Page 1 of 3
One Year RAPM and Weighted Ridge Regression
Posted: Thu Jan 03, 2013 4:05 am
by permaximum
EDIT (DSMok1): Split off of Permaximum Ratings and Rankings (Updated Often), since this is a different topic.
I think, it's better to ask it here. Does anyone have 1 year RAPM with the prior as zero for previous seasons? I mean, the very first RAPM which is just the regularized-APM. If nobody has it, the second question comes. How can I run a "weighted ridge regression" with R or another language/software? lm.ridge and ridge.cv functions don't have weight argument.
Re: Permaximum Ratings and Rankings (Updated Often)
Posted: Thu Jan 03, 2013 12:38 pm
by DSMok1
permaximum wrote:I think, it's better to ask it here. Does anyone have 1 year RAPM with the prior as zero for previous seasons? I mean, the very first RAPM which is just the regularized-APM. If nobody has it, the second question comes. How can I run a "weighted ridge regression" with R or another language/software? lm.ridge and ridge.cv functions don't have weight argument.
Right, they don't. What I found as the way to do it was to assign each player a single observation of 0 with a specified weight, and then run the normal weighted regression function.
You'd have to find that weight through leave-n out cross validation.
The data frame would likely have to be n*2 players wide (a column for O and a column for D), plus the offensive efficiency of the O side of the matchup & a column for weight (# of possessions), and would be 2*matchups+n players tall (for the O observation for each side of a matchup and the prior).
Make sense?
Re: Permaximum Ratings and Rankings (Updated Often)
Posted: Thu Jan 03, 2013 8:23 pm
by Crow
If J.E. does not offer them up to you directly, I probably have some previous year non-prior informed RAPM files I could dig up later. If you send me a private message with an e-mail address I could probably send what I find sometime in the next week.
Re: Permaximum Ratings and Rankings (Updated Often)
Posted: Fri Jan 04, 2013 3:20 am
by EvanZ
permaximum wrote:I think, it's better to ask it here. Does anyone have 1 year RAPM with the prior as zero for previous seasons? I mean, the very first RAPM which is just the regularized-APM. If nobody has it, the second question comes. How can I run a "weighted ridge regression" with R or another language/software? lm.ridge and ridge.cv functions don't have weight argument.
Use glmnet with
.
http://cran.r-project.org/web/packages/ ... index.html
Re: Permaximum Ratings and Rankings (Updated Often)
Posted: Fri Jan 04, 2013 3:52 am
by permaximum
@DSMok1
Thanks for the explanation but I think I'm gonna check glmnet function first.
@Crow
Thank you. I will probably send the pm
@EvanZ
Thank you so much. I'll try it.
@MikeG
Ed Davis started in the last 10 games of Toronto and Toronto went 8-2. Before that, Toronto went 4-18 with him coming of the bench
Let's see what happens when Drummond and McGee starts for their teams.
Edit: Although glmnet gives a little bit different results than other functions (lm, lm.ridge, ridge.cv) it's good to see it work with weights argument while the others can't. Ronnie Price tops 2007/08 1 year pure RAPM chart. Can anybody confirm this nonsense is not a problem of the data I took from basketballvalue.com? If it's accurate, I will realy start to doubt +/-, because with many years of data, priors, involvement of box score etc. any rating can pass laugh test. I hope there were errors with the data.
Re: Permaximum Ratings and Rankings (Updated Often)
Posted: Fri Jan 04, 2013 12:25 pm
by DSMok1
1 Year APM is extremely noisy due to very heavy collinearity issues (many players share the court most or all of the time).
Since players switch teams and rotations change, multi-year APM is way better.
See my exhaustive review on the subject at
http://godismyjudgeok.com/DStats/2011/n ... ilization/
Re: Permaximum Ratings and Rankings (Updated Often)
Posted: Fri Jan 04, 2013 3:44 pm
by permaximum
DSMok1 wrote:1 Year APM is extremely noisy due to very heavy collinearity issues (many players share the court most or all of the time).
Since players switch teams and rotations change, multi-year APM is way better.
See my exhaustive review on the subject at
http://godismyjudgeok.com/DStats/2011/n ... ilization/
It wasn't APM. Price tops all players in
RAPM chart that year. I thought RAPM was made exactly for collinearity issues.
I read that article before and found very informative. It was a great read.
Re: Permaximum Ratings and Rankings (Updated Often)
Posted: Fri Jan 04, 2013 4:20 pm
by DSMok1
permaximum wrote:DSMok1 wrote:1 Year APM is extremely noisy due to very heavy collinearity issues (many players share the court most or all of the time).
Since players switch teams and rotations change, multi-year APM is way better.
See my exhaustive review on the subject at
http://godismyjudgeok.com/DStats/2011/n ... ilization/
It wasn't APM. Price tops all players in
RAPM chart that year. I thought RAPM was made exactly for collinearity issues.
I read that article before and found very informative. It was a great read.
Even with the RAPM adjustment--I still am vary wary of anything single year. Regularization only helps the collinearity so much. If two players are perfectly collinear--you just can't disentangle them at all. Period.
Re: Permaximum Ratings and Rankings (Updated Often)
Posted: Fri Jan 04, 2013 6:20 pm
by talkingpractice
permaximum wrote: Edit: Although glmnet gives a little bit different results than other functions (lm, lm.ridge, ridge.cv) it's good to see it work with weights argument while the others can't. Ronnie Price tops 2007/08 1 year pure RAPM chart. Can anybody confirm this nonsense is not a problem of the data I took from basketballvalue.com? If it's accurate, I will realy start to doubt +/-, because with many years of data, priors, involvement of box score etc. any rating can pass laugh test. I hope there were errors with the data.
From testing we've done, there just isn't much of a predictive nature at all to a non prior informed 1yr apm model (regularized or not). You need years of data to get the n's big enough.
Re: Permaximum Ratings and Rankings (Updated Often)
Posted: Fri Jan 04, 2013 11:00 pm
by permaximum
talkingpractice wrote:permaximum wrote: Edit: Although glmnet gives a little bit different results than other functions (lm, lm.ridge, ridge.cv) it's good to see it work with weights argument while the others can't. Ronnie Price tops 2007/08 1 year pure RAPM chart. Can anybody confirm this nonsense is not a problem of the data I took from basketballvalue.com? If it's accurate, I will realy start to doubt +/-, because with many years of data, priors, involvement of box score etc. any rating can pass laugh test. I hope there were errors with the data.
From testing we've done, there just isn't much of a predictive nature at all to a non prior informed 1yr apm model (regularized or not). You need years of data to get the n's big enough.
Actually, I was looking for explanation of player performances in a season instead of prediction of future performances. That's why I wanted 1 year RAPM. But the results were very disappointing. So this Ronnie Price joke definetely confirms RAPM is very bad for explenation of performances in a season. As for prediction, what's the target? +/- of next season?
Re: Permaximum Ratings and Rankings (Updated Often)
Posted: Sat Jan 05, 2013 12:37 am
by talkingpractice
Ed Davis started in the last 10 games of Toronto and Toronto went 8-2. Before that, Toronto went 4-18 with him coming of the bench Let's see what happens when Drummond and McGee starts for their teams.
Per Rotoworld lol...
"Coach Dwane Casey said Friday that "it will be hard to pull Ed Davis out of the (starting) line up" when Andrea Bargnani (elbow) returns. Davis has started the last 10 games, averaging 12.4 points, 7.1 rebounds, 0.9 blocks and 0.9 steals. The Raptors have gone 8-2 during that stretch. "He's earned it," Casey said." Jan 4 - 6:08 PM
Re: Permaximum Ratings and Rankings (Updated Often)
Posted: Sat Jan 05, 2013 12:47 am
by EvanZ
permaximum wrote:
Actually, I was looking for explanation of player performances in a season instead of prediction of future performances. That's why I wanted 1 year RAPM. But the results were very disappointing. So this Ronnie Price joke definetely confirms RAPM is very bad for explenation of performances in a season. As for prediction, what's the target? +/- of next season?
Ronnie Price only played 10 mpg that season. Who's the highest rated player that played > 30 mpg? It's probably going to be more meaningful.
Re: Permaximum Ratings and Rankings (Updated Often)
Posted: Sat Jan 05, 2013 1:54 am
by Kevin Pelton
Despite its limitations, I wish there was single-season RAPM (or even APM) available to answer questions that multi-season RAPM isn't designed to answer, like how Dwight Howard's defense is different post-surgery.
Re: Permaximum Ratings and Rankings (Updated Often)
Posted: Sat Jan 05, 2013 2:42 am
by EvanZ
I'm planning to do it, KP. Maybe at the All-Star break.
Re: Permaximum Ratings and Rankings (Updated Often)
Posted: Sat Jan 05, 2013 3:43 am
by bchaikin
like how Dwight Howard's defense is different post-surgery
for howard Synergy shows excellent man defense, a very low PPP and eFG% allowed, almost as low as Cs joakim noah and marc gasol. he's also blocking shots at his highest rate in 3 seasons...