page 1 of 2
Author Message
Jose A. Martínez
Joined: 19 Jul 2009
Posts: 72
PostPosted: Thu Dec 17, 2009 3:49 pm Post subject: eWins, WinShares and EWA Reply with quote
Dear colleagues, I am sorry, but I have still problems to understand eWins. I have read many of your posts but I am still in trouble with eWins.
Would you be so kind to briefly explain (in plain English please, remember that I am Spanish Very Happy ), the difference between eWins and Win Shares and Hollinger's Estimated Wins Addes (EWA)?
They seem very similar indexes, because they are based on wins attributed to players.
In addition I have problems to correctly understand the concept of "Replacement player"
Finally, I would like to compare eWins, WinShares and EWA for the 2008/2009 season. WinSahres is available at basketball-reference.com, but I could not find eWins for that season. Maybe Mike G could provide the 2008/2009 results.
Thank you very much. I know some of these questions may disturb you because they are very clear for most of APBRmembers. I am sorry for that. But I am trying to make an effort to disseminate the APBR knowledge that comes from this marvellous site to Spanish basketball readers.
Thanks again!
_________________
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
Mike G
Joined: 14 Jan 2005
Posts: 3612
Location: Hendersonville, NC
PostPosted: Thu Dec 17, 2009 6:34 pm Post subject: Reply with quote
Jose,
No one is clear about eWins, so don't feel alone.
Here are last season's eW and the numbers behind them:
http://spreadsheets.google.com/pub?key= ... utput=html
A major distinction of eWins (equivalent wins) and others is that eW do not attempt to sum to the team's actual or expected (pythagorean, from point-differential) wins, because the talent, productivity, or performance of a team's players are not proportional to their win total.
A team's ratio of points scored/allowed is often nowhere near their win/loss ratio, unless they are nearly average. A team who's average score is 100-91 goes 66-16 (Cle), and a team with 100-109 scores goes 17-65 (Sac).
Is the team that scores .90 as much as the average team actually 90% as good? Or are they (17/41) just .41 as good?
I think it's somewhere in the middle, and eWins exists as a demonstration of that. A 17-65 team nets 29 eWins. They are 24 wins below .500, and their eWin total will be half of that disparity, or 12 less than 41.
A 65-17 team would total 53 eWins. They aren't really (65/17) 3.8 times as good as the 17-65 team, but (53/29) 1.8 times as good.
For an 82 game season, a team's expected wins would be
xW = eW*2 - 41
Why this works, I still don't know. But after scaling to opponent point and rebound totals, it's that way, year after year. It works in part-seasons and in playoff series.
_________________
`
36% of all statistics are wrong
Back to top
View user's profile Send private message Send e-mail
Mike G
Joined: 14 Jan 2005
Posts: 3612
Location: Hendersonville, NC
PostPosted: Fri Dec 18, 2009 8:20 am Post subject: Reply with quote
In a part season or a playoff series, the general form of the expected wins is:
xW = eW*2 - G/2
where G is the number of games played.
In last year's playoffs, round 1, the Hornets lost in 5 games to Denver, by an incredible average score of 86-111. The Pythagorean suggests that with this differential, Denver 'should have' won 4.87 to NO .13
Or, if they played 82 games like this, NO would go 6-76.
New Orleans players, ranked by minutes (divided by 5 games), their eWins and, from b-r.com, their WinShares.
http://www.basketball-reference.com/teams/NOH/2009.html
e480 is eWins/480 min. An average rate (without overtimes) is 1.0 .
w480 is, similarly, WS/480 min.
Code:
Hornets mpg eW e480 WS w480
Paul,Chris 40 .40 .97 -.1 -.2
West,David 35 .33 .90 -.3 -.8
Stojakovic,Peja 32 .10 .31 -.1 -.3
Butler,Rasual 31 .12 .35 .0 .0
Posey,James 24 .20 .80 .0 .0
Chandler,Tyson 18 .02 .10 -.2 -1.1
Marks,Sean 16 .07 .44 .1 .6
Daniels,Antonio 12 .02 .15 -.2 -1.6
Armstrong,Hilton 10 .02 .14 -.2 -1.9
Brown,Devin 6 -.01 -.13 -.1 -1.5
Wright,Julian 6 .02 .40 .0 .0
Peterson,Morris 4 .02 .49 .0 .0
total 1.32 -1.1
Even in a mismatch of such epic proportions, I think the losers consist of a few players battling heroically, but with inadequate support at too many positions. Thus the blowout.
A system that forces a player's performance metric to reflect his team's W% seems to suggest that an entire roster can be doing more to abet the opposition than they are doing to help their own team.
The 1.32 equivalent-wins produced in 5 games says the Hornets should win an expected
1.32*2 - 5/2 = 2.63-2.50 = 0.13 games
_________________
`
36% of all statistics are wrong
Back to top
View user's profile Send private message Send e-mail
jkubatko
Joined: 05 Jan 2005
Posts: 702
Location: Columbus, OH
PostPosted: Fri Dec 18, 2009 9:07 am Post subject: Reply with quote
Mike G wrote:
A system that forces a player's performance metric to reflect his team's W% seems to suggest that an entire roster can be doing more to abet the opposition than they are doing to help their own team.
Ridiculous small sample size issues aside, the Win Shares system does not force player wins to equal team wins.
_________________
Regards,
Justin Kubatko
Basketball-Reference.com
Back to top
View user's profile Send private message Send e-mail Visit poster's website
Mike G
Joined: 14 Jan 2005
Posts: 3612
Location: Hendersonville, NC
PostPosted: Fri Dec 18, 2009 6:50 pm Post subject: Reply with quote
OK, here are regular-season and playoff eWin rates for those hapless '09 Hornets. Per 480 min, average would be 1.00
Code:
eW/480 min mpg RS PO PO/RS
Paul,Chris 40 2.67 .97 .36
West,David 35 1.58 .90 .57
Stojakovic,Peja 32 .74 .31 .42
Butler,Rasual 31 .60 .35 .59
Posey,James 24 .51 .80 1.57
Chandler,Tyson 18 .83 .10 .12
Marks,Sean 16 .25 .44 1.79
Daniels,Antonio 12 .49 .15 .31
Armstrong,Hilton 10 .50 .14 .28
Brown,Devin 6 .41 -.13 -.31
Wright,Julian 6 .68 .40 .58
Peterson,Morris 4 .54 .49 .90
eWins says that when your top 4 players drop down to 36-59% of their normal effectiveness, you get slaughtered. Chandler was 1/8 of himself.
It isn't requisite that players are deemed to be negative contributors, to account for 25 ppg beatings. This happens when they are about half their norms.
_________________
`
36% of all statistics are wrong
Back to top
View user's profile Send private message Send e-mail
Mike G
Joined: 14 Jan 2005
Posts: 3612
Location: Hendersonville, NC
PostPosted: Sat Dec 19, 2009 7:16 am Post subject: Re: eWins, WinShares and EWA Reply with quote
josean.martinez@upct.es wrote:
... correctly understand the concept of "Replacement player"
The term is rather ambiguous, and there are as many variations as there are people who use it. I prefer to think of 'replacement level' player as one who can be easily replaced. In other words, someone who is filling a roster spot because he may be as good as other available players, but not necessarily better.
He may be a 'project' that may or may not develop, a fouling specialist, a good practice player, etc. I use a definition provided by the eWins spreadsheet: someone who adds no wins to an average team. His eW = 0.
Last year (2009) there were about 330 players with eW > 0.
This suggests an average 12th-man might be 'replacement-level'.
For 2009, a player's eWins were:
eW = (T - 12.95)*Min/5473
T is the summary total of other per36 rates
12.95 is the T Rate of a Replacement
5473 is an arbitrary divisor that makes player eWins sum to 1230 (41*30), the total games won by 30 teams in 82 games.
_________________
`
36% of all statistics are wrong
Back to top
View user's profile Send private message Send e-mail
Jose A. Martínez
Joined: 19 Jul 2009
Posts: 72
PostPosted: Tue Dec 22, 2009 2:15 am Post subject: Reply with quote
Many thanks for the explanations.
Ok, for example, NOH in 2008/2009 season (49-33), Following your reasoning, eWins would be 45.
If we sum the eWins for players that comes from this link:
http://spreadsheets.google.com/pub?key= ... DsTQ&gid=0
Then we have 42.68 eWins. In addition we have 47.9 WinShares (see basketball-reference.com)
It seems that WinShares performs better that eWins, because the error is
((47.9-49)^2)/49 = 0.025 for WinShares
((42.68-45)^2)/45= 0.12 for eWins
Another question would be how eWins are distributed among players, i.e. which is the procedure to obtain individual eWins?
_________________
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
Mike G
Joined: 14 Jan 2005
Posts: 3612
Location: Hendersonville, NC
PostPosted: Tue Dec 22, 2009 8:04 am Post subject: Reply with quote
Yes, there is generally a difference between the games a team wins and their Pythagorean -- which is what eWins is designed to approach (with minimal error across the league).
The team totals 42.68 eW, so expected wins would be
xW = 2*42.68 - 41 = 85.36-41 = 44.36
Using an exponent of 13 in the Pyth formula, NOH were expected to win 45.33 games, and they exceeded this with 49.
Both eWins and Pythagorean create a number of expected wins, but teams win or lose close games.
From above post:
Quote:
For 2009, a player's eWins were:
eW = (T - 12.95)*Min/5473
_________________
`
36% of all statistics are wrong
Back to top
View user's profile Send private message Send e-mail
Jose A. Martínez
Joined: 19 Jul 2009
Posts: 72
PostPosted: Tue Dec 22, 2009 8:23 am Post subject: Reply with quote
OK thanks. Just some additional questions on individual eWins
Quote:
For 2009, a player's eWins were:
eW = (T - 12.95)*Min/5473
T is the summary total of other per36 rates
12.95 is the T Rate of a Replacement
5473 is an arbitrary divisor that makes player eWins sum to 1230 (41*30), the total games won by 30 teams in 82 games.
Would you be so kind to explain (if possible) how to compute "T"? Maybe a weighted combination of points, rebounds, etc.
Another question would be how results would change if eWins were not computed per36minutes, but per48minutes, or as an absolute value...
Thanks
_________________
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
Mike G
Joined: 14 Jan 2005
Posts: 3612
Location: Hendersonville, NC
PostPosted: Tue Dec 22, 2009 8:34 am Post subject: Reply with quote
The 2009 T rates are :
T = Sco + Reb + 1.33*Ast -.24*PF + 1.5*Stl - 1.55*TO + 1.5*Blk
Sco, Reb, etc being the adjusted per36 rates appearing on the spreadsheet.
If these elements were per48, then T and Tr (replacement T) would just be 1/3 larger.
Multiplying by total minutes erases the per-minute influence.
Ranking players by eWins per minute (or as I prefer, per 484 min.) is exactly the ranking one gets when ranking by T rate.
If T is an 'absolute' productivity, and no replacement value is subtracted, then weak teams would have eWins even closer to that of the best teams. Rather like points scored are much closer, top to bottom, than actual Wins.
_________________
`
36% of all statistics are wrong
Back to top
View user's profile Send private message Send e-mail
Jose A. Martínez
Joined: 19 Jul 2009
Posts: 72
PostPosted: Tue Dec 22, 2009 11:51 am Post subject: Reply with quote
Wow Mike, nice explanation again!.
Maybe the last question Very Happy
Quote:
T = Sco + Reb + 1.33*Ast -.24*PF + 1.5*Stl - 1.55*TO + 1.5*Blk
How these weights are derived?
Thanks!!!!!!!
_________________
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
Mike G
Joined: 14 Jan 2005
Posts: 3612
Location: Hendersonville, NC
PostPosted: Tue Dec 22, 2009 12:11 pm Post subject: Reply with quote
Those are typical weights, and sometimes early in the season they'll be different; except that I never change them for Sco, Reb, Ast.
What determines the weights is that the spreadsheet returns minimal error at these weights. Error defined as teams' average difference between PythWins and xWins (eW*2-G/2).
It would be preferable if these weights never changed. But when doing eWins for specific playoff series, it seems to be necessary. Part of it may be in what is 'counted' as an assist or a block, for example; or the presence/scarcity of 'garbage time'.
The final 'weights' are also contingent on additional parameters, such as the weight on effective shooting%, starter/sub factor, unassisted%, and several others. To really see this, we'd have to sit down in front of my computer.
Until I can get a functioning eWins spreadsheet to anyone and be there to operate it, please just consider the basic principal: Player contributions are not proportional to success (W%). Only above a certain (replacement) level are they proportional, and then only relative to average (rather than to absolute).
_________________
`
36% of all statistics are wrong
Back to top
View user's profile Send private message Send e-mail
Jose A. Martínez
Joined: 19 Jul 2009
Posts: 72
PostPosted: Tue Dec 22, 2009 12:25 pm Post subject: Reply with quote
Thanks a lot Mike, that's ok. I can now understand eWins.
_________________
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
Mike G
Joined: 14 Jan 2005
Posts: 3612
Location: Hendersonville, NC
PostPosted: Mon Jan 11, 2010 8:13 am Post subject: Reply with quote
Thru games of Jan. 9 (Sat), some various systems' player wins allotments.
WS (Win Shares) are from - http://www.basketball-reference.com/pla ... i?id=IHcoP
EWA (Estimated Wins Added, the Hollinger stat) is at - http://insider.espn.go.com/nba/hollinger/statistics
WARP (Wins Above Replacement Player, Kevin Pelton) and WP (Wins Produced) are from Basketball Prospectus, gathered from individual player pages.
eW (Equivalent Wins) are mine.
Players ranked by average (5th root of the product of the 5).
Code:
wins attributing system per 484 minutes (1.0 = avg)
avg player eW WS WARP WP EWA eW WS WARP WP EWA
11.7 James,Lebron 8.2 8.8 12.1 17.1 14.4 2.7 2.9 4.0 5.7 4.8
8.3 Bryant,Kobe 6.4 6.1 7.4 13.3 10.1 2.2 2.1 2.6 4.6 3.5
7.4 Wade,Dwyane 5.7 4.8 8.4 9.9 10.0 2.2 1.8 3.2 3.8 3.8
7.3 Durant,Kevin 6.0 5.9 6.4 9.6 9.7 2.0 2.0 2.2 3.2 3.3
7.0 Bosh,Chris 5.7 5.5 8.0 6.8 9.7 2.1 2.0 3.0 2.5 3.6
6.8 Roy,Brandon 5.1 5.9 6.2 9.2 8.3 1.7 2.0 2.1 3.1 2.8
6.7 Nash,Steve 4.3 5.4 7.8 9.4 8.2 1.7 2.1 3.0 3.7 3.2
6.6 Duncan,Tim 5.1 5.5 7.8 6.2 8.9 2.4 2.6 3.7 2.9 4.2
6.5 Howard,Dwight 5.1 5.3 8.0 7.4 7.5 1.9 2.0 3.0 2.8 2.9
6.4 Nowitzki,Dirk 5.5 5.4 5.7 8.1 7.9 1.9 1.9 2.0 2.9 2.8
6.2 Anthony,Carmelo 5.2 4.9 5.5 7.5 8.8 2.1 2.0 2.2 3.0 3.5
5.7 Paul,Chris 3.8 5.0 7.1 6.1 7.4 1.9 2.5 3.5 3.0 3.7
5.4 Randolph,Zach 4.6 4.7 5.8 4.9 7.3 1.7 1.7 2.1 1.8 2.7
5.3 Williams,Deron 4.1 4.3 5.4 7.2 5.9 1.6 1.7 2.1 2.8 2.3
5.1 Wallace,Gerald 4.4 5.0 5.6 5.3 5.4 1.5 1.7 1.9 1.8 1.8
5.1 Rondo,Rajon 3.6 4.6 6.3 6.4 5.1 1.5 1.9 2.6 2.6 2.1
5.0 Lee,David 4.7 5.0 5.4 3.3 7.4 1.8 1.9 2.0 1.2 2.7
4.9 Boozer,Carlos 5.1 4.3 5.0 4.7 5.7 1.9 1.6 1.9 1.8 2.1
4.8 Smith,Josh 4.0 4.1 6.7 4.0 6.0 1.7 1.7 2.8 1.7 2.5
4.7 Johnson,Joe 4.0 3.8 3.8 6.9 6.0 1.5 1.4 1.4 2.5 2.2
top 20 avg 5.0 5.2 6.7 7.7 8.0 1.9 2.0 2.6 2.9 3.0
wins-attributing system per 484 minutes (1.0 = avg)
avg player eW WS WARP WP EWA eW WS WARP WP EWA
4.7 Horford,Al 3.7 5.1 5.1 3.9 5.9 1.5 2.0 2.0 1.5 2.3
4.6 Hilario,Nene 3.5 5.0 5.5 4.0 5.5 1.3 1.9 2.1 1.5 2.1
4.4 Gasol,Marc 3.5 4.2 5.1 3.8 5.6 1.3 1.6 1.9 1.4 2.1
4.3 Iguodala,Andre 3.9 3.1 4.3 5.1 5.9 1.3 1.0 1.5 1.7 2.0
4.3 Stoudemire,Amare 4.1 4.0 4.2 3.5 5.8 1.5 1.5 1.6 1.3 2.2
4.1 Williams,Mo 3.6 4.3 3.7 5.2 4.1 1.3 1.5 1.3 1.9 1.5
4.0 Landry,Carl 3.4 4.4 4.2 2.8 5.7 1.7 2.2 2.1 1.4 2.8
4.0 Pierce,Paul 3.2 4.2 3.7 4.8 4.2 1.5 2.0 1.7 2.2 2.0
4.0 Davis,Baron 3.4 3.0 4.7 4.4 4.6 1.4 1.3 2.0 1.8 1.9
3.9 Camby,Marcus 3.1 4.0 6.3 2.8 4.3 1.4 1.8 2.9 1.3 2.0
3.9 Kidd,Jason 2.7 3.9 4.4 5.8 3.4 1.0 1.4 1.6 2.1 1.2
3.8 Bynum,Andrew 3.3 4.3 4.2 3.0 4.7 1.5 1.9 1.9 1.3 2.1
3.7 Gasol,Pau 2.9 3.9 4.7 3.0 4.5 1.8 2.4 2.9 1.8 2.7
3.6 Lopez,Brook 4.0 3.3 5.4 1.3 6.5 1.5 1.2 2.0 .5 2.4
3.6 Garnett,Kevin 3.2 3.7 3.7 3.4 3.9 1.8 2.0 2.0 1.8 2.1
3.4 Westbrook,Russel 3.5 2.4 3.9 4.0 3.6 1.3 .9 1.5 1.5 1.4
3.4 Billups,Chauncey 2.7 3.6 3.6 3.0 4.3 1.4 1.9 1.9 1.6 2.2
3.4 Kaman,Chris 4.0 2.4 2.8 3.3 4.9 1.6 .9 1.1 1.3 1.9
3.3 Arenas,Gilbert 3.7 1.8 3.6 3.7 4.6 1.5 .7 1.5 1.5 1.9
3.2 Ginobili,Manu 2.3 3.1 3.8 3.5 3.8 1.5 2.0 2.5 2.3 2.5
21-40 avg 3.4 3.7 4.3 3.7 4.8 1.5 1.6 1.9 1.6 2.1
Systems ordered by their tendency to give heavy credit to superstars, seen in the order of the 'top 20'. Presumably, the tendency reverses for 'rank and file' players.
_________________
`
36% of all statistics are wrong
Back to top
View user's profile Send private message Send e-mail
Mike G
Joined: 14 Jan 2005
Posts: 3612
Location: Hendersonville, NC
PostPosted: Mon Jan 11, 2010 10:32 am Post subject: Reply with quote
Due to large differences in the magnitudes of various systems (table above, e.g. LeBron and Kobe have less than half as many eW as WP), it's difficult to discern some equivalencies.
An alternate table showing 'rankings' of course loses the magnitudes, but those aren't directly comparable anyway. You can see that WARP says Josh Smith is twice as good as Joe Johnson, while WP has them reversed. But, are Tim Duncan's 6.2 WP 'better' than his 5.5 WS?
In this table, the 'outliers' stand out. Each system has 'em, for better and for worse.
Code:
wins system per minute wins
rk rankings: eW WS WARP WP EWA eW WS WARP WP EWA
1 James,Lebron 1 1 1 1 1 1 1 1 1 1
3 Bryant,Kobe 2 2 7 2 2 3 6 13 2 7
4 Wade,Dwyane 4 16 2 3 3 4 24 4 3 3
4 Durant,Kevin 3 3 10 4 4 7 13 16 5 8
5 Bosh,Chris 5 5 3 12 5 5 8 8 14 5
7 Duncan,Tim 8 6 5 14 6 2 2 2 9 2
8 Roy,Brandon 10 4 13 6 8 16 14 20 6 14
8 Howard,Dwight 9 9 4 9 11 8 10 5 11 10
8 Nash,Steve 15 8 6 5 9 20 7 6 4 9
8 Nowitzki,Dirk 6 7 15 7 10 9 18 22 10 13
10 Anthony,Carmelo 7 15 17 8 7 6 15 15 8 6
14 Paul,Chris 23 13 8 15 13 11 3 3 7 4
16 Randolph,Zach 13 17 14 21 14 18 26 17 26 17
16 Lee,David 12 11 19 39 12 15 23 26 49 16
16 Wallace,Gerald 14 12 16 17 26 27 28 30 25 42
rk rankings: eW WS WARP WP EWA eW WS WARP WP EWA
17 Williams,Deron 16 21 21 10 19 22 29 21 12 23
18 Smith,Josh 19 26 9 26 16 19 27 11 29 18
18 Rondo,Rajon 26 18 11 13 28 30 21 12 13 31
19 Boozer,Carlos 11 20 24 23 23 10 30 31 27 27
20 Horford,Al 24 10 22 30 18 36 11 23 31 22
21 Johnson,Joe 18 31 35 11 17 34 40 44 15 26
22 Hilario,Nene 30 14 18 27 25 42 19 19 33 30
25 Stoudemire,Amare 17 27 29 35 21 25 36 39 42 25
25 Iguodala,Andre 22 41 28 20 20 45 53 42 28 35
27 Gasol,Marc 32 24 23 31 24 44 31 29 37 29
27 Lopez,Brook 21 36 20 59 15 32 49 25 60 21
28 Williams,Mo 28 22 38 18 41 47 34 46 20 50
30 Landry,Carl 35 19 30 51 22 17 5 18 38 11
31 Camby,Marcus 43 28 12 50 38 38 25 9 44 37
32 Davis,Baron 33 45 25 25 34 37 47 27 22 39
I've cut this table off at 30, because I don't really have any Top-50 for WARP or WP. B Lopez is shown at #59 in WP, but that's of the 60 who made top-50 in (at least 2 of) eW, WS, and EWA.
_________________
`
36% of all statistics are wrong
Last edited by Mike G on Mon Jan 11, 2010 1:39 pm; edited 1 time in total
eWins, WinShares and EWA (MikeG, 2009)
eWins, WinShares and EWA (MikeG, 2009)
Last edited by Crow on Mon Apr 25, 2011 6:57 pm, edited 1 time in total.
Re: eWins, WinShares and EWA
page 2 of 2
Author Message
Jose A. Martínez
Joined: 19 Jul 2009
Posts: 72
PostPosted: Mon Jan 11, 2010 12:11 pm Post subject: Reply with quote
Excellent stuff, Mike. I am sure this is very helpful for many of us when trying to understand the differences among these similar indexes.
I have computed the Pearson (low triangle) and Spearman (upper triangle) correlations among the 5 systems, using the first 20 players ranked by avg.
Code:
eW WS WARP WP EWA
eW 1,00 0,69 0,41 0,63 0,83
WS 0,85 1,00 0,57 0,59 0,77
WARP 0,68 0,79 1,00 0,47 0,66
WP 0,79 0,82 0,69 1,00 0,72
EWA 0,90 0,87 0,80 0,82 1,00
I think in this case Spearman correlations are more reliable than Pearson correlations to study associations among these variables. It seems that EWA is very similar to eW....
Thanks a lot Mike.
_________________
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
Neil Paine
Joined: 13 Oct 2005
Posts: 774
Location: Atlanta, GA
PostPosted: Mon Jan 11, 2010 1:14 pm Post subject: Reply with quote
Interesting that despite the whole Berri-Hollinger clash a few years ago, Berri's WP correlates closest with... EWA, which is PER converted into wins.
_________________
http://www.basketball-reference.com/blog/
Back to top
View user's profile Send private message Visit poster's website
Mike G
Joined: 14 Jan 2005
Posts: 3597
Location: Hendersonville, NC
PostPosted: Mon Jan 11, 2010 1:31 pm Post subject: Reply with quote
Jose A. Martínez wrote:
... correlations among the 5 systems, using the first 20 players ranked by avg.
Code:
eW WS WARP WP EWA
eW 1,00 0,69 0,41 0,63 0,83
WS 0,85 1,00 0,57 0,59 0,77
WARP 0,68 0,79 1,00 0,47 0,66
WP 0,79 0,82 0,69 1,00 0,72
EWA 0,90 0,87 0,80 0,82 1,00
It seems that EWA is very similar to eW....
.
Odd, given that EWA gives those top 20 an avg of 8.0 wins, vs 5.0 eW. They're at opposite ends on the 'magnitude' allotted to top players.
Neil Paine wrote:
Interesting that despite the whole Berri-Hollinger clash a few years ago, Berri's WP correlates closest with... EWA, which is PER converted into wins.
This WP is from BasketballProspectus. Does that have anything to do with Berri?
_________________
`
36% of all statistics are wrong
Back to top
View user's profile Send private message Send e-mail
Jose A. Martínez
Joined: 19 Jul 2009
Posts: 72
PostPosted: Mon Jan 11, 2010 2:00 pm Post subject: Reply with quote
Quote:
Odd, given that EWA gives those top 20 an avg of 8.0 wins, vs 5.0 eW. They're at opposite ends on the 'magnitude' allotted to top players
Ok, Ok, we must be cautious about the interpretation of correlation as a measure of similiarity between variables. Correlation, per se, is an incomplete measure of "distance" between variables. A 0,9 correlation coefficient does not mean that variables are the same. Height and weight correlates about 0,8 in the population but they are completely distinct variables.
There are some measures of distance that can be more attractive that correlation. I proposed a measure called "Discriminant Distance" in a paper that has been recently published. Is a form to summarize in a single coefficient the information provided by two effect size measures: Pearson correlation and standardized mean difference. This research is available upon request:
Martínez, J. A. (2009). Discriminant distance: A new form of evaluating the distance between variables. Methodology, 5 (4), 137-144.
Other alternatives are the use of entropy, K-L divergence, etc.
_________________
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
Jose A. Martínez
Joined: 19 Jul 2009
Posts: 72
PostPosted: Mon Jan 11, 2010 2:05 pm Post subject: Reply with quote
Quote:
Berri's WP correlates closest with... EWA, which is PER converted into wins.
Neil, it would be terrific that John Hollinger explains which is the rationale of EWA, a detailed explanation would be desirable, in order to make his index more understandable
_________________
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
Neil Paine
Joined: 13 Oct 2005
Posts: 774
Location: Atlanta, GA
PostPosted: Mon Jan 11, 2010 3:11 pm Post subject: Reply with quote
Mike G wrote:
This WP is from BasketballProspectus. Does that have anything to do with Berri?
Oh, I guess that's Brad Doolittle's Wins Produced then. I think the assumption is anytime someone mentions WP or Wins Produced, it's Berri's system, but I guess there's more than one metric out there called that.
Jose A. Martínez wrote:
Neil, it would be terrific that John Hollinger explains which is the rationale of EWA, a detailed explanation would be desirable, in order to make his index more understandable
He talked about it at length here:
http://sports.espn.go.com/nba/columns/s ... iem-090325
Not sure if that's Insider or not... Basically, it converts PER from a per-minute efficiency metric to a value metric by subtracting a positional replacement level and using a PER-to-wins converter. I guess the rationale is that people were complaining that PER is a pure rate stat and didn't take into account the fact that a guy with PER 20 in 2600 MP is probably more valuable than a guy with PER 22 in 1800 MP. Not to put words in JH's mouth or anything.
_________________
http://www.basketball-reference.com/blog/
Back to top
View user's profile Send private message Visit poster's website
Jose A. Martínez
Joined: 19 Jul 2009
Posts: 72
PostPosted: Mon Jan 11, 2010 4:10 pm Post subject: Reply with quote
Quote:
Oh, I guess that's Brad Doolittle's Wins Produced then. I think the assumption is anytime someone mentions WP or Wins Produced, it's Berri's system, but I guess there's more than one metric out there called that.
Neil, in their Basketball Prospectus 20092010 book, Doolittle and Pelton do not speak about Berri's WP system. Then, we have to assume that both metrics are different, but again, Doolittle and Pelton do not explain with detail their "Wins Produced"....
Quote:
He talked about it at length here:
http://sports.espn.go.com/nba/columns/s ... iem-090325
Many thanks for the link!!
_________________
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
Mike G
Joined: 14 Jan 2005
Posts: 3597
Location: Hendersonville, NC
PostPosted: Mon Feb 15, 2010 10:37 am Post subject: Reply with quote
In another thread ("Most Improved Players"), Neil wrote:
Quote:
Philosophically, WS is trying to say, "Given the actual circumstances this player played in, how many wins did he generate?" And eWins seems to be asking, "If this player played in neutral circumstances, how many wins would he have generated to the average team?" ...
The criticism of WS I hear often is that it's too dependent on context, that a player's defensive rating doesn't really capture his ability as much as it does his team's ability...
...
Then again, the ability to distinguish "value" from "ability" is a consistent blind spot with eWins -- wins are, by definition, a measure of pure value, that which happened in the past. But any metric which ignores context and attempts to minimize year-to-year fluctuation is philosophically attempting to measure "ability", not value...
So I guess I'm once again calling for clarification about the aims and intentions of the various win estimators. And until the realization is made that context-neutral ability is very different from context-dependent value, I suppose I'm also asking people to stop conflating metrics that have extremely different philosophical bases simply because they both have the word "wins" in their name.
Is it a given that "wins are, by definition, a measure of pure value"?
Or that eWins address a player's "ability", while Win Shares somehow gauge his "value"?
Let's look at a team that had some roster stability over a period of great instability in the Wins department: the Spurs of 1996, '97, '98.
In '97, Mr. Robinson only played 6 games. The team dropped from 59 wins to 20; bouncing back to 56 wins with the '98 addition of Mr. Duncan.
The Spurs had 8 players, other than DRob, who were on the team in '96 and still there in '98.
According to PER, eWins, and Win Shares, these guys were lesser players in '98 than in '96: Vinnie Del Negro, Sean Elliott, Avery Johnson, Chuck Person.
Another 4 players -- Will Perdue, Cory Alexander, Carl Herrera, and Monty Williams -- were there for the whole ordeal, too.
eWins and Win Shares are accompanied by per484 minute rates (1.00 = avg).
Code:
1996 1997 1998
Spurs eW e484 eW e484 eW e484
Robinson 17.7 2.83 .9 2.90 13.9 2.74
Avery J 7.0 1.10 4.5 .87 5.2 .94
Elliott 9.7 1.62 2.9 1.00 1.3 .64
Del Negro 7.2 1.27 4.5 .98 2.6 .72
C Person 4.8 1.09 dnp ... 1.3 .43
Perdue 1.3 .46 2.7 .68 2.1 .67
Alexander .1 .05 2.7 .90 .6 .61
Herrera -.6 -.71 1.4 .36 .2 .16
Monty W -.1 -.36 2.6 .95 1.7 .64
non-Rob 29.5 1.07 21.3 .81 15.0 .68
1996 1997 1998
Spurs Min PER Min PER Min PER
Robinson 3019 29.4 147 31.0* 2457 27.8
Avery J 3084* 16.3* 2472 15.0 2674 14.1
Elliott 2901 16.3 1393 14.0 1012 9.5
Del Negro 2766* 15.6* 2243 14.4 1721 11.1
C Person 2131 13.6 dnp ... 1455 8.3
Perdue 1396 13.7 1918* 16.7* 1491 14.8
Alexander 560 8.7 1454* 14.3 501 11.1*
Herrera 393 4.7 1837* 10.6 516 8.1
Monty W 122 8.7 1345 15.6* 1314 11.3
non-Rob 13353 14.7 12662 14.4 10684 11.7
1996 1997 1998
Spurs WS w484 WS w484 WS w484
Robinson 18.3 2.93 .7 2.30 13.8 2.72
Avery J 9.1 1.43 3.4 .67 6.7 1.21
Elliott 7.7 1.28 1.5 .52 1.4 .67
Del Negro 8.3 1.45 3.1 .67 3.7 1.04
C Person 5.5 1.25 dnp ... 2.1 .70
Perdue 3.4 1.18 4.4 1.11 4.7 1.53
Alexander .2 .17 1.1 .37 .7 .68
Herrera -.2 -.25 .1 .03 .1 .09
Monty W .0 .00 1.2 .43 2.3 .85
non-Rob 34.0 1.23 14.8 .57 21.7 .98
Listed players 'other than Robinson' are summed (non-Rob) at the bottom of each stat block.
According to PER and eWins, the top 4 non-Robs declined steadily and/or precipitously each year; and the next 4 all stepped up to the challenge of '97, all coming up with career-high (*) PER or minutes (or both) that year.
Win Shares suggests all 8 players created more Value in '98 than in '97; by WS, WS/Min, or both.
While there's an argument that Value = Wins, there's probably a stronger argument that better production * more minutes = more Value.
Author Message
Jose A. Martínez
Joined: 19 Jul 2009
Posts: 72
PostPosted: Mon Jan 11, 2010 12:11 pm Post subject: Reply with quote
Excellent stuff, Mike. I am sure this is very helpful for many of us when trying to understand the differences among these similar indexes.
I have computed the Pearson (low triangle) and Spearman (upper triangle) correlations among the 5 systems, using the first 20 players ranked by avg.
Code:
eW WS WARP WP EWA
eW 1,00 0,69 0,41 0,63 0,83
WS 0,85 1,00 0,57 0,59 0,77
WARP 0,68 0,79 1,00 0,47 0,66
WP 0,79 0,82 0,69 1,00 0,72
EWA 0,90 0,87 0,80 0,82 1,00
I think in this case Spearman correlations are more reliable than Pearson correlations to study associations among these variables. It seems that EWA is very similar to eW....
Thanks a lot Mike.
_________________
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
Neil Paine
Joined: 13 Oct 2005
Posts: 774
Location: Atlanta, GA
PostPosted: Mon Jan 11, 2010 1:14 pm Post subject: Reply with quote
Interesting that despite the whole Berri-Hollinger clash a few years ago, Berri's WP correlates closest with... EWA, which is PER converted into wins.
_________________
http://www.basketball-reference.com/blog/
Back to top
View user's profile Send private message Visit poster's website
Mike G
Joined: 14 Jan 2005
Posts: 3597
Location: Hendersonville, NC
PostPosted: Mon Jan 11, 2010 1:31 pm Post subject: Reply with quote
Jose A. Martínez wrote:
... correlations among the 5 systems, using the first 20 players ranked by avg.
Code:
eW WS WARP WP EWA
eW 1,00 0,69 0,41 0,63 0,83
WS 0,85 1,00 0,57 0,59 0,77
WARP 0,68 0,79 1,00 0,47 0,66
WP 0,79 0,82 0,69 1,00 0,72
EWA 0,90 0,87 0,80 0,82 1,00
It seems that EWA is very similar to eW....
.
Odd, given that EWA gives those top 20 an avg of 8.0 wins, vs 5.0 eW. They're at opposite ends on the 'magnitude' allotted to top players.
Neil Paine wrote:
Interesting that despite the whole Berri-Hollinger clash a few years ago, Berri's WP correlates closest with... EWA, which is PER converted into wins.
This WP is from BasketballProspectus. Does that have anything to do with Berri?
_________________
`
36% of all statistics are wrong
Back to top
View user's profile Send private message Send e-mail
Jose A. Martínez
Joined: 19 Jul 2009
Posts: 72
PostPosted: Mon Jan 11, 2010 2:00 pm Post subject: Reply with quote
Quote:
Odd, given that EWA gives those top 20 an avg of 8.0 wins, vs 5.0 eW. They're at opposite ends on the 'magnitude' allotted to top players
Ok, Ok, we must be cautious about the interpretation of correlation as a measure of similiarity between variables. Correlation, per se, is an incomplete measure of "distance" between variables. A 0,9 correlation coefficient does not mean that variables are the same. Height and weight correlates about 0,8 in the population but they are completely distinct variables.
There are some measures of distance that can be more attractive that correlation. I proposed a measure called "Discriminant Distance" in a paper that has been recently published. Is a form to summarize in a single coefficient the information provided by two effect size measures: Pearson correlation and standardized mean difference. This research is available upon request:
Martínez, J. A. (2009). Discriminant distance: A new form of evaluating the distance between variables. Methodology, 5 (4), 137-144.
Other alternatives are the use of entropy, K-L divergence, etc.
_________________
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
Jose A. Martínez
Joined: 19 Jul 2009
Posts: 72
PostPosted: Mon Jan 11, 2010 2:05 pm Post subject: Reply with quote
Quote:
Berri's WP correlates closest with... EWA, which is PER converted into wins.
Neil, it would be terrific that John Hollinger explains which is the rationale of EWA, a detailed explanation would be desirable, in order to make his index more understandable
_________________
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
Neil Paine
Joined: 13 Oct 2005
Posts: 774
Location: Atlanta, GA
PostPosted: Mon Jan 11, 2010 3:11 pm Post subject: Reply with quote
Mike G wrote:
This WP is from BasketballProspectus. Does that have anything to do with Berri?
Oh, I guess that's Brad Doolittle's Wins Produced then. I think the assumption is anytime someone mentions WP or Wins Produced, it's Berri's system, but I guess there's more than one metric out there called that.
Jose A. Martínez wrote:
Neil, it would be terrific that John Hollinger explains which is the rationale of EWA, a detailed explanation would be desirable, in order to make his index more understandable
He talked about it at length here:
http://sports.espn.go.com/nba/columns/s ... iem-090325
Not sure if that's Insider or not... Basically, it converts PER from a per-minute efficiency metric to a value metric by subtracting a positional replacement level and using a PER-to-wins converter. I guess the rationale is that people were complaining that PER is a pure rate stat and didn't take into account the fact that a guy with PER 20 in 2600 MP is probably more valuable than a guy with PER 22 in 1800 MP. Not to put words in JH's mouth or anything.
_________________
http://www.basketball-reference.com/blog/
Back to top
View user's profile Send private message Visit poster's website
Jose A. Martínez
Joined: 19 Jul 2009
Posts: 72
PostPosted: Mon Jan 11, 2010 4:10 pm Post subject: Reply with quote
Quote:
Oh, I guess that's Brad Doolittle's Wins Produced then. I think the assumption is anytime someone mentions WP or Wins Produced, it's Berri's system, but I guess there's more than one metric out there called that.
Neil, in their Basketball Prospectus 20092010 book, Doolittle and Pelton do not speak about Berri's WP system. Then, we have to assume that both metrics are different, but again, Doolittle and Pelton do not explain with detail their "Wins Produced"....
Quote:
He talked about it at length here:
http://sports.espn.go.com/nba/columns/s ... iem-090325
Many thanks for the link!!
_________________
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
Mike G
Joined: 14 Jan 2005
Posts: 3597
Location: Hendersonville, NC
PostPosted: Mon Feb 15, 2010 10:37 am Post subject: Reply with quote
In another thread ("Most Improved Players"), Neil wrote:
Quote:
Philosophically, WS is trying to say, "Given the actual circumstances this player played in, how many wins did he generate?" And eWins seems to be asking, "If this player played in neutral circumstances, how many wins would he have generated to the average team?" ...
The criticism of WS I hear often is that it's too dependent on context, that a player's defensive rating doesn't really capture his ability as much as it does his team's ability...
...
Then again, the ability to distinguish "value" from "ability" is a consistent blind spot with eWins -- wins are, by definition, a measure of pure value, that which happened in the past. But any metric which ignores context and attempts to minimize year-to-year fluctuation is philosophically attempting to measure "ability", not value...
So I guess I'm once again calling for clarification about the aims and intentions of the various win estimators. And until the realization is made that context-neutral ability is very different from context-dependent value, I suppose I'm also asking people to stop conflating metrics that have extremely different philosophical bases simply because they both have the word "wins" in their name.
Is it a given that "wins are, by definition, a measure of pure value"?
Or that eWins address a player's "ability", while Win Shares somehow gauge his "value"?
Let's look at a team that had some roster stability over a period of great instability in the Wins department: the Spurs of 1996, '97, '98.
In '97, Mr. Robinson only played 6 games. The team dropped from 59 wins to 20; bouncing back to 56 wins with the '98 addition of Mr. Duncan.
The Spurs had 8 players, other than DRob, who were on the team in '96 and still there in '98.
According to PER, eWins, and Win Shares, these guys were lesser players in '98 than in '96: Vinnie Del Negro, Sean Elliott, Avery Johnson, Chuck Person.
Another 4 players -- Will Perdue, Cory Alexander, Carl Herrera, and Monty Williams -- were there for the whole ordeal, too.
eWins and Win Shares are accompanied by per484 minute rates (1.00 = avg).
Code:
1996 1997 1998
Spurs eW e484 eW e484 eW e484
Robinson 17.7 2.83 .9 2.90 13.9 2.74
Avery J 7.0 1.10 4.5 .87 5.2 .94
Elliott 9.7 1.62 2.9 1.00 1.3 .64
Del Negro 7.2 1.27 4.5 .98 2.6 .72
C Person 4.8 1.09 dnp ... 1.3 .43
Perdue 1.3 .46 2.7 .68 2.1 .67
Alexander .1 .05 2.7 .90 .6 .61
Herrera -.6 -.71 1.4 .36 .2 .16
Monty W -.1 -.36 2.6 .95 1.7 .64
non-Rob 29.5 1.07 21.3 .81 15.0 .68
1996 1997 1998
Spurs Min PER Min PER Min PER
Robinson 3019 29.4 147 31.0* 2457 27.8
Avery J 3084* 16.3* 2472 15.0 2674 14.1
Elliott 2901 16.3 1393 14.0 1012 9.5
Del Negro 2766* 15.6* 2243 14.4 1721 11.1
C Person 2131 13.6 dnp ... 1455 8.3
Perdue 1396 13.7 1918* 16.7* 1491 14.8
Alexander 560 8.7 1454* 14.3 501 11.1*
Herrera 393 4.7 1837* 10.6 516 8.1
Monty W 122 8.7 1345 15.6* 1314 11.3
non-Rob 13353 14.7 12662 14.4 10684 11.7
1996 1997 1998
Spurs WS w484 WS w484 WS w484
Robinson 18.3 2.93 .7 2.30 13.8 2.72
Avery J 9.1 1.43 3.4 .67 6.7 1.21
Elliott 7.7 1.28 1.5 .52 1.4 .67
Del Negro 8.3 1.45 3.1 .67 3.7 1.04
C Person 5.5 1.25 dnp ... 2.1 .70
Perdue 3.4 1.18 4.4 1.11 4.7 1.53
Alexander .2 .17 1.1 .37 .7 .68
Herrera -.2 -.25 .1 .03 .1 .09
Monty W .0 .00 1.2 .43 2.3 .85
non-Rob 34.0 1.23 14.8 .57 21.7 .98
Listed players 'other than Robinson' are summed (non-Rob) at the bottom of each stat block.
According to PER and eWins, the top 4 non-Robs declined steadily and/or precipitously each year; and the next 4 all stepped up to the challenge of '97, all coming up with career-high (*) PER or minutes (or both) that year.
Win Shares suggests all 8 players created more Value in '98 than in '97; by WS, WS/Min, or both.
While there's an argument that Value = Wins, there's probably a stronger argument that better production * more minutes = more Value.