Sagarin Ratings now "connected"

[MChambers]

I think this means that they are based solely on this season's performance. We're No. 1 in the overall ratings, but No. 2 in the pure points predictor (behind WVa). Another ACC team is No. 2 in the overall ratings.

http://www.kiva.net/~jsagarin/sports/cbsend.htm

[cspan37421]

Do you know how Sagarin does his ratings? I take seriously anyone who can do Bayesian statistics (wonder how many AP writers who vote can?) but the method is still an unknown to me.

In terms of the meaning of connected, I can only guess that it means that, through the transitive property, every team has played every other team, through however many degrees of separation that requires. Is that the case, or is it something else?

Of course, being superior by way of the transitive property doesn't really assure anything since strategy and outcomes often come back to exploiting matchup problems.

Are his Bayesian prior weights based on last season's results, or does he tweak them for freshman recruits and graduating seniors?

I knew of one system - prob. not Sagarin - which simply took 2/3rds of your own win % and 1/3 of your opponents' win %. Or something like that. Very simple shortcut rating system, not without merit, but probably not as reliable as Sagarin.

[darthur]

Being connected means that the Sagarin rankings no longer depend on previous years at all.

The other method you mentioned sounds similar to the RPI. I believe Sagarin is much more accurate.

I don't know the details either though - they are probably secret.

[hurleyfor3]

So they mean something. Well, more than they do when unconnected.

Duke is #1 Sagarin, but not #1 in either of the subpolls.

http://www.usatoday.com/sports/sagarin/bkt0708.htm

ACC is the #3 conference, behind the Pac 10 and the Big XII.

http://www.usatoday.com/sports/sagarin/bkc0708.htm

Duke is #3 Pomeroy, with unc... sixteenth?

http://www.kenpom.com/rate.php

[mus074]

Do you know how Sagarin does his ratings? I take seriously anyone who can do Bayesian statistics (wonder how many AP writers who vote can?) but the method is still an unknown to me.

In terms of the meaning of connected, I can only guess that it means that, through the transitive property, every team has played every other team, through however many degrees of separation that requires. Is that the case, or is it something else?

Of course, being superior by way of the transitive property doesn't really assure anything since strategy and outcomes often come back to exploiting matchup problems.

Are his Bayesian prior weights based on last season's results, or does he tweak them for freshman recruits and graduating seniors?

I knew of one system - prob. not Sagarin - which simply took 2/3rds of your own win % and 1/3 of your opponents' win %. Or something like that. Very simple shortcut rating system, not without merit, but probably not as reliable as Sagarin.

Sagarin's ratings have two components. The predictor component is basic linear progression based on margin of victory and level of competition. The Elo component borrows from Chess ratings, which rates you on your "best" wins, without regard to margin of victory - it is more PC in not rewarding running up the score. Neither component is very accurate and together they give us a very dull tool, good for seeing who is relatively good and who is relatively bad.

Pomeroy's ratings actually think about the game of basketball and model its ratings based upon the way teh game is played. To my mind, it is far superior, but still not perfect in identifying which team, among those in the top 20 or so, would win a given contest. Pomeroy's ratings give you a good look at how a team would fair in a season conference play against the average conference slate. So team-by-team comparisons for one game are not perfect. However, it would likely do very well in predicting the level of success in a 16-game or so season.

[gw67]

mus074 - I've followed both Sagarin and Pomeroy ratings for several years. Before I would say that one or the other is superior in predicting results, I would compare results from a large sampling of games. I am always surprised that the RPI, Sagarin, Pomeroy and other ranking systems just prior to the NCAA tourney are as close as they are in ranking teams.

gw67

[cspan37421]

Sagarin's ratings have two components. The predictor component is basic linear progression based on margin of victory and level of competition.

Do you mean simple linear regression? Just curious.

Do you know how Pomeroy's rankings are done? I'd be interested in a general summary of the method. I don't know why, but I find this stuff interesting. Who's #1 (or #10) and why, that kind of thing.

[gw67]

Ken Pomeroy explains his ratings on his site:

http://kenpom.com/rate.php

The general explanation is provided by clicking on 'Read this!' and the efficiencies are described on his Stats page.

gw67

[Clipsfan]

Another interesting note on Pomeroy's ratings is that Wisconsin (a team we handled fairly easily) is rated just behind us at #4 (despite their two losses).

[mus074]

mus074 - I've followed both Sagarin and Pomeroy ratings for several years. Before I would say that one or the other is superior in predicting results, I would compare results from a large sampling of games. I am always surprised that the RPI, Sagarin, Pomeroy and other ranking systems just prior to the NCAA tourney are as close as they are in ranking teams.

gw67

The proof is always in the pudding. I just find the theory behind Pomeroy so much more satisfying. That said, I think there has been a general acceptance that Pomeroy better identifies likely FF success.

[mus074]

Do you mean simple linear regression? Just curious.


Thanks for the clarification. :)

[mus074]

Do you know how Pomeroy's rankings are done? I'd be interested in a general summary of the method. I don't know why, but I find this stuff interesting. Who's #1 (or #10) and why, that kind of thing.

Another interesting note is the weighting of recent games in Pomeroy. I believe it is a geometric weighting, with the 10th game worth 10 times the 1st game (weighting = 10/(1+2+3+4+5+6+7+8+9+10)), the thirtieth game worth 30 times game one and 3 times game 10. It puts a real premium on what you have done lately, especially by the end of the season.

I don't know how/if Sagarin weights recency, but I doubt its this strongly.