Author Message
Jose A. Martínez
Joined: 19 Jul 2009
Posts: 72
PostPosted: Tue Dec 08, 2009 3:55 pm Post subject: Question on WINVAL and "other" Adjusted plus minus Reply with quote
Hi, maybe you could help me with some doubts arising from the book of Winston: Mathletics.
He briefly explains his WINVAL system and he reports 2006-2007 Adjusted +/- ratings (page 207).
I still do not know enough the procedure implemented by Rosenbaum and others regarding their Adjusted +/- computations, but it seems there are some differences in the final result. Is this correct?
I was trying to compare the results depicted by Winston in his page 207 with the information provided by 82games.com (http://www.82games.com/ilardi1.htm) , but the latter included the play-offs, so the comparison is not feasible.
In addition, in basketballvalue.com there is no data available from the 20062007 season.
In sum, how many different procedures for computing Adjusted plus/minus are available? WINVAL and the other procedures are in agreement?
Thank you very very much.
_________________
Jose A. Martínez
http://www.upct.es/~beside/jose
http://basket-research.blogspot.com
Back to top
View user's profile Send private message Send e-mail Visit poster's website
jsill
Joined: 19 Aug 2009
Posts: 73
PostPosted: Tue Dec 08, 2009 4:49 pm Post subject: Reply with quote
At his talk at NESSIS, Winston indicated that he uses something other than standard linear regression, but I believe this "something else" is propietary and is not something he's revealed publicly up until this point.
I also advise against using standard linear regression, since ridge regression (a.k.a. linear regression + regularization) does a much better job at predicting future out-of-sample games. You can read more about this on my site if you're interested.
In his talk, Winston did openly describe two specific areas where his approach differs from other adjusted +/- approaches. These differences are in addition to whatever proprietary twist on linear regression he's using:
1) He weights each game snippet by the game time elapsed rather than by the number of possessions. I gave this a try myself and it didn't seem to change the results very much, but I only looked at it very briefly, so I can't say that with a lot of confidence.
2) He does not use the "reference player" approach where players whose total number of minutes played falls below a certain threshold are lumped together and modelled as a single "reference" player. I believe every player in the league is represented individually in the regression he does, if I understood him correctly from the video of his talk.
Back to top
View user's profile Send private message
Jose A. Martínez
Joined: 19 Jul 2009
Posts: 72
PostPosted: Wed Dec 09, 2009 11:39 am Post subject: Reply with quote
Thanks Joe.
I have heard his talk, but honestly is somewhat hard to follow for a non-Englihs speaker. However, I do notice that he compares his method wiht the Adj+/- results of basketballvalue.com. They are different (he speak about Dirk Nowitzki).
It seems that he use an alterantive approach for this Adj+/- ratings. Thanks for identifying these two different items where his method differs from others.
In addition, Winston explains in his book (page 206) that he uses the "Excel Solver" tool for minimizing the following:
iSum of (Sum of Team 1 player ratings for game i - Sum of Team 2 player ratings for game- Poionts team 1 wins game i by)^2
I am not very much familiar with optimization methods but it seems that he uses a linear programming approach or similar to find the solution of the problem. However he uses a very simple example and maybe is not illustrative enough to characterize his procedure...
Let me tell that I am impressed for the method you (JSill) develop in your site: Regularized Adjusted Plus Minus. It is a great work. You follow some of the principles of data mining approach (training sample and cross validation) in order to give robustness to your results. However, your procedure is so complex that we have to be very familiar with concepts such as regularization, bayesian estimation, etc (I think I would need several weeks of study to fully understand your method...and I will do it).
_________________
Jose A. Martínez
http://www.upct.es/~beside/jose
http://basket-research.blogspot.com