Anchors

[daal]

... and how come they never seem to get adjusted up?

Looking at the KGS archive graphs for some of the most active players seems to show that the similar overnight adjustment at the end of last December *was* up.

Thanks for pointing this out. Where can one see these archived graphs?

[ez4u]

... and how come they never seem to get adjusted up?

Looking at the KGS archive graphs for some of the most active players seems to show that the similar overnight adjustment at the end of last December *was* up.

Thanks for pointing this out. Where can one see these archived graphs?

The KGS archives page. The bottom box is for the graphs. Try 'TheCaptain'.

[Mef]

Counter-example: I did not play any KGS games in November, but my rank magically jumped from 13k to 8k! Of course, then I played a few games and quickly dropped back to 10k But there was definitely some sort of odd drift there, maybe that was anchoring? Our just the players I beat getting better?

http://www.gokgs.com/graphPage.jsp?user=sligocki
http://www.gokgs.com/gameArchives.jsp?u ... 3&month=10
http://www.gokgs.com/gameArchives.jsp?u ... 3&month=12

This isn't technically an "Anchoring" issue, though in your case it might seem that way. Your first loss was to a person who was ranked [15k?] when you played them, but later it turned out they were actually a 5k. As a result, your rating calculation went from being "Two wins and a loss rated at 15k" to being "Two wins rated at 15k and a loss rated at 5k". Basically the only upper bound the rating system had was that 5k player, so it gave you an estimate of [8k?].

[badukJr]

KGS has been up adjusted before. Once I was around 20k and got up adjusted to about 9k, it was insane. I was never teen kyu.

[SmoothOper]

Something has to be said for the transparency of a point based rating system, though it sounds like the anchoring is a step in the right direction. When I log into IGS, there are X number of people of above me and if I win Y more than I lose, then I go up a rank(or lose Z more than I win, I go down a rank).

[hyperpape]

Yeah, that's one of the major tradeoffs rating systems have to deal with. Transparency is great, and is the most important thing for some people, but all the transparent systems we know about lose some accuracy.

[Kirby]

Yeah, that's one of the major tradeoffs rating systems have to deal with. Transparency is great, and is the most important thing for some people, but all the transparent systems we know about lose some accuracy.

For some definitions of accurate

[skydyr]

Yeah, that's one of the major tradeoffs rating systems have to deal with. Transparency is great, and is the most important thing for some people, but all the transparent systems we know about lose some accuracy.

For some definitions of accurate

The real question is 'what is accuracy and how does that compare to the goals of the rating system?'

[Kirby]

The real question is 'what is accuracy and how does that compare to the goals of the rating system?'

Sure. From a personal perspective, the goals that I have for a rating system to have is for games against players of the same rank to be enjoyable. But what is enjoyable can also vary from person to person, so even if someone shares this goal, they might have different rating system preferences.

[oren]

Something has to be said for the transparency of a point based rating system, though it sounds like the anchoring is a step in the right direction. When I log into IGS, there are X number of people of above me and if I win Y more than I lose, then I go up a rank(or lose Z more than I win, I go down a rank).

IGS requires a bit of honesty to make it work. If everybody joined stating they were a high dan, it would be broken compared to other ranking systems fairly quickly.

[Mef]

Yeah, that's one of the major tradeoffs rating systems have to deal with. Transparency is great, and is the most important thing for some people, but all the transparent systems we know about lose some accuracy.

For some definitions of accurate

The real question is 'what is accuracy and how does that compare to the goals of the rating system?'

Generally for a rating system, accuracy is the ability to correctly predict the outcome of a game between two players. To that end KGS does very well, especially compared to most other systems.

As far as the goals of a rating system, accuracy is a primary goal of KGS's and not necessarily that important to some others.

[Kirby]

Generally for a rating system, accuracy is the ability to correctly predict the outcome of a game between two players. To that end KGS does very well, especially compared to most other systems.

KGS does well, assuming that someone has an unchanging rank. If someone is 5k, and they're going to be 5k forever, then the maximum likelihood approach correctly uses history to estimate a rank. If someone is improving, KGS does not do so well compared to some of the other online systems, because past history makes it such that playing a lot of games makes it harder to adjust your rank, even if you have genuinely improved.

While it's true that KGS can do well with players that are not improving (or declining in skill), it is misleading to say that it universally does a good job of correctly predicting game outcome, when other systems that are not weighted so heavily by the history of past games do better in some circumstances.

Case in point: If KGS were accurate, people would not be inclined to create new accounts just to get a correct rank. We see this happen on a regular basis.

[UnclMartin]

When the drop in ratings was reported, the super admins checked the anchors and found one of them had stopped playing. The anchors have been adjusted, effective December 26.

[daal]

When the drop in ratings was reported, the super admins checked the anchors and found one of them had stopped playing. The anchors have been adjusted, effective December 26.

Why wasn't the adjustment done so that drop in ratings got corrected? Just askin...

[DrStraw]

When the drop in ratings was reported, the super admins checked the anchors and found one of them had stopped playing. The anchors have been adjusted, effective December 26.

Why wasn't the adjustment done so that drop in ratings got corrected? Just askin...

Why bother? Ratings are relative anyway. The same handicaps will be generated. No need to go to a lot of extra work.