Barry's Glicko Calculators (was Question For Barry Cox)

[Kevin Bonham]

Originally Posted by drug

That means for ratings 800-1100 the function is monotonic but the other way around. The higher is your original rating, the lower is your final rating.

That seems to bottom out at an initial rating of about 1150 and then pick up again for higher initial ratings.

I don't know if this can also happen under the current Glicko-2 but I think there is actually some kind of argument for it even if it can. That is that the greater the disparity between the original rating and the observed performance, the stronger the statistical evidence becomes that the original rating is simply total junk, as opposed to (say) a player with erratic runs of form having a good few months.

However an 1100! who gets 50% from 30 games vs 1400!s should go closer to 1400 than 1354. I expect that under G2 which is more dynamic they would go closer (maybe 1375 or so).

Next I changed ! for the player X to be an empty space and I got
f(800)=1765.
So we have 800 player who drew 30 times a 1400 player and as a result becomes 1700 player.

I had an overshoot like this under the old ELO system where because of a backlog of dozens of tournaments being processed at once in a period in which I had improved, I went to a rating that was about 60 points higher than my PR for the tournaments. However I was able to maintain that rating from that point.

I think overshoot/undershoot protection should be applied in such cases generally. Not sure if it is at the moment or not. Perhaps there are some cases though, with fast improving juniors, where an overshoot could be more predictive of future results than the player's actual performance rating for the period.

[Bill Gletsos]

Originally Posted by Kevin Bonham

I think overshoot/undershoot protection should be applied in such cases generally. Not sure if it is at the moment or not.

No players rating can change such that their new rating for the period exceeds or drops below their performance rating for the period.

[Vlad]

Sorry Bill, are you saying that Barry's calculator has an error?

[Rincewind]

Originally Posted by drug

Sorry Bill, are you saying that Barry's calculator has an error?

My calculator does a straight Glicko-1 calculation and (as far as I am aware) works as per its design. The ACF rating system is Glicko-2 with a number of tweaks and so my calculator is only an estimate of what might happen.

[Kevin Bonham]

Originally Posted by Bill Gletsos

No players rating can change such that their new rating for the period exceeds or drops below their performance rating for the period.

That is the sort of protection I had in mind.

[Bill Gletsos]

Originally Posted by drug

Sorry Bill, are you saying that Barry's calculator has an error?

Not at all. I'm just pointing out the ACF system has additional features to the standard Glicko system.
Barry's calculator does what standard Glicko-1 does. However even back when the ACF was running Glicko-1 and not Glicko-2 the ACF system the overshoot/undershoot protection in place.

[Vlad]

Originally Posted by Bill Gletsos

Not at all. I'm just pointing out the ACF system has additional features to the standard Glicko system.
Barry's calculator does what standard Glicko-1 does. However even back when the ACF was running Glicko-1 and not Glicko-2 the ACF system the overshoot/undershoot protection in place.

Thanks for the reply. Regarding the non-monotonicity, is it different for Glicko-2?

The way I thought the system is working is that your new rating is a weighted sum of your current rating (your performances in the past) and your performance in the current period. If your current rating is higher then the new rating has to be higher as well. I do not buy the argument that if you can perform significantly better than your current rating then your former performances should not be taken into account.

[Bill Gletsos]

Originally Posted by drug

Thanks for the reply. Regarding the non-monotonicity, is it different for Glicko-2?

I believe when I looked at it ages years ago that it was similar to Glicko-1. BTW we also do the overshoot/undershoot protection in the current ACF Glicko-2 system.

Originally Posted by drug

The way I thought the system is working is that your new rating is a weighted sum of your current rating (your performances in the past) and your performance in the current period. If your current rating is higher then the new rating has to be higher as well. I do not buy the argument that if you can perform significantly better than your current rating then your former performances should not be taken into account.

If your performance rating in the period is higher than your current rating then your new rating should be higher than your current rating and if your performance rating in the period is lower than your current rating then your new rating should be lower than your current rating.

[Vlad]

Originally Posted by Bill Gletsos

If your performance rating in the period is higher than your current rating then your new rating should be higher than your current rating and if your performance rating in the period is lower than your current rating then your new rating should be lower than your current rating.

Sorry Bill, by non-monotonicity I mean the fact that
f(800)=1397,
f(900)=1389,
f(1000)=1369,
f(1100)=1354.
Your higher current rating results in lower new rating.

[Bill Gletsos]

Originally Posted by drug

Sorry Bill, by non-monotonicity I mean the fact that
f(800)=1397,
f(900)=1389,
f(1000)=1369,
f(1100)=1354.
Your higher current rating results in lower new rating.

Only to a certain point.

Using the settings you described in your original post then it stops getting lower at 1148 and starts increasing again.

Using Barry's Calculator you get

f(1148) = 1352
f(1200) = 1354
f(1300) = 1372
f(1350) = 1385
f(1400) = 1400

[Vlad]

Thanks Bill. Yes, I understand what it is doing. I am just saying that I do not think that it makes sense.
f(800)=1397,
f(1100)=1354,
f(1390)=1397.

Why is 800 player able to jump as high as 1390 player after exactly the same performance?

[Rincewind]

June 2008 ratings have been loaded into the calculator.

We now return you to your normal programme.

[Rincewind]

After several version updates of mail and web servers as well as a couple of releases upgrade on the CMS my website is hopefully again operational.

http://www.bjcox.com/modules.php?name=Glicko_Calc

Should still work. If you have time try it out and let me know if there are issues. (PS I know the ratings are still the December 2008 ones).

[Rincewind]

I migrated the webserver to new hardware today. Seems to work but if you notice anything funny happening, let me know.

[Rincewind]

The website will be going offline Wednesday 23 June, 2010 and will not be coming online again until some time in July (hopefully before the 15th). In the interim I will be moving the last few rating lists and some software downloads onto a temporary site in Google Sites.

To see what the temporary site will look like go to the temporary website.