This is Google's cache of viewtopic.php?p=32142&sid=eeb1b0e3f82bd7404c9e10590bbe1a14. It is a snapshot of the page as it appeared on Apr 11, 2011 16:47:02 GMT. The current page could have changed in the meantime. Learn more

Text-only version
These search terms are highlighted: jsill  
APBRmetrics :: View topic - A new player rating system
APBRmetrics Forum Index APBRmetrics
The statistical revolution will not be televised.
 
 FAQFAQ   SearchSearch   MemberlistMemberlist   UsergroupsUsergroups   RegisterRegister 
 ProfileProfile   Log in to check your private messagesLog in to check your private messages   Log inLog in 

A new player rating system
Goto page Previous  1, 2
 
Post new topic   Reply to topic    APBRmetrics Forum Index -> General discussion
View previous topic :: View next topic  
Author Message
back2newbelf



Joined: 21 Jun 2005
Posts: 276

PostPosted: Sat Apr 24, 2010 5:39 pm    Post subject: Reply with quote

Even though this metric will undergo some changes in the (hopefully near) future I still decided to publish the rankings from this regular season, simply because last years' rankings were really helpful to predict team wins for the 2009-2010 regular season, even though it made really bad predictions for Oklahoma and New Jersey.

I feel like it helped the most with
-Boston (not seeing a valuable addition in Rasheed Wallace*)
-Philly (the fact that Andre Miller was importand for them*)
-Milwaukee (didn't see valuable players in Charlie V and Jefferson*)
-Detroit (Charlie V again*)
-New Orleans

*you can obviously argue that those teams did bad for enterely different reason, this is just my take on it

https://docs.google.com/leaf?id=0B31RF2rWzwSwNmNjOGM5YjYtMjc5Ni00MzViLWFkYTEtYWZmODQyNmI2MmE1&hl=en
(let me know if it doesn't work)
most files have the following format:
name|differential created through specific action|sum of possessions|differential created by specific action per possession

Unfurtunately this metric has a crush on low post players and point guards, so that will be the story for most teams.
For some weird reason white Power Forwards look really strong, there are three of them in the top 5 (possession based).

What surprised me the most is that, even though his TS% is really high, this metric still doesn't like Durant. He made a big jump from last year though. A lot of high volume scorers don't look good: Bryant, Carmelo, Ellis, Nowitzki.

This years' all NBA team would be Lebron, Rondo, Ginobili, Ben Wallace and Dwight Howard.


Comments on some of the files:
-assists vs turnovers:
This is credit for assisting added to the negative credit for turning the ball over. If you want to look for why Durant doesn't look strong , this is the main reason. (Possession based) he is surrounded by players like Tyson Chandler, Tyrus Thomas and Amare'. His TOV% is low but his turnovers seem to result in alot of opponent points (through fast breaks).

-creating extra possessions/offensive rebounding:
I had never heard of Jon Brockman before but his OREB% is pretty sick (18.2%)

-negative from opponent offensive rebounds:
Dallas, Cleveland, Orlando and the Spurs shining, with GSW, PHX, DET on the other side of the spectrum. Surprised to see Chris Andersen and J.R.Smith near the bottom.

-shot defense(opponent TS% in regular possessions):
Spurs, Orlando and Cleveland shining again. Antonio McDyess is 3rd best in preventing offensive rebounds and/or preventing baskets after offensive rebounds and he's 2nd best in this category. Quite impressive


Team Comments (the words "efficient" and "productive" are used with regard to this metric):

Atlanta seems to be about Horford and Smith, fortunately with no one to drag them down.
Boston seems to be pretty much all Rondo, with Wallace dragging.
Chicago seems to be Noah and Gibson, with a variety of wings and guards dragging.
Cleveland seems to be Lebron and Varejao, with no one really dragging, Jamison looks bad though, especially in comparison with Varejao.
Dallas managed to ship off super inefficient Howard for the super efficient Haywood, but Butler is negative.
Lawson looks awesome for a rookie in Denver, with Smith being the only big negative.
Detroit has Wallace and Jerebko in the green, but everybody else is negative. 4 players in deep red, two of them being offseason additions and one of them being Rip Hamilton who just barely missed being the worst player in the league (production wise). He has been that bad for the last several years.
Houston with several productive players: Lowry, Hayes, Landry and Battier but unfortunately just as many being quite unproductive, "led" by Ariza and MIP Aaron Brooks.
Earl Watson is awesome for Indiana. Foster has been one of the most efficient for several years but he's never playing much.
Camby was pretty much the only positive guy in LAC, Kaman and Rasual Butler are among the 6 most unproductive players in the league.
Miami has several guys in the positive with Wade leading the way. Unfortunately Beasley is dragging them way down
Minnesota with only Love significantly positive, Brewer being the 4th worst player in the league, Flynn being the absolute worst (in overall production).
The blame in New Jersey is split between a whole bunch of guys, with no one being in the deep red (mostly, though, because of limited playing time)
This metric says Oklahoma improved because everybody on their team improved. Collison, Westbrook and Ibaka looking strong.
Orlando has a gazillion players in the positive, only Lewis seems to be dragging a tiny bit.
Roy, Miller and Camby looking good in Portland, Webster being the only slight negative.
Nocioni drags the Kings down, he's been doing that for a couple of years now(in Chicago too)
Blair looks good for for a rookie on the Spurs, who get their biggest lift from Ginobili and Duncan.
Johnson looks really awesome for Toronto but he's not playing that much (foul trouble?). Instead, Bargnani is playing. He's the 5th worst rated player.
Both Brewer and Maynor looked pretty strong in Utah, both were shipped, Maynor now looks strong in Oklahoma.
Haywood seemed to be the lone productive player in Washington. He was in the top10 in overall production this year. That's quite a loss.
Back to top
View user's profile Send private message
Crow



Joined: 20 Jan 2009
Posts: 825

PostPosted: Sat Apr 24, 2010 6:04 pm    Post subject: Reply with quote

What is the difference between "overall" and "rating"? Do both include all of defense consistent with your earlier statement?

"Shot defense" is strictly shot defense and nothing beyond it, right?
Back to top
View user's profile Send private message
back2newbelf



Joined: 21 Jun 2005
Posts: 276

PostPosted: Sat Apr 24, 2010 7:29 pm    Post subject: Reply with quote

Crow wrote:
What is the difference between "overall" and "rating"?

None, should have made that clearer, sorry. "Overall" is sorted by production over the entire season, while "rating by possession" is sorted by production/possessions
Quote:

Do both include all of defense consistent with your earlier statement?

Both include defense as was explained in my first post.
Quote:

"Shot defense" is strictly shot defense and nothing beyond it, right?

It's credit for opponents missing shots added to (negative credit) for opponents making shots. Regular possessions only
Back to top
View user's profile Send private message
Mike G



Joined: 14 Jan 2005
Posts: 3625
Location: Hendersonville, NC

PostPosted: Sun Apr 25, 2010 7:01 am    Post subject: Reply with quote

Quote:
Kaman and Rasual Butler are among the 6 most unproductive players in the league.

Can you offer a layman's description of how Kaman is "most unproductive"?

He's easily his team's most prolific scorer. He rebounds.

Could an ESPN talking head explain to a viewing audience how it is that Kaman is really "unproductive"?
_________________
`
36% of all statistics are wrong
Back to top
View user's profile Send private message Send e-mail
back2newbelf



Joined: 21 Jun 2005
Posts: 276

PostPosted: Mon Apr 26, 2010 12:44 pm    Post subject: Reply with quote

Mike G wrote:

Can you offer a layman's description of how Kaman is "most unproductive"?

I would say:
If he's asked to create shots for himself or others the end result is worse than average. Unfortunately, either because he either thinks he's doing a good job at it, or because he's frequently asked to do it , he's trying to create shots alot which probably hurts his team more than if he tried it less often


(he is 12th on the Clippers roster in ORtg, yet has the highest Usg% of everybody)
Back to top
View user's profile Send private message
back2newbelf



Joined: 21 Jun 2005
Posts: 276

PostPosted: Sun Sep 12, 2010 10:59 am    Post subject: Reply with quote

finally got around doing some tests with this metric. retrodiction performance was tested on march & april for the last four years, only using data through february from the season we want to retrodict.

x = performance if we had predicted the home team to win by 3.61 points every game
y = performance of my method
z = performance of homecourt-adjusted point differential

all players were included, no matter how many minutes they had played before.
Code:

2007:
games 342
              x      y     z
RMSE         13.8  13.52  13.05
r-squared     0    0.048  0.11


2008:
games 350
              x      y     z
RMSE         14.4  13.41   11.81
r-squared    0     0.133   0.33


2009:
games 339
              x      y     z
RMSE        12.7 12.05*  11.56
r-squared    0   0.098*   0.18


2010:
games 337
              x      y     z
RMSE         13.63 12.24 11.92
r-squared     0    0.22  0.24

*Joe Sill reported an RMSE of 11.47 (R-squared 0.169) for the same time period for regularized adjusted +/-.
He also reported an RMSE of 12.76 for non-regularized adj +/- with a single season of data, and 12.01 (R-squared 0.088) with multiple seasons of data
Unfortunately it seems I don't have all the games he had. I know this because the RMSE I get for the simple method(x) differs from his(my 12.7 vs his 12.58 ).

Still it seems that my method seems worse than regularized adj. +/- and better than non-regularized adj. +/-. It never outperforms point differential, which is ultimately what I want. Nevertheless, I find the results quite promising since I haven't messed with the weights at all so far


Last edited by back2newbelf on Mon Sep 13, 2010 2:29 pm; edited 2 times in total
Back to top
View user's profile Send private message
jsill



Joined: 19 Aug 2009
Posts: 73

PostPosted: Sun Sep 12, 2010 3:45 pm    Post subject: Reply with quote

Quote:
Unfortunately it seems I don't have all the games he had. I know this because the RMSE I get for the simple method(x) differs from his(my 12.0 vs his 12.58 ).


How many games do you have in your dataset for March/April of 2009? You should have 348 games.

By the way, I had extracted the final score from play-by-play logs, and there was a small error in the log for one of the games,. With that error fixed, the RMSE that I get if you simply predict the mean home margin of 3.61 points is actually 12.565.
Back to top
View user's profile Send private message
back2newbelf



Joined: 21 Jun 2005
Posts: 276

PostPosted: Sun Sep 12, 2010 6:26 pm    Post subject: Reply with quote

jsill wrote:

How many games do you have in your dataset for March/April of 2009? You should have 348 games.

As listed in my earlier post I had 314 games. I found some more game logs and now have:
- 339 games (so still some games missing)
- a base RMSE (method X) of 12.70
- R-Squared of 0.125 for my method

Updating my earlier post..

jsill, do you have the R-Square numbers for 2007/2008/2010 for regularized adjusted +/-?
Back to top
View user's profile Send private message
jsill



Joined: 19 Aug 2009
Posts: 73

PostPosted: Sun Sep 12, 2010 9:07 pm    Post subject: Reply with quote

Quote:
jsill, do you have the R-Square numbers for 2007/2008/2010 for regularized adjusted +/-?


First, let me check something. Are you generating your ratings using data from the entire season (including March and April) and then testing the predictions on March and April, or are you generating your ratings only using data through February and then testing on March and April? The question also applies for your method z (homecourt-adjusted point differential).
Back to top
View user's profile Send private message
back2newbelf



Joined: 21 Jun 2005
Posts: 276

PostPosted: Mon Sep 13, 2010 4:16 am    Post subject: Reply with quote

jsill wrote:

First, let me check something. Are you generating your ratings using data from the entire season (including March and April) and then testing the predictions on March and April, or are you generating your ratings only using data through February and then testing on March and April? The question also applies for your method z (homecourt-adjusted point differential).

unless there is a bug in my code I'm generating through February to "forecast" March 1st, then generate through March 1st to "forecast" March 2nd and so on
Back to top
View user's profile Send private message
jsill



Joined: 19 Aug 2009
Posts: 73

PostPosted: Mon Sep 13, 2010 9:35 am    Post subject: Reply with quote

OK. Well, it's not an apples-to-apples comparison to what I did. I obtained the player ratings one time, based on data through February, and then forecast March and April.
Back to top
View user's profile Send private message
back2newbelf



Joined: 21 Jun 2005
Posts: 276

PostPosted: Mon Sep 13, 2010 2:27 pm    Post subject: Reply with quote

you're right. I updated my earlier post with the correct numbers. here they are again
Code:
2007:
games 342
              x      y     z
RMSE         13.8  13.52  13.05
r-squared     0    0.048  0.11


2008:
games 350
              x      y     z
RMSE         14.4  13.41   11.81
r-squared    0     0.133   0.33


2009:
games 339
              x      y     z
RMSE        12.7 12.05  11.56
r-squared    0   0.098   0.18


2010:
games 337
              x      y     z
RMSE         13.63 12.24 11.92
r-squared     0    0.22  0.24
Back to top
View user's profile Send private message
gabefarkas



Joined: 31 Dec 2004
Posts: 1313
Location: Durham, NC

PostPosted: Sat Sep 18, 2010 9:47 pm    Post subject: Reply with quote

I was reading something else recently, and it reminded me of this discussion and the whole notion of model validation. So, I wanted to share it with the group:
http://www.stat.columbia.edu/~cook/movabletype/archives/2010/08/useful_models_m.html

Gelman's blog is usually pretty good for some stimulating tidbits. And I'm not just saying that because I'm a Columbia alum...
Back to top
View user's profile Send private message Send e-mail AIM Address
Display posts from previous:   
Post new topic   Reply to topic    APBRmetrics Forum Index -> General discussion All times are GMT - 5 Hours
Goto page Previous  1, 2
Page 2 of 2

 
Jump to:  
You cannot post new topics in this forum
You cannot reply to topics in this forum
You cannot edit your posts in this forum
You cannot delete your posts in this forum
You cannot vote in polls in this forum


Powered by phpBB © 2001, 2005 phpBB Group