings stay the same, or go up.  Those are two entirely separate things.

Put it another way:  Depending on why it's happening, the average going up or down may or may not mean your own personal rating goes up or down.

However, an enormous influx of new players normally will do two things:  (1) lower the aggregate rating of the distribution overall and (2) drive existing players' individual ratings up or down depending on exactly how fast the new, bad players are either learning to play (which would cause deflation) or churning (which would pump points into the existing players, causing inflation.)

There are lots of moving parts and a number of possible scenarios with different results, both aggregate and individual.

Paleobotanical
CooloutAC wrote:

I thought lichess uses glicko and chess.com uses elo.  

And again, I'm not disagreeing with any of that and it doesn't need repeating.   The question now is why you don't think the opposite is also true.  When lots of new players then stop playing that the avg goes back up.

 

Chess.com uses Glicko, lichess uses the updated Glicko-2, FIDE uses Elo, USCF uses a hybrid Glicko-Elo system that Mark Glickman (creator of Glicko and Glicko-2) helped create, which is documented here.

I'm not saying the average rating may not go up if lots of new players leave, depending on who leaves and how that changes the distribution.  I am saying that individuals left behind are more likely to see their personal scores decrease when low-rated new players leave.  The reason is because play at or around the rating floor creates upward pressure on all the ratings above, and new players leaving will tend to relieve that pressure.

Paleobotanical
CooloutAC wrote:

So you are saying they decrease when new players join, and also decrease when they leave?   So players coming in going have no effect on inflation,  only deflation?   I'm not understanding the logic there.

 

No, I never said the high rated players' ratings decrease when new players come in.  The distribution can move lower while individuals stay the same or move higher (because the distribution shifts when you add more people at the bottom.)

Paleobotanical
CooloutAC wrote:

First of all we are talking about the average rating. 

 

What?  I'm not talking about average rating at all.  Read my first post, it's all about my individual rating.  I'm talking about whether other individual existing players feel they're moving down or up, not what's happening to the distribution or the average (which could be going the opposite direction, depending on who's leaving.)

Paleobotanical
CooloutAC wrote:

I would say the way you judge if the playerbase is affected,  is by the average rating.   I guess I was wrong to assume that as common sense.  Maybe i'm wrong.   lol

 

Yeah, ok.  Here's an example.  Let's say you have a distribution with two groups, low rated and high rated.  Low rated have ratings 1-5 and high rated have ratings 10-14.  The average is 7.7.  Here's a plot of how many people are at each rating:

 

Now, let's say all the high players move down one rating point.  The person at 10 goes to 9, person at 14 goes to 13.  And, all the low players leave.  Then you get this.  Even though all those people moved down in rating, the average went up from 7.7 to 11:

 

That's a trivial example showing that the way the average moves doesn't tell you how individual ratings move.

Duckfest
CooloutAC wrote:

The first graph is not based in any kind of reality though.  Again you are making this too complicated and getting lost in the weeds bud.

 

I'm confounded by this exchange. It has a semblance of an interesting conversation. I'm also amused.

@Paleobotanical , can you please explain why you would use graphs that are not based in any kind of reality? Also, why would you use it at all? Apparently, explaining an abstract concept with a simplified visual representation is making it too complicated . It seems you are getting lost in the weeds. GL explaining this.

Paleobotanical
Duckfest wrote:

@Paleobotanical , can you please explain why you would use graphs that are not based in any kind of reality? Also, why would you use it at all? Apparently, explaining an abstract concept with a simplified visual representation is making it too complicated . It seems you are getting lost in the weeds. GL explaining this.

 

Sure.  He says: "I would say the way you judge if the playerbase is affected, is by the average rating. I guess I was wrong to assume that as common sense. Maybe i'm wrong. lol"

The charts make the purely mathematical point that if a bunch of people drop out at the low end, the average can move up due to their leaving at the very same time that the ratings of everyone who remains drop (or stay the same, or increase.)  In other words, the average doesn't tell you what's happening to any particular subgroup of the whole, even if that subgroup is "everyone who hasn't quit."

In other words, if "common sense" tells you that the average is a good proxy for what individuals are experiencing with their own rating movement, either individually or in the aggregate, you can't necessarily trust it.

All this, by the way, is a response to "no the average is moving X direction so people's ratings must be moving that direction too."  That would be true only if nobody could join or leave.

(BTW I have to say I'm really baffled by this idea that ratings, which are entirely an artifact of modern statistics, should behave completely intuitively and not require any mathematical concepts at all to understand and discuss.  Why would you ever think that were the case?)

Paleobotanical

Revisiting this thread after a while, it looks like I may just have been out of practice.  Here's what happened:

The first couple weeks I was playing a lot post-break, my rating bounced around near 900 (from my April peak of 1014.)  Then, around the third week, when I wrote this, I started doing some focused tactics study alongside still playing games, and my rating tanked to a low of about 835.

Then, over the last 7 days, I've had 39 rapid games with a win rate of about 66%, carrying my rating up to 988 (as I write this.)

Through all of this, my accuracy score hasn't really budged, but my rate of blunders, as measured by that 3rd-party meta-analysis site I use, has oscillated with my rating and is now at a low since I started tracking it.

All this leads me to believe:

1)  At least at my rating, accuracy is not a great predictor of game outcomes (which kind of makes sense, since mistakes and blunders still dominate game outcomes and they're a minority of all moves.)

2)  Measured blunder rate is a much better predictor of game outcomes for me.

3)  Returning from a long break, my chess skill, such as it is, had somewhat lapsed due to neglect and took a while to kick back in.

I've seriously found myself wondering this week whether all the underrated players are off watching the world championship instead of playing online.  I guess it's possible!  But, blunder count per game is closer to an absolute measure that tells me that I really am doing better than I was a few weeks ago.

It's interesting to me that it's so easy to get out of practice, but also that practicing more can help my focus return.

Paleobotanical
CooloutAC wrote:

that happens to all of us.  And I think has to do with the way they default mmr is chess.com.

 

Thanks.  Looking over my opponents, most of them have hundreds of rapid games played and seem to play reasonably often, so I don't think default ratings are having much effect.  Less sure about stale ratings (but then a bunch of people were getting free wins from me when my rating was stale, so it works both ways.)

Paleobotanical
CooloutAC wrote:

Something is very different about both sites thats for sure.

 

Chess.com uses the Glicko rating system while Lichess uses Glicko-2.  The primary difference between these systems is that, if you take a break or become inactive for a while, Glicko-2 assumes increasingly greater uncertainty about the rating, while Glicko does not.

This means that games against people with stale ratings or those who play infrequently will tend to move your score less (and theirs more).  I could totally imagine this making Glicko-2 ratings significantly more stable over the long run.  If Glicko-2 were in use on chess.com, I would have seen a much steeper rating drop when I arrived and started doing badly, and probably would have recovered more quickly too.

There might also be differences in active management of the scoring systems on the two sites that affect this.

justbefair
CooloutAC wrote:
Paleobotanical wrote:
CooloutAC wrote:

that happens to all of us.  And I think has to do with the way they default mmr is chess.com.

 

Thanks.  Looking over my opponents, most of them have hundreds of rapid games played and seem to play reasonably often, so I don't think default ratings are having much effect.  Less sure about stale ratings (but then a bunch of people were getting free wins from me when my rating was stale, so it works both ways.)

I'll take your word for it.  so far i been playing 100s of games on lichess and my rating has always been between 950 and 1050 since day 1.   On chess.com i been up and down 400 points and go on very long losing and winning streaks.  It just happens too often for me to think its natural.  yes I do go on long tilts and just start resigning games out of frustration I do it on lichess too,  but even still my rating and the matches seem very stable on lichess.  Something is very different about both sites thats for sure.

Your rating here on Chess.com has been fairly stable too-  mainly staying within 100 points of 400 since a few weeks after you got started.

/ Except for a period in late October when it dropped to 100, the minimum rating possible.   

Paleobotanical
CooloutAC wrote:

because of the amount of blunders which again can be based on your opponents and not necessary your own level of play.

 

Possible in principle, but I know the stuff I'm blundering and while my win rate and blunder rate have swung wildly over the last three weeks, my opponents aren't really doing anything more clever.  (I'm just remembering better what I've lost to recently and looking for similar things more carefully.)

happy140411

Yes but if you get more rating you will be more famous.