Why I'm not a WP fan

[floppymoose]

I suppose you could always write a good article on the subject and point Gladwell at it. Everyone makes mistakes. Maybe he just needs a push.

[Doctor MJ]

I suppose you could always right a good article on the subject and point Gladwell at it. Everyone makes mistakes. Maybe he just needs a push.

Maybe I will floppy. It's a good idea.

Anyone aware of good articles already written on Berri's folly? I know of the Rosenbaum paper. Anything else?

[sp6r=underrated]

I don't think being radical is a significant factor for his appeal amongst the Gladwell types. For them Berri's work has significantly more credibility because it has been published in peer-reviewed journals.

Oh god. Yeah, THAT aspect too.

Gladwell bothers me so much because when he talks on other subjects, he makes a ton of sense to me. Yet, he comes to an area that I really know a lot about, and he's utterly duped by a hack with an academic backing. The unfairness to other basketball analysts is one thing...but how am I supposed to take seriously Gladwell's research in any other area when he has such analytical vertigo?

I have the same feeling when I read Matt Yglesias's writings on basketball. I generally enjoy his political commentary even when I don't agree with his conclusion because of his analysis. Yet, I noticed the only stat he regularly cites is wins produced and discusses it in an extremely uncritical manner.

I'm convinced it's the peer review issue that hooks Matt in. It makes me wonder about other topics Yglesias writes about.

[sp6r=underrated]

As if on cue, Yglesias on hoops

The idea of a system like Dave Berri’s “wins produced” is that if you add up all the “wins produced” of the individual 2009-2010 Los Angeles Lakers you get a number that’s approximately equal to the total wins of the Los Angeles Lakers. People can (and have, and do, and should continue to) raise questions about whether the Berri formula is accurately allocating credit for these wins to individuals, and also can (and have, and do, and should continue to) raise questions about the predictive value of these quantities. But there’s no question of what’s being measured.

Does Berri even consider value over replacement?

[ElGee]

I don't think being radical is a significant factor for his appeal amongst the Gladwell types. For them Berri's work has significantly more credibility because it has been published in peer-reviewed journals.

Oh god. Yeah, THAT aspect too.

Gladwell bothers me so much because when he talks on other subjects, he makes a ton of sense to me. Yet, he comes to an area that I really know a lot about, and he's utterly duped by a hack with an academic backing. The unfairness to other basketball analysts is one thing...but how am I supposed to take seriously Gladwell's research in any other area when he has such analytical vertigo?

I love Gladwell the writer. He's a great story teller too. But he does a lot of pseudo-science ("social scientist?"). I think I've said this before on realgm, but he tries to go more than an inch deep into esoteric topics and that's a tricky venture.

As for WP, here's the way I understand the stat w a quick marginal value example: http://elgee35.wordpress.com/2011/01/24 ... nsProduced

So it makes sense Love is the best player in the league. Frankly, I think it's a horrible stat, and echo Doc's sentiments about the bizarre behavior of people who cling to it. From a "scientific" point of view, it's weird to me that Berri can't acknowledge the problems with it (or even understand why there would be problems in the first place).

[mysticbb]

As for WP, here's the way I understand the stat w a quick marginal value example: http://elgee35.wordpress.com/2011/01/24 ... nsProduced

You are rasing a good point here. For Berri a scorer breaks even, if he shoots 50 fg% for 2pt field goals or 33.3 % for 3pt field goals or 46.9 FT%. Yes, somehow someone can make his free throws below that 50% threshold while the league average is currently at 76.3 FT%, but he has to score above average on 2pt field goals (currently at 48.5%).

But the major concern here are the rebounding number and those would reflect even more, if it weren't for the positional adjustment (he is using value above league average for position, sp6r=underrated). He gives an offensive rebound as much value as a 2pt field goal (well, a 3pt shoot is worth double the amount of 2pt field goal and making two free throws is worth like scoring 2.125 with a 2pt shot). Now, imagine someone gets his own miss, he would get still the value of the 2pt shot (offensive rebound and the miss are cancelling each other out). Sounds fair. But if someone takes a shot while his teammate makes the tip in, Berri gives out the value of TWO 2pt field goals made, while punishing the other guy with the value of a 2pt field goal. Now, we all now that someone has to take the shot to avoid a turnover. The shooter also draws attention away from the basket, which can give the rebounder more room to go for that rebound. In reality there is cause for that offensive rebound + made field goal, in Berri's calculation there isn't. A player will not get any credit for shot creation, if he is not credited with an assist, even though is partly responsible for the team score even with that missed shot. That a missed shot means at least a turnover less, isn't in the calculation somewhere, but in the way Berri is calculating the scoring, he needs to include this. Well, and when someone turns the ball over, it is also spread around the team via the various team adjustments Berri makes, but that is a different story.

For a defensive rebound Berri gives out more credit than for a 2pt field goal. The rebounder gets full credit for the rebound and later again via the team defensive adjustment. What Berri doesn't take into account is the fact that in average 73.7% of the rebounds are defensive rebounds anyway. There are dimishing returns. When you play with 5 great rebounders and the other teams has 5 weak rebounders, they will still not rebound more than 100% of the opponents missed shots. He is somehow accounting for that via the positional adjustment, but that means in the end that a rebound by a guard is worth more than a rebound from a center or power forward. That is an interesting thing here, because in reality the big guy in the middle is doing the work regardless whether he gets the board or not. A prime example would be the mid 00's Nets. Kidd was getting rebounds a lot. We know that he is a great rebounder, but that was also by design. Kidd got especially more rebounds on the defensive end when he played together with one specific forward/center, Jason Collins. Collins job was it to boxout (that and playing man-to-man defense under the basket). It makes a lot of sense, because you want the ball in the hand of your point guard as soon as possible, especially when a guy like Kidd needs the open court to be really efficient as a playmaker. The poor Collins didn't create any wins for the Nets according to Berri, he in fact is credited with "losses". But in reality the Nets played always better with Collins than without him. They even dropped below average in rebounding without him on the court. Collins is a +6 defensive adjusted PM guy from 2003 to 2009 (the average value from Iliardi).

Well, every metric will underrate a guy like Collins, that is true, but even with my system and my underrating him, he still should get a couple of millions for playing basketball. Berri concludes that Collins shouldn't get any money and he should not even be on the court. It is not just a hypothesis which would needed to be tested, no he concludes this by using that metric. Because Collins doesn't grab as many rebounds as an average center, he gets punished, even though his work was the reason for the defensive of the Nets. Here is a basic example why Berri also completely ignores the defensive aspect of the game in his metric on an individual level. Claiming he would do it, is just a straight damn lie.

The best rebounders, going by an adjusted rebounding value recently someone posted on the APBR forum, suggest, that he gives the team 1.6 rebounds per game more than an average replacement. That is the impact of a Kevin Love due to his rebounding. The values are rather slim, because teams are getting their defensive rebounds most of the time anyway due to their positioning on defense.

Thus Berri overvalues rebounding by lot while, as you pointed out, there is no negative rebounding impact for an individual player. The missed rebounds aren't included.

From a "scientific" point of view, it's weird to me that Berri can't acknowledge the problems with it (or even understand why there would be problems in the first place).

Because Berri thinks his rating has some "scientific" value. That's what a lot of economy guys think. They are using regressions and making conclusions out of the results. While they don't need to really prove that their is also a causality. Especailly when it comes to sport where a lot of "amateurs" know just more about it than those "scientists". (Well, let me say it straight, economy has imho nothing to do with science at all, I'm a geophysicist and if I would base my conclusions on some weak correlation coefficients, I could look for another job).

Regarding the discussion about Gladwell I read some comments by a guy name Steven Pinker. Pinker is a Harvard professor, and he points out some obvious flaws. For example Berri once made a study about NFl quarterbacks and his basic solution was that there is a unfounded bias to choose the higher picked QB over the lower pick. Well, Berri basically suggested that each QB should get an equal chance. Pinker pointed out that this idea has a logical flaw, because the value of an assessment is that it is less costly than trial&error. Berri wants the NFL teams to choose their quarterbacks for the season after the season is already over and the teams already know how each quarterback performed in game situations. A similar thing happens with NBA players here. Berri is using his rather high correlation to make conclusions about players, but he is using the current season data to make a prediction about the current season. Well, great job.

His rating has value to determine an outcome of a playoffs series, but it is not doing a better job than point differential. What Berri saves his life is the short selection each coach has to setup a lineup. He needs some of the bigs, he needs the guards and the playing time is distributed accordingly and most times rather consistent for the top players anyway. We have around 300 not-injured players overall who are more or less not just garbage time players. That makes basically 2 for each position for each team. Well, you can distribute the minutes via a coin flip and you will have basically as good as a chance to get the right lineup as Berri (as I showed in my analysis). What we are facing here is a sampling bias. Berri can use already choosen players, who are picked by coaches and scouts due to their own analysis and concludes that those coaches and scouts are dumb, because they aren't distributing the minutes right based on his metric. But he doesn't show (neither can he) that with his suggested minutes distribution the teams would have more success. In fact he is using league average values to determine his marginal values. Do you know what that means? Even with Berri's minutes distribution the result of his correlation analysis wouldn't change a bit. He can't even prove his claim in any sort or way.
Why does his metric is still doing a rather good job for predicting the future (even though it is the weakest tested metric by Rosenbaum&Lewin (2007))? Because most times the players with major minutes are those with major minutes again next season in a similar role. He real test would be specific examples of coaching and player movements in which his metric is doing better than others. Well, he predicted with the current minutes distribution on the Warriors that they should have 32.1 wins right now. They have 10 less. How many wins would my rating predict by using the ratings from last season? 23.0! Their expected wins right now are 20. Based on WS48 we get 18.4 by using last years numbers for all players who played last season and this years numbers for the rest (I've done the same for WP48 and my rating). Now, Berri is again way off while WS48 and my rating are doing a rather good job at predicting the success of the current Warriors by using old data of the players.

But NOBODY in the scientific world would do such a study, because the cases in which you have heavy player movement and coaching changes are rather small, especially those where the minutes distribution changes heavily also due to injuries. Berri can always counter with some excuses and the small sample size.

So, overall we have a case where Berri doesn't have to prove his claims (nor can he really do this)to the scientific community, nobody will test his hypothesis and a sampling bias which helps him for his correlation. There is no causality involved in his metric on an individual player basis, but nobody in the scientific community can really show that without using +/- minutes numbers, which can be refute due to their inconsistency in short term analysises.

That all is well known and understood among teams, that is the reason no team would make a decision for a lineup or for aquiring a player based on WP48. They can use other boxscore metrics which are also stable and are more consistent with the result on the individual player level. Berri can call those decision makers dumb, but he is evaluating things in hindsight. Well, that is something everyone can do.

[floppymoose]

Well, every metric will underrate a guy like Collins[...]

But you just quoted the one that didn't: +-

I think player evaluation metrics should be largely built around team performance rather than individual box score numbers. When I am trying to figure out if a player is a good rebounder, for instance, I look at how they affect the team rebounding percentages in addition to the percentage of available boards they get.

[Vinsanity420]

Damn, Mystic or Tsherkin, who can put up the longer wall of text?

[Idunkon1stdates]

As for WP, here's the way I understand the stat w a quick marginal value example: http://elgee35.wordpress.com/2011/01/24 ... nsProduced

You are rasing a good point here. For Berri a scorer breaks even, if he shoots 50 fg% for 2pt field goals or 33.3 % for 3pt field goals or 46.9 FT%. Yes, somehow someone can make his free throws below that 50% threshold while the league average is currently at 76.3 FT%, but he has to score above average on 2pt field goals (currently at 48.5%).

But the major concern here are the rebounding number and those would reflect even more, if it weren't for the positional adjustment (he is using value above league average for position, sp6r=underrated). He gives an offensive rebound as much value as a 2pt field goal (well, a 3pt shoot is worth double the amount of 2pt field goal and making two free throws is worth like scoring 2.125 with a 2pt shot). Now, imagine someone gets his own miss, he would get still the value of the 2pt shot (offensive rebound and the miss are cancelling each other out). Sounds fair. But if someone takes a shot while his teammate makes the tip in, Berri gives out the value of TWO 2pt field goals made, while punishing the other guy with the value of a 2pt field goal. Now, we all now that someone has to take the shot to avoid a turnover. The shooter also draws attention away from the basket, which can give the rebounder more room to go for that rebound. In reality there is cause for that offensive rebound + made field goal, in Berri's calculation there isn't. A player will not get any credit for shot creation, if he is not credited with an assist, even though is partly responsible for the team score even with that missed shot. That a missed shot means at least a turnover less, isn't in the calculation somewhere, but in the way Berri is calculating the scoring, he needs to include this. Well, and when someone turns the ball over, it is also spread around the team via the various team adjustments Berri makes, but that is a different story.

For a defensive rebound Berri gives out more credit than for a 2pt field goal. The rebounder gets full credit for the rebound and later again via the team defensive adjustment. What Berri doesn't take into account is the fact that in average 73.7% of the rebounds are defensive rebounds anyway. There are dimishing returns. When you play with 5 great rebounders and the other teams has 5 weak rebounders, they will still not rebound more than 100% of the opponents missed shots. He is somehow accounting for that via the positional adjustment, but that means in the end that a rebound by a guard is worth more than a rebound from a center or power forward. That is an interesting thing here, because in reality the big guy in the middle is doing the work regardless whether he gets the board or not. A prime example would be the mid 00's Nets. Kidd was getting rebounds a lot. We know that he is a great rebounder, but that was also by design. Kidd got especially more rebounds on the defensive end when he played together with one specific forward/center, Jason Collins. Collins job was it to boxout (that and playing man-to-man defense under the basket). It makes a lot of sense, because you want the ball in the hand of your point guard as soon as possible, especially when a guy like Kidd needs the open court to be really efficient as a playmaker. The poor Collins didn't create any wins for the Nets according to Berri, he in fact is credited with "losses". But in reality the Nets played always better with Collins than without him. They even dropped below average in rebounding without him on the court. Collins is a +6 defensive adjusted PM guy from 2003 to 2009 (the average value from Iliardi).

Well, every metric will underrate a guy like Collins, that is true, but even with my system and my underrating him, he still should get a couple of millions for playing basketball. Berri concludes that Collins shouldn't get any money and he should not even be on the court. It is not just a hypothesis which would needed to be tested, no he concludes this by using that metric. Because Collins doesn't grab as many rebounds as an average center, he gets punished, even though his work was the reason for the defensive of the Nets. Here is a basic example why Berri also completely ignores the defensive aspect of the game in his metric on an individual level. Claiming he would do it, is just a straight damn lie.

The best rebounders, going by an adjusted rebounding value recently someone posted on the APBR forum, suggest, that he gives the team 1.6 rebounds per game more than an average replacement. That is the impact of a Kevin Love due to his rebounding. The values are rather slim, because teams are getting their defensive rebounds most of the time anyway due to their positioning on defense.

Thus Berri overvalues rebounding by lot while, as you pointed out, there is no negative rebounding impact for an individual player. The missed rebounds aren't included.

From a "scientific" point of view, it's weird to me that Berri can't acknowledge the problems with it (or even understand why there would be problems in the first place).

Because Berri thinks his rating has some "scientific" value. That's what a lot of economy guys think. They are using regressions and making conclusions out of the results. While they don't need to really prove that their is also a causality. Especailly when it comes to sport where a lot of "amateurs" know just more about it than those "scientists". (Well, let me say it straight, economy has imho nothing to do with science at all, I'm a geophysicist and if I would base my conclusions on some weak correlation coefficients, I could look for another job).

Regarding the discussion about Gladwell I read some comments by a guy name Steven Pinker. Pinker is a Harvard professor, and he points out some obvious flaws. For example Berri once made a study about NFl quarterbacks and his basic solution was that there is a unfounded bias to choose the higher picked QB over the lower pick. Well, Berri basically suggested that each QB should get an equal chance. Pinker pointed out that this idea has a logical flaw, because the value of an assessment is that it is less costly than trial&error. Berri wants the NFL teams to choose their quarterbacks for the season after the season is already over and the teams already know how each quarterback performed in game situations. A similar thing happens with NBA players here. Berri is using his rather high correlation to make conclusions about players, but he is using the current season data to make a prediction about the current season. Well, great job.

His rating has value to determine an outcome of a playoffs series, but it is not doing a better job than point differential. What Berri saves his life is the short selection each coach has to setup a lineup. He needs some of the bigs, he needs the guards and the playing time is distributed accordingly and most times rather consistent for the top players anyway. We have around 300 not-injured players overall who are more or less not just garbage time players. That makes basically 2 for each position for each team. Well, you can distribute the minutes via a coin flip and you will have basically as good as a chance to get the right lineup as Berri (as I showed in my analysis). What we are facing here is a sampling bias. Berri can use already choosen players, who are picked by coaches and scouts due to their own analysis and concludes that those coaches and scouts are dumb, because they aren't distributing the minutes right based on his metric. But he doesn't show (neither can he) that with his suggested minutes distribution the teams would have more success. In fact he is using league average values to determine his marginal values. Do you know what that means? Even with Berri's minutes distribution the result of his correlation analysis wouldn't change a bit. He can't even prove his claim in any sort or way.
Why does his metric is still doing a rather good job for predicting the future (even though it is the weakest tested metric by Rosenbaum&Lewin (2007))? Because most times the players with major minutes are those with major minutes again next season in a similar role. He real test would be specific examples of coaching and player movements in which his metric is doing better than others. Well, he predicted with the current minutes distribution on the Warriors that they should have 32.1 wins right now. They have 10 less. How many wins would my rating predict by using the ratings from last season? 23.0! Their expected wins right now are 20. Based on WS48 we get 18.4 by using last years numbers for all players who played last season and this years numbers for the rest (I've done the same for WP48 and my rating). Now, Berri is again way off while WS48 and my rating are doing a rather good job at predicting the success of the current Warriors by using old data of the players.

But NOBODY in the scientific world would do such a study, because the cases in which you have heavy player movement and coaching changes are rather small, especially those where the minutes distribution changes heavily also due to injuries. Berri can always counter with some excuses and the small sample size.

So, overall we have a case where Berri doesn't have to prove his claims (nor can he really do this)to the scientific community, nobody will test his hypothesis and a sampling bias which helps him for his correlation. There is no causality involved in his metric on an individual player basis, but nobody in the scientific community can really show that without using +/- minutes numbers, which can be refute due to their inconsistency in short term analysises.

That all is well known and understood among teams, that is the reason no team would make a decision for a lineup or for aquiring a player based on WP48. They can use other boxscore metrics which are also stable and are more consistent with the result on the individual player level. Berri can call those decision makers dumb, but he is evaluating things in hindsight. Well, that is something everyone can do.

So basically,

[lorak]

WP = Wins Produced, as per the book Wages of Wins and the associated website http://dberri.wordpress.com/

Those who have discussed nba stat with me before will know I'm much more a fan of metrics that measure on/off court distinctions, than I am of metrics that are purely box score based.

WP is box score based, and as such is flawed in my opinion. It's difficult to prove this to the WP crowd, though. Finally a specific prediction based on WP came up that I think will reveal the problem: David Lee's trade to the Warriors and how that will impact their record. WP measurements suggest that his presence alone, holding other factors constant (such as the rest of the team being injured, Biedrins suddenly sucking, no improvement from young players, etc), will lead to *14* more wins by the Warriors.

There is an article on this here:
http://dberri.wordpress.com/2010/07/14/ ... on-50-wins

The article goes on to make some not unreasonable assumptions about the state of the team that gets the WP prediction all the way up to 56.9 wins.

As the best case scenario - so no injuries, minutes allocations as assumed and so on.
Now we know that Warriors had several injuries to key players - first Biedrins wasn't fully recovery, Lee misseed almost 10 games and Curry is all the time bothered by injuries (and also missed games).

It also worth to note that this prediction (56.9) is a aberration. There are other predictions based on WP and most of them agree that Warriors would be below 50 wins, for example: http://thenbehteam.blogspot.com/2010/10 ... tions.html
49 wins isn't far off if we keep in mind injuries.
http://nerdnumbers.com/archives/859
50 wins


It would be also nice if some of you would provide other predictions, for comparison sake, because it's always easy to criticize but how many predictions are better than these based on WP?

[mysticbb]

It would be also nice if some of you would provide other predictions, for comparison sake, because it's always easy to criticize but how many predictions are better than these based on WP?

I wrote already that with the current minutes distribution on the Warriors they should get 32.1 wins so far this season, if we use last seasons data for all players possible and this seasons data for the rest. Doing the same with my rating gives me 23 wins and using WS48 is giving out 18.4 wins. Well, we have to add that we are basically doing similar stuff by trying to reproduce point differentials, thus we should look at the expected wins so far and not on the real wins. Well, until now the Warriors are expected to have 20 wins, which puts WS48 on top and WP48 way off.

And that is the real test here. Everything else is just excuses.

Well, in fact we can make a lineup analysis and we find that for the Top50 lineups in terms of minutes played WP48 is doing the worst job at predicting their point differential. That basically means WP48 is worth less when it comes down to evaluating individual players or basing coaching decisions on it. Not surprising at all, because who would pick a lineup of Reggie Evans, Kris Humphries, Matt Barnes, Landry Fields and Jose Calderon over a lineup with Tim Duncan, Dirk Nowitzki, Paul Pierce, Ray Allen and Derrick Rose? The first lineup has 1.509 WP48 and is nearly twice as good as the 2nd 0.857 WP48?

[floppymoose]

As the best case scenario - so no injuries, minutes allocations as assumed and so on.
Now we know that Warriors had several injuries to key players - first Biedrins wasn't fully recovery, Lee misseed almost 10 games and Curry is all the time bothered by injuries (and also missed games).

It also worth to note that this prediction (56.9) is a aberration. There are other predictions based on WP and most of them agree that Warriors would be below 50 wins, for example: http://thenbehteam.blogspot.com/2010/10 ... tions.html
49 wins isn't far off if we keep in mind injuries.
http://nerdnumbers.com/archives/859
50 wins

Check the comments on the article. I'm giving WP the victory if the Warriors hit 45 wins. That's below either of the numbers you quote above. So saying that I'm cherry picking here won't really fly.

It would be also nice if some of you would provide other predictions, for comparison sake, because it's always easy to criticize but how many predictions are better than these based on WP?

It would also have been nice if dberri had told us in his comments whether or not he agreed with the prediction. But he didn't. He just told me in so many words that I was wrong, without ever making a coherent argument.

As for my predictions, I was on record as taking 44 or less and giving WP 45 or more. I could pick a specific number now but that would be hindsight, because the season is well underway. I did pick the Warriors to finish 13th in the western conference before the season began, though. That can be seen here:
http://pattyanddave.com/dave/warriors/p ... ndings.php
http://pattyanddave.com/dave/warriors/p ... p?owner=28

[lorak]

It would be also nice if some of you would provide other predictions, for comparison sake, because it's always easy to criticize but how many predictions are better than these based on WP?

I wrote already

Where?
Show some predictions wrote BEFORE the season, because only that way it makes sense.

That can be seen here:
http://pattyanddave.com/dave/warriors/p ... ndings.php
http://pattyanddave.com/dave/warriors/p ... p?owner=28

Where is your predicted win-loss record for every team?

[mysticbb]

Where?
Show some predictions wrote BEFORE the season, because only that way it makes sense.

I guess you couldn't quite follow what I wrote. I used as a basis the minutes distribution of this season. That means all the injury excuses are included in that calculation. Thus WP48 isn't getting a worse result due to Lee missing games, he is just included with his current value. Now I want to test how good a metric can predict the success of a team by using data from LAST season for the players.

That makes no sense? Well, check out Rosenbaum&Lewin (2007), "The Pot Calling the Kettle Black". They are doing exactly this and Wins Produced is the worst predictor. The same happens here with the Warriors.

[floppymoose]

Where is your predicted win-loss record for every team?

Don't have one. That's the great thing about theory verification. You don't have to have an alternate theory before you demonstrate problems with a current one.

Nobody has to propose another metric in order to discuss problems with WP. It doesn't handle defense correctly. It doesn't handle fake rebounders correctly. This can be seen when guys like Lee and Murphy switch teams.

It's an obvious limitation of boxscore-based metrics.

[Paydro70]

It doesn't handle ANY rebounders correctly. Nor does it handle shooting efficiency correctly. It's honestly just an absurd setup, I don't know how he continues to be taken seriously.

[mysticbb]

I don't know how he continues to be taken seriously.

He has a published paper in an economy journal, that is the reason. That is the reason Forbes is using his WP48 to determine over- and underpaid players. How absurd is it to think that Rashard Lewis shouldn't have get any money last season? Well, you can only think this, if you believe the following things: Right now Kevin Love has 0.451 WP48, an average player has 0.099 WP48 right now in the league (slightly below 0.100 due to overtime minutes). The conclusion here: Love is 4.57 times more valuable than an average player! Who in hell believes that? WS48 has the highest rated player with 2.85 times better than average, PER with 1.77 and my rating with 2.13. Just to understand that point: If Love would play with average players around him, he would be enough to put that team at a level of 0.863 win%, if he would play all team minutes. Ok, playing him so many minutes is absurd, thus we just go by the current share of minutes he gets. That gives us a 0.775 win% in case Love's backup is also an average player. That means Love is as good as bringing an average team to 63.5 wins in a season! That is 22.5 wins above an average replacement player. That value has to show up somewhere, right? But why in hell are the Timberswolves at -6.45 with him and -6.5 without him per 100 possessions? Adjusting his PM value gives us +2 by Barzilai or -0.8 by another method.

Well, where is the problem? Linear thinking. Berri thinks linear. He thinks he can look at every event isolated and gets in sum the correct answer. Well, that works on the team level, but the distribution to the players fails badly. Why? Because an offensive rebound for example is not the result of an isolated event. To understand that point better: Berri gives -0.032 for a turnover, but he is also giving -0.032 for a missed field goal. In which basketball game is the probability to score after a turnover the same as after a missed field goal? In fact, in no basketball game. A missed field goal has a chance of going back to the offensive team by around 30% (26% of the times via individual offensive rebound, 4% of the time via team offensive rebound). After an offensive rebound the offensive efficiency is higher by about 1.3 times. That means in a possession with a missed field goal a team will still put around 0.42 points per possession in average on the scoreboard, after a turnover it will always be 0!

An offensive rebound doesn't happen without context, it doesn't happen without that missed field goal. And if no shooter is taking a shot, it will be a turnover (a team turnover, which is NOT showing up in the indvidual boxscore!). And that is what studies showed, a player with a high usage reduces the overall amount of turnovers for his team. Some are doing that more, some are doing it less, but in the NBA high usage player are usually doing this.
In Berri's linear model an offensive rebound is an isolated event without context. Well, that is part of the thinking process, that's why Berri thinks a linear regression on team data will give him anything to evaluate the impact of individual players. Yes, for some players it seems to be working, but that is not a proof that his system is working, it just shows that talented players can put some numbers into the boxscore, but so can less talented and lower impact players (especially on bad teams).

Another part of the problem is defense. The boxscore numbers can explain around 50% of the variance in defense, the rest is not showing up in the boxscore. 50% of the defensive variance is not correlated to anything in the boxscore. And what is showing up in the boxscore isn't even highly correlated to the real impact by an individual player. For example, the data from the regression for rebounds has an around 0.35 correlation coefficient to the rebound percentages for individual players. That basically means that your conclusion via TRB% that a player is or is not helping a team with his rebounding is in most cases just wrong. But to go in with defense here: Berri is distributing the team overall defense among players accoring to their playing time. That means, he just assumes that for those 50% of the defense every player has the same value per minute on a team. Well, via that adjustments he is pushing his correlation coefficient from around 0.7 to 0.9ish, and he is justifying that with the comment that this small adjustment doesn't change the ranking of the players much. No kidding, Berri, for sure it doesn't change much for the player ranking, because you just assumed that every player is basically the same defender in this league. The high impact defender on a great defensive team will see his real defensive value get closer to average, while even the worst defender on the worst defensive team will also get closer to average. That doesn't change much of the ranking among players, because it is not accounting for individual defense impact. The point is: You can do that with EVERY boxscore based metric and it will show a higher correlation to winning on a team basis, but that doesn't mean that it becomes a more useful tool to evaluate players. In fact doing that with PER will give us a better predictor than WP48 for the future team success (as Rosenbaum&Lewin showed).
What's needed to be tested is how Berri's model without the team adjustment is doing, because in that case you can say something about the value of the players. And it is doing AWFUL.

So, and in that way Berri ends up with Kris Humphries as franchise player material. In that way one of his lemmings concludes he would take Humphries over Nowitzki.

@floppymoose

Regarding the +/- comment you made: I wasn't specific enough, I meant every boxscore based metric will underrate a guy like Collins.

[floppymoose]

And here is the Wages of Wins damage control thread:

http://dberri.wordpress.com/2011/02/08/ ... d-instead/

[Sleepy51]

I piggybacked your discussion to the Warrior board, where I'm sure it will get dumber and more hostile, but this has been a great conversation so far, and might be eye opening for some of our fans.

viewtopic.php?f=15&t=1090805&p=26735081#p26735081


My take?
WP is inherently vulnerable in that it is based on a formula of boxscore stats and therefor CAN NOT account for defense in any meaningful way. Because the traditional boxscore stats do not account for defense in any meaningful way. We don't count the right things to measure defensive performance. We don't count how often you need help or switches. We don't count how often you change shots without blocking them. We don't count box-outs. We don't count blown rotations. We don't even count things as simple as shooting fouls conceded vs. non-shooting fouls or and+1's conceded.

WP is based upon regression analysis of boxscore statistics combined with an adjustment factor for team defensive performance. While regression analysis a powerful analytical tool, the quality of conclusions you can draw from regression analysis is HIGHLY dependent upon the quality of your data-set (and the quality of your UNDERSTANDING of the data being described.) If you do regression analysis on the wrong data points, your results will still correlate highly to wins, but not for the reasons you think. When data that you don't count starts to figure prominently into a particular player's game to game performance profile you get aberrations and major flaws in the analysis, like the prediction that David Lee would contribute 19 wins to this Warriors season.


The gaping hole in the WP analysis comes down to how the formula addresses defense. After accounting for a player's individual offensive production, an adjustment is made based on an incorporation of team defensive boxscore stats. It makes no distinction between team and individual contribution.

http://www.wagesofwins.com/CalculatingWinsProduced.html

It would probably produce a significant improvement in the WP formula to incorporate on/off differential in that defensive adjustment and would probably reduce the margin of error for the kind of player that has torpedoed the Warriors prediction, but Berri & cult are extremely hostile to 82games.com data in any way shape or form. That blindspot is probably the biggest Achilles heel in their understanding of the game. Some things you HAVE to account for by differential because of the nature of team concepts (like rebounding and defense.)

Folks who don't realize the grave inadequacy of the traditional boxscore in describing individual defensive contributions to the team result on the defensive end are missing the boat, and building a model based on bad/highly incomplete data set. The ability to disregard or overlook the inadequacy of that data set implies a certain lack of understanding of the game.

[Paydro70]

That WP fans are overly confident in its deeply flawed vision of defense is perhaps its biggest flaw from an overall perspective. But the factor that I find most baffling, and with least connection to real basketball, is providing zero value for shot creation. The notion that 5 Tyson Chandlers (or his small-man equivalent) would win 80 games just reflects a complete lack of understanding of how basketball works. The sheer arrogance of believing that every NBA personnel decision-maker has no idea what's going on at all, ever, is ridiculous. Yes, they make mistakes, and get blinded by certain things, but for the most part, players' minutes and shooting opportunities make sense.