Hacker News Reader

> Far far below the lofty heights of Carlsen &co., I play a fair bit on Lichess. Sometimes I'll play someone with a similar ranking to myself and beat them and will earn 3 or 4 points. Then we'll have a re-match which I lose and they'll gain 12 or 15 points. I just can't fathom how them beating me earns them more points than I earn from beating them --if we were on the same or similar rankings to start with. But it almost always seems to pan out this way.

Lichess works on a different system than professional ratings. It takes into account the RD (sort of a standard deviation on your rating). Playing against a player with a more stable rating will cause the rating to change more and vice-versa. Professional ratings are always assumed to be stable.

> I've also noticed that there doesn't seem to be any relation between the circumstances of a win and the points awarded. For example, shouldn't a win or draw when playing black earn more points than when playing white --given that white had the advantage of moving first?

This is something that is sometimes used as a tiebreaker in tournaments. Sort of an away goals rule, if you're familiar with soccer. But as far as ratings go, in the long run you're bound to play ~50% of the games with white and ~50% with black, so it smooths out itself.

> And what about the manner of victory? It seems there's no more ranking points to be gained from; snatching a victory where both players ended up down to a single pawn each and the winner just managed to queen one square ahead of the loser... than there are from completely decimating your opponent.

I'm sorry but this makes absolutely no sense. There are three results: Win - Loss - Draw, everything else is just discourse around the game.

Endgames are something players trade into, reducing to a won endgame is part of the skill involved in the game. Why would that count as less of a win?

Also, material is a pretty stupid way to determine how close a game was. Should I get less rating points because I won with a queen sac?

Lichess doesn't use ELO rating but a different rating system. And yes, I don't understand that difference either.

About white versus black, there is some consensus that after about 20 moves, white's advantage is gone. I doubt that statement is supported by statistics. I guess the rating just assumes everyone plays white and black just as often, so it will even out in the end. (edit) By the way, if you are strong with white and weak with black, how should your ELO rating be calculated? It is only one number, not two. Assuming an average seems fine with me.

And the way a win is made, those are for the beauty contests :)

A simple calculation for ELO is just to see it as both players placing a bet, the K factor is the amount of points in the jar. If the K factor is 25, you have 1600 ELO and I have 1400 ELO, you might put in the jar 16 points, while I put in 9 points. Winner takes the jar, on a draw it is plit in half. Difference between win and draw is always half the K factor. When both players have a different K factor, the calculation for both players will be different, it's not zero sum.

> About white versus black, there is some consensus that after about 20 moves, white's advantage is gone. I doubt that statement is supported by statistics.

That statement is supported by statistics, however, the problem with that statement is that the average chess game is 25 moves.

except in top-level classical play.

But it's not relevant, because virtually all experts agree that chess is a theoretical draw. It's extremely difficult to prove it, however.

I wonder if there's even a way to be sure of a minimum length? Like suppose I'm White, I know how to play Chess perfectly and so does my opponent. So (if the popular assumption is correct) we will eventually draw, but how many moves can I draw the game out for, assuming my opponent isn't willing to concede, but is trying to shorten the game ?

> but how many moves can I draw the game out for

depends on the position, from not a lot, to 100+ moves.

but don't forget that there are forced draws in chess, namely, repeated position 3 times (doesn't have to be consecutive), 50 move rule (50 moves without a pawn move or a capture), and insufficient material.

So trading off material would shorten the game.

Lichess doesn't use ELO rating but a different rating system.

That explains some odditites I've observed, like different strength with the same rating at different hours of the day or different days.

Yes!!! I've long been frustrated by this. I'm around 1800 if I play at certain times, but quickly get down to 1700 when playing at others. For some reason 1700s are stronger during night.

And something even worse: sometimes it seems that lower rated players are consistently stronger, so if you fall below some threshold, you get stuck there. On the other hand, if you can get above it, you suddenly win and win...

Lichess' page on rating systems: https://lichess.org/page/rating-systems

They say that Glicko 2 is more accurate than Glicko 1, and they reference a kaggle leaderboard. On that private leaderboard, I only see "Glicko benchmark" and "Real Glicko," so I don't see how they made that claim. I find it dubious that Glicko 2 would be more accurate out of sample.

> Sometimes I'll play someone with a similar ranking to myself and beat them and > will earn 3 or 4 points. Then we'll have a re-match which I lose and they'll > gain 12 or 15 points.

It's probably a sign that they have a new account. If a player has played very few games, then his rating is only approximate. It's likely his true rating is far from his current rating. The rating added or subtracted due to a win/loss is on purpose made large, so he quickly can get to his true rating. When he gets to his true rating, he will presumably win and lose the same amount of games and stay at the rating.

Lichess and Chess.com both use (variants of) Glicko (instead of ELO). The first step is to calculate the rating deviance. A high deviance means that the player hasn't settled on a given rating. This factors into the calculation of the new rating.

The sites have chosen this system so that higher (or lower) rated players that make a new account quickly get close to their "true" rating.

The Glicko[0] rating system is different from Elo[1], and it's what Lichess and Chess.com use

In Elo, the rating gain for one player equals the rating loss for another, because each player's rating is the only independent variable.

Glicko calculates a kind of 'certainty', aka 'rating deviation' which is based on the number of games played in some period of time. The idea is that the ratings of players who have not played in a while are more likely to be wrong.

[0] https://en.wikipedia.org/wiki/Elo_rating_system

[1] https://en.wikipedia.org/wiki/Glicko_rating_system

About the manner of victory, I make an example from the game of go, which being points based makes this case more straightforward to evaluate. In go it's either win or loss, there is no draw. You can win by one point or by one hundred and you still score only one point in the tournament table. Why is that?

Because if you count to be losing by five points you could either resign (I'm assuming a game between strong players) or start playing more adventurous moves to turn the game around. Adventurous moves usually make for a larger win, let's say fifty points. Should players be punished for that and only play games that minimize the difference between the scores of the players? How boring. I guess this applies also to chess. Sacrifice, sacrifice, win. Or sacrifice, sacrifice, mistake, loss. I won't punish a mistake so much more than any other one.

In Baduk gambling in Korea, you pay for points, so you might expect people to just minimize differences to pay out less, but they end up fighting all game instead. This is because it's capped for resignation, it might be cheaper to resign if you're too far behind (why count and pay more?)

As a note, baduk is the name of go in Korea.

On chess websites gains and losses swing more extreme until you play enough games to stabilise. It's a modified elo system or elo-adjascent. This way people reach their "true" rating much quicker e.g. a grand master who signs up will reach 2400 over a smaller amount of games than if he could only gain 2 or 3 points for every game. The same goes for the beginner who starts at 1500 but should be 800, he'll reach that quickly and then play against people around his ability.

If you have roughly the same rating both of you should be gaining the same amount on a win. However many rating systems have some kind of aging factor. Meaning the more games you play the more stable your rating gets.

So it could be that your opponent has played significantly less games than you. However if he loses he also should lose more points than you would.

The color you play has no influence on the rating numbers. And also the type of win/draw/loss does not matter at all. Ratings are just statistics of your results.

Elo also depends on how many games you have played in a time format.

It is likely that the person with 12/15 point gains has less games, thus his delta variation is higher than yours. It takes tens of games before you take "normal" points.