Who had best the one-two punch in the NBA last season? Previously I looked at the best player on each team (see Kobe Myths) and also the best trio (see the Pareto Principle). What about the best duo? For an answer, let’s turn to Table One:
Table One: The Best Duos in 2006-07
If we look at Wins Produced, the top duo in 2006-07 was Shawn Marion and Steve Nash. These two players combined to produce 38.9 wins for the Suns last season. But if we focus on WP48 [Wins Produced per 48 minutes] – which is how the duos are ranked in Table One – the leaders were Tim Duncan and Manu Ginobili of the Spurs. The WP48 of these two stood at 0.344 in 2006-07.
The Top Duo in the Big Easy
Table One not only provides an interesting angle on last season, it also allows us to better appreciate the wonder that is the New Orleans Hornets this season. The Hornets didn’t make the playoffs 2006-07. This season, with many of the same players, this team is suddenly among the top four squads in the West. For many, this is a surprise. When we turn to Table Two, though, the play from New Orleans is actually something we should have expected.
Table Two: The New Orleans Hornets in 2007-08 after 34 games
Table Two provides two projections for this team. The first presumes that every player will play as well as he did last season. The second projects what will happen if each player continues to play as well as he has in 2007-08. As one can see, each projection yields a very similar result. The projection based on 2006-07 values indicates this team should win 51 games. When we look at this year’s performance, we project 52 victories.
Although the totals are the same, there are some differences. Chris Paul has clearly improved upon what he did last year while Jannero Pargo is quite a bit worse. In fact, Pargo’s play is clearly holding this team back. If Pargo were replaced by an average player, or a player with a 0.100 WP48, the Hornets would be on pace to win 59games. Yes, Pargo stands in the way of this team rising to the top in the West.
Even with Pargo, though, this team is pretty good. And although Adrian Wojnarowski of Yahoo Sports thinks it’s because of the play of Paul and David West, the Wins Produced answer is a bit different. Certainly Paul is a big part of this team’s success. But the other half of the leading duo is the play of Tyson Chandler.
Chandler’s WP48 stands at 0.322, a mark quite similar to what he did last year. And when you combine his play with Paul, we see a duo with a combined WP48 of 0.354. Such a mark surpasses the WP48 posted by Duncan and Ginobili last season. Although I have not looked at every team this year, it’s possible that Paul and Chandler are the best 0ne-two punch in the Association in 2007-08.
Unfortunately, unlike the Spurs last season, the Hornets don’t have much after these two. The combined WP48 of everyone else on the Hornets is 0.036. Last year, only five teams had a non-top two that were less productive.
Of course, one of these five was the Hornets. In 2006-07 every player on the Hornets not named Paul or Chandler combined to offer a 0.021 WP48. So the supporting cast on this team has improved.
And much of this improvement has been due to the play of Peja Stojakovic and Morris Peterson. When we look at Table Two we see that Stojakovic’s WP48 currently stands at 0.076 while Peterson’s WP48 is 0.103. Neither mark is outstanding. But these players have replaced Desmond Mason, whose WP48 was well below zero last season. And when you replace awful with average, your team tends to improve a bit.
The David West Story
Again, Wojnarowski argued that the play of David West was key to this team’s rise towards the top. This argument appears to be motivated by the scoring average of West. Currently only Paul averages more points per game.
When we look past scoring, though, we see the story I told in the following column offered last September.
How Fewer Wins Led to More Pay for David West
This post indicated that West – despite his scoring average — was the quintessential average power forward. As Table Three indicates, last season he was near average with respect to shooting efficiency, rebounds, steals, blocked shots, and assists.
Table Three: The Career of David West Prior to 2007-08
This season West has improved with respect to rebounds. But his turnovers are a bit higher while his steals a bit lower. As a result, his WP48 of 0.125 is not far from average.
Still given this mark, the minutes West plays, and a lack of many other outstanding players on the roster, he’s on pace to finish third on this team in Wins Produced.
This placing, though, is a distant third. Again, West is average at just about everything. In contrast, Paul and Chandler are absolutely amazing.
And it’s this dynamic duo that is truly leading the Hornets drive to the playoffs. How far this drive lasts, though, might depend upon the help Paul and Chandler get from the collection of average and below average performers they know as teammates.
– DJ
Our research on the NBA was summarized HERE.
The Technical Notes at wagesofwins.com provides substantially more information on the published research behind Wins Produced and Win Score
Wins Produced, Win Score, and PAWSmin are also discussed in the following posts:
Simple Models of Player Performance
What Wins Produced Says and What It Does Not Say
Introducing PAWSmin — and a Defense of Box Score Statistics
Finally, A Guide to Evaluating Models contains useful hints on how to interpret and evaluate statistical models.
Rashad
January 9, 2008
Did you Hollinger’s article on the all-star selections using PER as his main point of analysis? Are you planning a Wins Produced analysis of who should be going to the all-star game?
Ryan Schwan
January 9, 2008
Nice, DJ
I’m looking forward to seeing how David West ends the season in your rankings. About three weeks ago, after he had a rough stretch offensively(36% shooting), the Hornets started giving it to him in the low post, and stopped setting him up on the elbow for long jumpers and two-dribble drives. He’s been near amazing since then(60% shooting for 1.39 PPS), since he has more room to operate, and only takes his 20-footer in transition. As for the bench, they make me want to cry.
Oh and you forgot to mention that Hilton went from serviceable(.05 WP48) to awful. Considering the Hornets expected him to be our primary big man backup, that hurts.
TG Randini
January 9, 2008
Camby and Wallace again! I made a New Year’s resolution to speak of rebounds only one more time and then be silent. Herewith —
THE SO-CALLED ART OF REBOUNDING: ON THE NATURE AND VALUATION OF CERTAIN POSSESSION TERMINATIONS IN BASKETBALL
There are many classifications, or typologies, of painting as art. Among them are classical art, impressionism, cubism, post-modernism, and dadaism (among many others). These various types represent different artist world-views, or, different ways of seeing the world. Below are presented various ways of looking at the valuation of rebounds in basketball.
#1: A Classical Approach to Valuation
Berri, Schmidt & Brook, in The WAGES OF WINS (2007, updated paperback edition) indicate on page 101 that opponents’ possessions are terminated five different ways: turnovers, made field goal attempts, made free throw attempts, team rebounds, and defensive rebounds. They indicate on pp. 248-249 (footnote 32 to Chapter 6)…
“We do not consider the impact of an opponent missing a field goal or free throw. When the opponent misses a shot the team can collect the defensive rebound, which we include in our evaluation.” …
“So opponent’s missed shot add nothing to our story.”
Okay… I beg to differ. It may not add anything to BSB’s regressions, but causing an opponent to miss a shot utterly DOES matter in the proper allocation of credit when terminating opponents’ possessions.
That’s the crux of the matter, isn’t it? How do we terminate the opponent’s possessions and how do we allocate the termination credits?
Under Berri, Schmidt & Brooks (henceforth, ‘BSB’) model… the allocation of credit (for missed shots) goes 100% to the person who fetches the ball after the opponent misses the shot.
Unlike BSB, I think the activity that terminates the possession is composed of two distinct elements:
Activity A: the causing of the shot to be missed
Activity B: the fetching of the ball after the shot is missed
Thus, (Activity A) + (Activity B) = Opponent Possession Termination
I believe that Activity A is just as important (if not more so) than Activity B… but for simplicity… let’s say they are equal.
Then, the allocation of credit for the total activity (activity A + activity B) would be as follows when valuing the possession termination due to a missed field goal (and utilizing defense as a team construct):
Point Guard .1
Shooting Guard .1
Small Forward .1
Power Forward .1
Center .1
Rebounder .5
The person who actual rebounds the ball would end up with .6, or 60% of the credit because .1, or 10% of the credit would be incremental to the 50% due to his team defense.
It would be easy to allocate credit in this manner: simply attribute rebounds as a .5 weight instead of a 1.0 weight to the rebounder… and then allocate the other 50% of individual ‘rebounds’ to a team pool (constituting team defense) and further allocate based on minutes played.
In this manner, you will more properly allocate “the causing of the shot to be missed” PLUS “the fetching of the missed shot”… instead of BSB’s attribution of 100% of the credit to Activity B, the person who fetches the ball.
Pertinent: The causing of the missed shot is an important activity because field goal percentages are almost always much lower than free throw percentages. If ‘causing the missed shot’ was not an important activity, then field goal percentages would be much higher and would reach equilibrium with free throw percentages. In fact, they could even exceed free throw percentages because you are forced to stand at least 15 feet from the basket when attempting a free throw. An un-defensed field goal could be as easy as a slam dunk or layup.
In summary to my classical approach to valuation… I believe a proper model would allocate credit to the skills employed in causing the opponent to miss a field goal attempt (and the fetching of the ball) instead of giving ALL the credit to the person who fetches the ball after the missed field goal.
Note: In answer to those who would say BSB’s ‘position adjustments’ would negate the effect of 100% attribution of (Activity A + Activity B) to merely the ball-fetcher (Activity B)… it wouldn’t… at least, not properly.
BSB’s methodology relatively overrates high-rebounding guards over lower-rebounding guards (and higher-rebounding forwards/centers to lower-rebounding ones). In effect, the BSB methodology overrates Jason Kidd compared to Steve Nash.
I shall not go into BSB’s conversion of a regressed .67 assist value to a .50 here, but will note that it is interesting they cite (p. 254, footnote 12 to Chapter 7, 2007 paperback) Hollinger’s three part deconstruction of a scoring play that is very analogous to my two-part deconstruction of a missed field goal/rebound, and then they complement themselves that their .67 regression value for an assist corroborates Hollinger’s intuition!
BSB, it is time you intuit the proper attributions of credit on the defensive side also.
Please note, the regressed value of the TOTAL activity (Activity A + Activity B) would be the same under my system as in BSB’s regressed value for Activity B alone. My system would not change the amount BSB’s regressed variables are ‘explained’ (using ‘explained’ merely in the academic statistical sense).
But my system would explain (capitalize EXPLAIN) what is going on a heck of a lot better than BSB’s attribution of 100% defense to the ball-fetchers. It would have the same statistically relevant explanatory power, but with the added enhancement of a much better way of attributing credit.
Explaining, instead of ‘explaining’.
TG Randini
January 9, 2008
#2: A Cubist Approach (to Valuation)
Let’s have a little more fun now. Let’s compare the ‘missed FG and rebound’ to baseball. In baseball, the batter is part of the offense, and the pitcher, catcher, and fielders constitute the defense.
The pitcher throws. The batter swings. Strike three. He’s out! Another strike out for Randy Johnson!
What happened here?
The batter (shooter) missed the ball (basket). The catcher caught the missed attempt. But the fans are cheering for Randy Johnson, the pitcher! What are they… nuts? The catcher caught (fetched) the missed attempt by the batter! Why are they not cheering for the catcher… (the fetcher)? (A mini-poem, sorry…)
Because the fans know something that BSB do not know. That defense is not 100% attributable to the ball fetchers and catchers.
The fans know that the speed, location, and spin of the ball by the pitcher (analogous to the movement and placing of basketball players when playing defense) constitute at least 50% of the value of getting the hitter to miss the ball when he swings (or the shooter missing the basket when he arches one up toward the rim).
TG Randini
January 9, 2008
#3: Another Cubist Approach to Valuation
I once watched the Dipper play tennis at a public park in LA in the late eighties. Great guy. Here’s the thing: Wilt was a graceful player. Two steps: forehand down the line. Three steps the other way: backhand down the line. Accomplished? Yes. And surprisingly graceful. (By the way, he was a track star in high school… and dberri, he did not like to be called ‘The Stilt’. He did like ‘The Dipper’, however…)
BSB (page 149, WOW, 2007 paperback edition) indicate that Wilt may have been responsible for 52-odd wins in the early 60’s when his team (the Warriors) won only 49 games. They, of course, see something wrong with this calculation.
If this calculation makes sense, the Warriors would have been better off if the other four guys did nothing at all! Just stand in a corner of the court and do nothing!
I love Wilt but he couldn’t do it all. Ergo: the Celtics dynasty.
The calculation is non-sensical because BSB attribute 100% of the credit to the rebounder for ‘making the opponent miss a shot’ + ‘grabbing the rebound of the missed shot’.
The attribution of 100% of the credit to Wilt for getting a rebound is what makes the 52-odd wins non-sensical… and by logical extension… … makes their attribution of credit in their Player Wins model non-sensical.
Dberri, there was a REASON for those four guys to be on the floor with Wilt.
(Note: Once again, I am saying BSB’s attribution of CREDIT is non-sensical, not the regression value for their mis-defined concept of ‘rebounding’ itself. The BSB regression value for (and terminology of) ‘rebounding’ is equivalent to ‘making them miss the shot + rebounding the loose ball’ in the construction I have presented above.)
TG Randini
January 9, 2008
#4: A Dadaist (and Final) Approach to Valuation
Let’s imagine a league where FG % is even worse than it is in the present NBA. (Heck, we could go back in history to the NBA itself… the 50’s… especially the early 50’s…)
Let’s imagine a 30% conversion rate. On average, the ‘expected outcome’ (sorry, people, that’s econ-speak)… would be negative for shooters.
A rational decision-maker (sorry again… more econ-speak), in this case, the coach, would instruct his players NOT TO SHOOT. “Don’t shoot!” the coach says to his team in the huddle. “Don’t EVER shoot!” Then, the expected negative outcomes through shooting would never occur, and his team would have lots of positive player win value through rebounding.
Except the other team’s coach is ALSO a rational decision maker and he instructs his team to do the same thing.
So… neither team shoots. Neither teams scores. The fans stop coming. The game of basketball dies.
All because BSB overvalue rebounding.
IN SUM: The above represents my thinking on the nature and valuation of certain possession terminations in the game of basketball as played in the late 20th and early 21st centuries using various classical, modernist and dadaist constructions (or dadaist ‘destructions’). Please consider the term ball ‘fetching’ as a dadaist intrusion into the classical and cubist schemes.
As Dorothy once sang in The Wizard of Oz…
Fetchers and catchers and bears, oh my!
The value of a rebound is much less than ‘pi’!
Always,
Tomas Giuseppe Randini
Mike H
January 9, 2008
Can someone who still reads anything Randini posts please summarize his points for me – assuming that they have any merit? I stopped reading him long ago when he was just trolling, but I’ve heard that he actually says something interesting every once in a while now.
Animal
January 9, 2008
Mike H, the conclusion of his 3rd post kind of recapitulates everything:
As Dorothy once sang in The Wizard of Oz…
Fetchers and catchers and bears, oh my!
The value of a rebound is much less than ‘pi’!
Jason
January 9, 2008
Randini does not seem to be aware of the concept of signal to noise. I might have read a single post. As it was, I just got annoyed that I had to scroll down as much as I did.
Patrick Minton
January 9, 2008
Randini,
I stopped reading when you said:
“I believe that Activity A (forcing a missed shot) is just as important (if not more so) than Activity B… but for simplicity… let’s say they are equal.”
This assumption is pretty flawed, but even if I agreed with the assumption, it falls under a specific category of theories that are doomed to failure: ones that are not disprivable.
The scientific method dictates pretty clearly that if a theory cannot be disproven, it is useless as a basis for experimental analysis.
By the way, the major flaw in your assumption is that if this were true, rebounders would perform better on teams that force their opponents to shoot a lower FG% (there would be more rebounds to collect, according to you), or in games where FG% is collectively low. I’m not sure any evidence for such correlation exists. For instance, take Kevin Garnett, who led the league in rebounding several years in Minnesota. He now plays for Boston, who’s opponent FG% is significantly lower than Minnesota’s last year, and yet his rebounding has declined.
Rasta
January 9, 2008
Dave,
There appears to be a misprint in the David West career page. Under “NBA Avg PF” column, the WP48 is listed as 10.3, and the Wins Produced is .215.
I’m guessing these numbers reversed.
If so, that leads to a question: does the “average PF” really produced WP48 of .215? WP48 is the result AFTER the position adjustment, right?
TG Randini
January 9, 2008
Patrick,
Your “For instance, take Garnett…”:
The logic of your retort is wrong. Garnett’s rebounds by themselves prove nothing. You would have to look at the total rebounds of the entire team (Celtics) vs. the other entire team (Minnesota). One individual’s rebounds differential means nothing.
TG Randini
January 9, 2008
Jason, the people who still like “Dick and Jane” get annoyed at, and will never comprehend, “War and Peace”.
And that’s okay.
Mountain
January 9, 2008
I agree with TG’s fundamental points and found his presentation a good read. I am not well grounded in artistic world views so I’ll have to catch up a bit on that to capture that part of the presentation better but I think do think that the application of different world views is a key part of the rebounding / formula / basketball analysis debate going on. One view isn’t necessarily right to the exclusion of all others. They are different ways of seeing, parts of the whole or the whole.
Mountain
January 9, 2008
The relative weighting of forced (or not so much) shots vs defensive rebounding is a good topic for further discussion to get closer (or at least more confidently opinionated as one wants) than 50/50 somewhat arbitrary starting split or near 0/near 100 or near 100/near 0 splits as some prefer.
Jason
January 9, 2008
The people who present diarrheic prose and believe it to be “War and Peace” should seek immediate psychiatric attention.
andrew
January 9, 2008
But Jason, don’t you realize the more big words you employ, the better your argument?
Randini is an *academic* name-dropper.
Harold Almonte
January 9, 2008
TG. Dean Oliver apportion 50% to the stop and 50% to the rebound. I was convinced that it was too much for the stop, because a lot of shot defenses are not as closed to be the effect balanced to the defender, but your free thrower analogy made me think that the Oliver’s approach was correct, and you can let some shooters to shoot alone from certain distance, and they won’t be 100%FG%. Keeping the distance is part of the defense even if you didn’t contest the shot.
Animal
January 9, 2008
I think “Mountain” is just “TG Randini” writing in disguise!
This reminds me of the time Rosenbaum made a faux account in order to penetrate the comments thread here and praise his own work!
Mark
January 9, 2008
I like TG Randini’s posts here as well.
Jason and DBerri, please don’t just dismiss detailed posts like this.
Harold Almonte
January 9, 2008
Keeping the distance, in effect, is what zone defense is based in.
Jason
January 9, 2008
I have published papers and written book chapters. I appreciate the value of a good editor, even if it is one’s self. Randini appears never to have learned such a lesson.
This is not simply blaming the messenger; it is possible for the message to get lost in the noise, as happens with three posts, all rather long, two of which appear to be more satire than anything else. They do not encourage discussion. The volume and tone appears to me to be more of a troll or one who is posting solely for his own entertainment. I think he’s overly impressed with how clever he perceives himself to be, though I’ve yet to see him come up with anything particularly novel, anything that someone else hasn’t pointed out and responded to before. I don’t think he actually understands the model he is critical of, or perhaps the purpose of models in general. Whatever points “Randini” may have are lost on me as a result. He is free to prove me wrong, but unless he changes his style, he won’t accomplish this.
Mark
January 9, 2008
Ok then, please ignore him and respond to me.
Is giving the person who rebounds the ball essentially giving them credit for all o fthe defense that is played on that possession?
I agree that the correct weighting for terminating the opponent’s possession is and should be 1. But how much of that should be apportioned to the reboudner and how much should be apportioned to the team for defense played?
“How do we terminate the opponent’s possessions and how do we allocate the termination credits?
Under Berri, Schmidt & Brooks (henceforth, ‘BSB’) model… the allocation of credit (for missed shots) goes 100% to the person who fetches the ball after the opponent misses the shot.
Unlike BSB, I think the activity that terminates the possession is composed of two distinct elements:
Activity A: the causing of the shot to be missed
Activity B: the fetching of the ball after the shot is missed
Thus, (Activity A) + (Activity B) = Opponent Possession Termination
I believe that Activity A is just as important (if not more so) than Activity B… but for simplicity… let’s say they are equal.
Then, the allocation of credit for the total activity (activity A + activity B) would be as follows when valuing the possession termination due to a missed field goal (and utilizing defense as a team construct):
Point Guard .1
Shooting Guard .1
Small Forward .1
Power Forward .1
Center .1
Rebounder .5
The person who actual rebounds the ball would end up with .6, or 60% of the credit because .1, or 10% of the credit would be incremental to the 50% due to his team defense.”
Is this reasoning flawed? Is the weight of the defense played 0 and the weight of the rebound grabbed 1?
Please help me to understand.
Harold Almonte
January 9, 2008
oppFGMissed could have been included in the team defense adjust, with an appropiate weight of course if just it could be done the regression to the possession gaining action: FGMissed defended + DReb, as a whole. But the problem began with getting accomodated to what the boxscore can give, and obviate logic before the regression in this situation.
Mountain
January 9, 2008
Nope Animal I am a separate person from TG.
Mike H
January 9, 2008
So, the issue that Mr. Locuacious is getting at is that he thinks that the rebounder currently gets too much credit for defensive rebounds? I imagine that the proper amount of credit to give the team vs. the player could be empirically derived and could be any value (this may have already been done by dberri). It’s certainly an interesting point. It’s also a point that can be easily conveyed in just a few sentences. So until Randini pumps up his information density and dials down his smugness, many – if not most – will continue to ignore his posts.
Jason
January 9, 2008
There are a few ways to look at the apportionment of the credit for a defensive stop. One is to try to reason out a value by suggesting that X% goes to the guy who got the rebound and Y% goes to everyone else and this Y does not need to be divided equall. Reasoning this through via some “logical” interpretation of the game and what we know happens runs the very significant risk of simply perpetuating a subjective opinion already held; true or not, it’s not evidence based.
Another is to empirically test whether or not dividing credit is necessary and, if so, who gets what. I suspect that other players do have influence on players rebounding, but empirically, as I’ve measured it, it is not great. Either defensive ability does not vary significantly and the outliers who are great or lousy are too few to make a difference in generalization of formulae or others really do not have that much influence (which is a subset of the abilities being not terribly variable, but also not terribly important) or the methodology corrects for the difference in some other way already.
On empirical grounds, we can address things in a number of ways. One is to look at how rebounds vary with respect to other aspects of the game. At the team level, rebounding percentage does covary with defensive FG%. But it isn’t that tight a correlation. Some large part of rebounding belongs to the guy who got the rebound.
But how much? The 50-50 split has some support. The r_sq between defensive FG% and defensive rebounding % is about .45 over the single year I looked. I don’t know if this year was typical or not, but it suggests that about half of the miss is explained by looking at the propensity to grab available defensive rebounds. This isn’t suggesting cause. This simply suggests that, on average, they covary, though not perfectly. Raising one raises the other, but there’s noise. This 50-50 split also presumes that the defensive effort that went into the miss was independent of the rebounder, that anyone could have done this and the rebounder would have had the same chance at the ball had a guard played his ass off on defense vs. playing well himself. More on that below, but the assumption of equal effort divided on defense seems at least as shaky than assigning credit to the rebounder.
The correlation for any particular player between how *he* rebounds from year to year is much tighter and isn’t explained particularly well by defensive FG%, indicating that this 50-50 split isn’t necessarily a good way to go, as it underpredicts rebounding change with player movement. If teammates are largely responsible for the defensive misses, we would expect changing teams to alter rebound rates for players significantly. It doesn’t. Something about the player seems to be largely responsible for his per-minute rebound rate.
It’s possible that big men who get more rebounds are also more active in producing misses. My entirely subjective reasoning on this is that guards guarding guys away from the basket are guarding shots that had a better chance of missing even if there was no pressure. Guys defending inside are defending potentially higher percentage shots that, uncontested, would go in. Big men who get more rebounds may actually be generating more misses that they then gather. I do not presently have direct evidence of this, but if true, it’s consistent with the observation that rebound rates follow players independent of teammates.
I have more thoughts on this, many more, but that’s enough to chew on at this point and in the interest of avoiding losing signal to more noise…
Harold Almonte
January 9, 2008
It’s amazing how this site (and others) just turns around the same this again, and again, and again, and in the end is Berri who has the last word if he keeps the metric as it is, or upgrade it.
The DRebs stealing credits to shot defense in the regression is old stuff from two years ago; The lack of punishing (rebounding rating) by opponent’s rebounds is old stuff from two years ago; The individual punishing of scorers’s FGMade as an extreme logic that can be compared with the justification that because other metrics rewards by taking more shots, and the fact that neither of them fit with the Lg. ave., producing and justificating an “in between” theory, is also old stuff.
I think we just must stop and wait for the WOW II.
dberri
January 9, 2008
As I noted a few days ago, I am posting less. I am also commenting less. But my sports econ class just got out, so before I go home I would note…
1. I agree with Jason. Rebounds are very consistent over time which suggests that a player’s rebounds are about the player, not his teammates. This may result in players being ranked differently than you would like, but such is life.
2. As noted before, changing the value on rebounds to 0.7 does not substantially change the rankings. So even if you make an effort to divide credit, it doesn’t seem to matter.
3. The key differences between the WoW metrics and stuff like PERs isn’t rebounding, but how you treat shooting efficiency.
4. Finally, I am not inclined to change any of this in WoW II. Our second book is not going to be about new basketball metrics. And it is not going to be strictly about basketball. I might give more details as we go through the process of writing this book.
As I can, I am reading the comments. But my ability to participate is limited. There will be no post tonight, but look for something new posted on Thursday night.
Once again, thanks for all the comments.
TG Randini
January 9, 2008
Martin and Mark,
Thanks for your support. I have nothing more to say on the matter of rebounding. Changing the rebounding weight to .5 does matter, even if dberri doesn’t think so. You may also find the discussion at APBRmetrics to be very useful. If you haven’t been there there is a thread that discusses WoW…
http://sonicscentral.com/apbrmetrics/viewtopic.php?t=1232&postdays=0&postorder=asc&start=195
And there are threads that discuss just about everything related to basketball.
The discussions are generally on a high level and people actually use big words there.
Enjoy.
Mountain
January 9, 2008
Jason I was interested in more discussion and appreciate your post.
There are several topics in play here and elsewhere: division of rating system credit between shot defense and rebounding, credit split for shot defense among defenders, rebounding credit among the actual rebounder and others, and rebounding credits as a measure of underlying individual rebounding skill (not needed or expressed the same in different contexts).
Taking these one at time fully might be better than all at once but ultimately they all need to be addressed as well as possible and fit together.
My post in support of TG’s main points came out of and was mainly acknowledging my support (at apbrmetrics, recently and on-going from the past) players getting a visible individual share of defensive credit in a rating system from the missed shot / rebound sequence
for shot defense.
In your presentation you start by saying :
“There are a few ways to look at the apportionment of the credit for a defensive stop. One is to try to reason out a value by suggesting that X% goes to the guy who got the rebound and Y% goes to everyone else and this Y does not need to be divided equall. Reasoning this through via some “logical” interpretation of the game and what we know happens runs the very significant risk of simply perpetuating a subjective opinion already held; true or not, it’s not evidence based.”
True but later in your presentation you say
” the assumption of equal effort divided on defense seems at least as shaky than assigning credit to the rebounder.”
Now I am correct you support (at least to date) all credit to the rebounder? And if both positions are shaky, after the empirical study, why is that position favored by you? Does it boil down to your “logical interpretation of the game” or is it arbitrary?
Maybe the alternative is not better but isn’t there room for alternative views? Ultimately I think it may be that folks do have to choose for themselves and that gets back to TG’s theme of worldviews. It was elaborate (with extras) but I think he made that point.
Even if we don’t reach consensus I think those with different views can understand the topic better and their own view and the views of others better from the interaction. It is tough because it is a complicated topic and the worldviews used in discussing it are different. I might have more later but those for my thoughts for now.
TG Randini
January 9, 2008
Mountain… thanks for your support, I meant to thank you earlier when I called you ‘Martin’.
Sorry. (Good luck on this site).
Animal
January 9, 2008
Mountain, I am sorry for falsely accusing you of being a fake account of someone else. Welcome to the site!
Jason
January 9, 2008
I think the null is that the whole of the credit goes to the individual who grabbed the rebound. It may be arbitrary, but anything other than whole credit or zero credit is substantially more arbitrary. We can reject the notion of zero credit on empirical grounds because a) rebounds clearly *do* correlate and raise win probability and b) rebound rates *do* follow players. Some credit need be assigned else you are clearly missing something.
My position is that the default is to assign credit to the guy who got the rebound. There is empirical support for this being a reasonable way to deal with it because rebounds do follow players so well, more than most other aspects of the game. My point about dividing it up in some manner (e.g. rebounder gets .6, everyone else splits the remaining .4) is arbitrary in the assumption that everyone else who did not rebound was equally responsible. If we divide it, why .6/4x.1? Why not .6/.2/.1/.1/and zero for the guy who was staring and Jessica Alba on the play? If the rebounder isn’t responsible for almost everything (which may or may not be true) it is almost certainly not true that all players who didn’t get the rebound contributed equally.
There is plenty of room for other ways of dealing with it, but the method of trying to ‘reason’ why it needs to be divided is not going to convince me. I have not seen a good argument here nor a APBRmetrics and I have seen a whole lot of ignoring how consistent rebounding for particular players are, independent of many of the variables proposed to influence them.
I am also not convinced that this ‘partial credit’ is not already accounted for elsewhere in the model sufficiently to make the model accurate enough. Most of the critiques seem to assert that it overvalues rebounding largely because the critic sees rebound specialists rated “too high.” I will not be convinced because someone believes that the player rankings are off unless those rankings are directly correlated with predictive ability of the model, and that predictive ability that matters is the ability to predict future victories based on past player performance. Rankings based on other aspects are arbitrary and it seems that at the heart of most criticisms is some subjective sense that the rankings “can’t be right” because rebounds seem to influence things so much.
If it can be demonstrated that modification results in superior prediction, I would be convinced that the modification was superior. I have NEVER tried to silence anyone from trying to finding a better method. *If* it is demonstrated that credit undivided results in inferior predictions for some reason, then we can reject the null (the null again being that the rebound belongs to the rebounder).
But I do not think that this has been done. I also do not believe it can be done in any manner other than examining the issue *empirically*. Analogies to baseball and the catcher (a rather poor analogy that draws a strawman) is not going to do it. Insistence that the model is flawed based on ridiculous extremes that do not ever occur will not convince either.
Your own “worldview” is fine, but if in your worldview, you believe you are ranking players based on their ability to add to victories for their teams, all opinions are not equal.
The criticism I have seen has, for the most part, not been objective criticism, but some pseudo-logic explaining why it has to be different; it does not actually demonstrate that this is so. I do not have *ANY* problem with different views. I have problems with the assertion that something is wrong with an empirical model without reasonable empirical support. I have problems with declaring a problem because it doesn’t fit with a deep held notion about what is important, supported largely with the observation that this deep held notion is also widely held. This latter point bothers me the most as it is entirely anti-science.
I do not strive for consensus. I do strive for reason. I seem a whole lot of anti-reason masquerading as criticism.
SHC
January 9, 2008
‘Chris Paul has clearly improved upon what he did last year while Jannero Pargo is quite a bit worse. In fact, Pargo’s play is clearly holding this team back.’.
Jannero Pargo is holding this team back? I like ‘wins produced’, but there are too many articles in this blog that have gone overboard with this ‘wins produced’ == ‘player’s value’ assumption. I’ll give credit to wins-produced for what it is – linear regression learning of co-efficient for each category. But it’s too simplistic to interpret each player’s value – (Bruce Bowen, Stephen Jackson are completely useless? Come on.)
Nobody knows why Jannero’s WP48 are that way – it may be chemistry issues, or reduced minutes, or maybe all his production got dispersed among the team members because he sacrificed himself. It’s for coaches to decide.
It’ll be interesting if anyone decides to take upon more complex research (maybe applying simpler graph models such as Naive Bayes) to find correlation between ‘team’s production and a player’s presence.
antonio
January 9, 2008
Jason, what is your opinion on what Rosenbaum wrote here. The paper this comes from was not written by just him, but other great basketball minds including Dean Oliver. Rosenbaum also claims it to be peer reviewed. Him and Nick S continue to have a great discussion, but I would be interested in the WP defense of the criticisms they bring up.
http://www.sonicscentral.com/apbrmetrics/viewtopic.php?t=1595&start=0
SHC
January 9, 2008
antonio,
This is really interesting. So my take is that dberri solved a regularized linear regression, but assumptions made in the constraints were not adequately explained.
However, it seems that wins-produced model does a very good job on test set, so I’m not sure if we’ll get a wildly different result even with better constraint.
Mountain
January 10, 2008
Maybe the presence of multiple issues is affecting the communication.
“I think the null is that the whole of the credit goes to the individual who grabbed the rebound. ”
I can understand / accept that stance- if we are talking about just the rebounding part of the defensive play, giving the whole of that credit to the actual rebounder.
Giving a small share of credit to other teammates for the rebound capture is not a big deal and is not clearcut how much to give and can be dealt with thru other means.
Giving the whole of the credit for the entire shot defense /rebounding sequence to the individual who grabbed the rebound is not necessarily the null. I do not consider it the null.
I hear your points regarding the criticisms of Wins Produced treatment of rebounds. There are stronger and weaker criticisms made of it. I don’t think the case for rejecting WP treatment of rebounds is case closed / nothing more to say as some treat it, if you step back and look at the whole method.
I don’t like everything about the PER accounting system for rating players. I am not making a simple WP vs PER rebounding argument. My primary focus at the moment was on trying to get individual shot defense (only roughly estimated right now) fairly and visibly included in a player rating formula and was trying to think thru how that fits with an rebounding credit system and if it offered an opportunity to address issues with rebounds.
Simple PER ignores shot defense but counterpart PER or combined into Roland Rating corrects for that. Shot defense comes in via the team adjustment for Wins Produced. I accept and use these approaches (on the WP side I find PAWS/min a fine metric). But maybe there could be another way that might suit at least me better. So I was tinkering with the concept of something new. TG and others were too. So I piped up briefly in alliance with that. But now I am going to bed.
Jason
January 10, 2008
Dave has said that he used regression to derive his coefficients. Rosenbaum asserts otherwise. I’m inclined to believe the person who made the model used the method he claimed rather than a counter-assertion.
Mountain
January 10, 2008
Well, in a minute.
I do agree that it would be incumbent on a new system that broke defensive credit out into shot defense and rebounding shares to show it predicted wins better- if that was its purpose. It doesn’t have to have that purpose. That wasn’t PER original purpose. But I would certainly be interested in if it could pass that test. I am interested in both player comparisons and team win predictions. Doing both from the same method would make sense.
If such a method gets built, I’d hope it got tested head to head with the popular metrics and variants & combinations of metrics and would want to see what the test shows. If it proved stronger, great. If not, I would try to understand why and assess where to go from there.
SHC
January 10, 2008
No, Rosenbaum is not saying that dberri didn’t use regression. What he’s saying is that there’s a constraint on the model (which is very reasonable, to avoid overfitting) and instead of using a reasonable constraint, dberri made some arbitrary choices.
You’ll have to have some knowledge of statistical learning to follow their arguments.
TG Randini
January 10, 2008
“If we divide it, why .6/4x.1?”
Why? For two reasons. Because a) basketball is a team sport, and b) basketball is a continuous activity or sport (rather than a more ‘discrete’ -type activity or sport like baseball).
The individual defensive credits for the ‘shot defense/rebounding sequence’ that terminates the opponent possession would be more fairly allocated than merely giving 1.0 or 100% credit for team defense to one man.
Basketball is a team sport and it is teams, not one individual, that play defense.
It is not “pseudo-logic” or “anti-science”… but photons (of light) striking my retina (isn’t that science?) that make my eyes see five men on a court playing defense/collecting rebounds… not just one man.
“Insistence that the model is flawed based on ridiculous extremes that do not ever occur will not convince either.”
They do occur. Berri even put one in his book.(Chamberlain’s Philadelphia Warriors. )
Kevin Broom
January 10, 2008
I think that this conversation is running up against the limitations of the statistics actually collected and published by the NBA. In a sense, both sides of this conversation are correct.
Assigning full credit for a defensive stop to the guy who gets the defensive rebound is usually going to give too much “credit” to the rebounder. I have hand-tracked defensive data from Wizards games in some detail over the past few seasons, and the busiest defenders (bigs) are typically involved in only about 25% of defensive possessions when they’re in the game. That includes forcing misses, forcing turnovers, fouling, and grabbing defensive rebounds. (That looks at the outcome of each possession and doesn’t get into other valuable defensive contributions such as ball denial or shot prevention.)
However, the method being suggested to improve this over-crediting isn’t a good one, in my opinion, because it ends up crediting guys who basically did nothing on a possession. As an example, in 04-05 Antawn Jamison was the Wizards’ least busy defender, and one of their worst. The Wizards effectively hid him on the defensive end by putting Jeffries on the opposing team’s best forward scorer. Jamison almost always had the luxury of being matched against a non-scorer. So, Jamison would end up getting credit for defending all the misses Jeffries forces. And Haywood, who was in fact the team’s best defender that season (and every season since, but that’s a different issue).
The problem, of course, is that the individual defensive data just isn’t being tracked and published. Yet. When/if it is, I suspect Dr. Berri and his co-authors will have some material to examine a revision of WP.
As it stands, the numbers tell us that opponent shooting percentage (efg) is by far the most important determinant of a team’s overall defense. Defensive rebounding is important, but the correlation between defensive efg and defensive efficiency is quite strong year after year after year. Which further suggests to me that assigning nearly full credit for a defensive stop to the rebounder is giving too much credit to the rebounder. (I say “nearly full credit” because the defensive adjustment in WP spreads some of that credit to the rest of players based on minutes played. Until there’s more data available, I do think dividing credit based on minutes played is as reasonable a way to do it as any.)
Westy
January 10, 2008
So the discussion becomes is a null of 1.0 or 0.6+0.1*4 more reasonable. Now Jason has previously shown work that 1.0 is slightly more predictive than the latter in tracking player movement from year to year. On APBR, Mountain suggested using 0.76+0.06*4. I would suspect that’s pretty close to the optimal value.
Mr. Berri notes that this wouldn’t “substantially” change the individual player rankings. I guess I would beg to differ. I think it would serve to flatten the peak value given to players who are only rebounders. And I think this would serve to improve the player valuations.
Jason, as I’ve thought further about the (surprising to me) consistency from year to year of reb./min., I’m wondering if this may be a bit misleading. The point in trying to show this consistency is to say that rebounding is a result of a skill that’s consistent. However, isn’t it possible that rebounding is more correlated with height, body type, position on the floor, and coaching strategy? If these are all consistent, which I think they are, rebounding also remains consistent. This means that a player will tend to rebound the same, and garner the reward (in the WP system) of reflecting his team’s defensive prowess because that’s the role he continues to fill. The fact that it’s so much more highly correlated from year to year than FT%, a ‘skill’ that should be extremely consistent, seems to show that there are further factors pushing players to the place in which they garner rebounds that are static.
The Franchise
January 10, 2008
Westy-
Even if the consistency in rebounding is due more to physical attributes than to skills, it doesn’t matter, as long as the attribute helps the team. Height, strength, and speed are positives.
TG Randini
January 10, 2008
I am certainly not wed to using 0.6+0.1*4 as the optimal distribution for rebounding+defense. The Mountain’s 0.76+0.06*4 may indeed be a better distribution.
It would be great to see what distribution works the best. It certainly cannot be 1.0+0.0.
And no distribution of the defense portion will be perfect because even if teams play defense, some individuals are better at it than their teammates. And until we have better NBA breakdowns as Kevin suggests, no construct will be perfect.
It’s just that any breakdown for defense + rebounding that gives less than 100% credit to the rebounder ALONE will be more ‘perfect’ than one that does.
Jason
January 10, 2008
“Perfect is the enemy of good.”
Perfect is also, in this sense subjective towards the intended task, and I’m getting the feeling that many, many are trying to force the model into doing something it wasn’t designed to do. I’m very certain of this when someone brings up 5 Jordans vs. 5 Rodmans. I’m very certain of this when someone questions the purpose of predicting wins that they aren’t evaluating the model by its intended purpose.
I am not wedded in any way to the notion of assigning full credit to the rebounder, but I am convinced that other allocation will result in better predictions *even if* others received credit. Again, there’s no reason to believe that all players on the floor equally contribute to the missed shot nor is there any necessity that the input is regularly divided or even on average regularly apportioned such that formulaic division will result in better predictions. Showing that that something makes better predictions requires empirical evidence that it results in better predictions, not some insistence that because other things went into the rebound, you cannot assign full credit. Assigning rebound full credit may not correctly show up what happened on the court, but this doesn’t mean that this correlation will not result in accurate predictions. I think I’ve explained why this can be.
WP balances points with FG attempts. A criticism is that rebounds are not similarly balanced with some opportunity. The defensive adjustment does track *made* baskets (really opposition scoring) by the opposition. In this way tracks opposition baskets vs opposition misses (recorded as rebounds). I’m curious to see how pronounced this component of “team defense” is. The “team defense” correction also incorporates opposition turnovers and since turnovers are rather valuable in the method, I’m curious how much of the correction is influenced by this alone. It may be that the scoring vs. rebounds assigned with a position average correction that assumes a baseline of position average production *already* accomplishes a weight of ‘opportunity cost’ such that altering the model will not result in substantially different rankings *or* predictions. Again, empirical questions that beg empirical answers. It is not a given that dividing the rebound must result in more accurate predictions.
I have looked at a variation of the model where rebounds are divided up by floor time, exactly like opposition scoring is. The predictive ability here is far, far, far worse than giving credit to rebounders.
Mountain, if predicting future wins isn’t your criterion for ranking players, what is? Assigning credit to past wins independent of future wins starts becoming a very iffy affair. Whether or not it is correct, it isn’t *testable* and suddenly becomes rather unscientific. Why bother with methodology? Why not just declare to be better than and be done with it?
Mountain
January 10, 2008
0.76+0.06*4 was my revised thought on credit split for rebounding portion of shot defense / rebound sequence. I was still tinkering with the shot defense portion, thinking initially that 50% of that value should go to the individual counterpart to the opponent and rest spread out to other teammates but willing to lower the individual share. Shot defense is perhaps more a collective enterprise than rebounding? I sense some people talking about defensive credit split are thinking of one split. I was thinking of two- for both shot defense and rebounding.
I know there are serious and deserved reservations about the individual shot defense data. Although I think it could still represent an improvement over the simple formulas that lack individual shot defense data, the prospect that such a formula might be ignored by many or most all reduced my incentive to finalize a formula, incorporate the shot defense and present for scrutiny. If I had more time I might anyways but I haven’t yet.
If you defined possession end at release of shot all rebounds would be worth a full possession. But the conventional possession framework doesn’t. This is a worldview choice. Years ago I was for possession end at release of shot. But I put that down to converse in that universe.
WP has full credit for all rebounds. From a scoreboard and winning standpoint they are of course worth the value of possession. Grabbing rebounds is a significant way is which bigs contribute to team. All other things equal a player who rebounds more than another is performing more valuable work for team. But rebound is position and role. PAWs/min is one way to handle that and level the field so to speak- for player comparisons. PER is another.
In a world where every team uses bigs and in roughly the same proportion defensivc rebounds can get taken somewhat for granted. But they are very important to secure.
It might be worth try to estimate team offensive and defensive rebounding impacts of “going small” (though going small is usually just a couple of inches for one or a couple positions and would not represent the extreme of playing 5 guards or 5 PGs- that would fully capture the value of fielding height for rebounding but there isn’t NBA empirical data for that.)
TG Randini
January 10, 2008
Mountain, I think a two level split actually describes the world we see better than a one level split. But Berri has a zero level split… and I was just taking a simplified ‘one small step for mankind’ approach. I think there are two levels: one, the level that basketball defense is played as a team… (where guarding off the ball better makes it harder for the shooter/ball handler on offense has options, for example)… and two, the individual defense on the ball/handler shooter. So, I think you are entirely right. But I didn’t think these guys at WoW would be ready for the entire two-step approach when they argue so defensively against the one-step approach.
Jason
January 10, 2008
Actually, Randini, your grey matter needs an upgrade. You are confusing “arguing against” something with calling your assertion that another method has to be better exactly what it is: an assertion. An empirical model requires empirical evidence. You cannot declare an alternative better without providing empirical evidence to back this.
Mark
January 10, 2008
Jason,
All of the factors that go into producing Wins Produced certainly imply that the player with X Wins Produced actually is reqponsible for giving his team that proportion of wins.
But with rebounds, surely this would only be the case if the rebounds collected by a player were rebounds gained by the team?
For example, if you replace a player A (who pulls down 5 rebounds a game and no other statistics at all) with player B (who pulls down 10 rebounds per game but has no other statistics at all) that you would expect the total team rebounds to go up by +5 per game?
You’ve said many many times (please forgive me if i’m mis-paraphrasing you) that players reboudn averages are highly correlated year to year. But isn’t the key factory how much it affects total TEAM rebounds?
For example, take player X, a power forward who is average except that they get 5 rebounds per game more than an equivalent player at their position. If a team currently has a completely average power forward (player Y) and signs player X, and gives them their regular minutes. Doesn’t WoW reward player X in terms of ‘Wins Produced’ predictions by implying that the team will have a net gain of +5 rebounds per game with their new improved rebounding proficiency?
Said another way, doesn’t the Wins Produced difference between player X and player Y suggest that the 5 rebound difference between the players has a direct linear impact on the total number of wins produced?
Mountain
January 10, 2008
Jason I said I was interested in both player comparisons and team win predictions.
I also said not every player rating system has or has to have prediction of wins as it goal. But that was a side point and I probably should have kept the focus on the main point.
If /when a formula with shot defense and rebounding splits incorporated in the fashion I described is finalized and data is compiled and ready for comparison to predictive power of other formulas then I may return to the topic. This was a preliminary brainstorming. Thanks for the discussion.
P.S. The method I sketched is not and won’t be “perfect” (even if you went off Synergy video) and I don’t claim it to be. It does not have to be the enemy of the good.
Multiple perspectives can refine the image and relay different information about it.
Jason
January 10, 2008
“For example, if you replace a player A (who pulls down 5 rebounds a game and no other statistics at all) with player B (who pulls down 10 rebounds per game but has no other statistics at all) that you would expect the total team rebounds to go up by +5 per game?”
That’s the ideal situation. I do not know presently how closely this fits team data. But see below as to whether or not this is unreasonable.
“You’ve said many many times (please forgive me if i’m mis-paraphrasing you) that players reboudn averages are highly correlated year to year. But isn’t the key factory how much it affects total TEAM rebounds?”
Yes, that’s the key. If a player adds to his total but the team total doesn’t change, then the additional ‘wins’ don’t materialize. It would give the illusion that other players got worse as the rebounds the new player got must be coming from some deficit by them. Does this happen regularly? Is there a way that player rebound rates can be highly correlated *without* a player adding nearly arithmetically to rebound totals?
I’ve repeated this over and over: these are empirical questions. What I have seen is anecdotal cases to refute it. Anecdotal evidence is that and its highly flawed. I’ve also seen the thought experiments that rather often border on the absurd.
Replacing a 5 rebound player with a 10 rebound player means replacing a below average player with someone who is at the very least good. 10 rebounds per game (given reasonable starter minutes) is a respectable PF/C. 5 rebounds at the same position means someone who is below competent. It could also mean replacing a SF who is shading below average with an elite in the game at that position. I’m presenting this because I want to make sure that someone trying for a ‘smell test’ understands that this hypothetical is a rather monumental upgrade.
What would a monumental upgrade mean to the team though? Would it be absurd at the team level?
It seems like there’s a gut feeling that replacing a guy who pulls down 5 rebounds over the course of his game with a guy who pulls down twice as many somehow cannot mean a net gain of 5 rebounds. There seems to be a gut that some must be coming from somewhere else. Is this true? What does a net increase in 5 rebounds *actually* mean in terms of the team data.
Here’s a reality check to measure it against: Last year, there were ~82 rebounds in each game. 27% of these were offensive rebounds, or about 11.1 a game compared with 29.9 defensive boards. If a net gain in 5 rebounds came at the same proportion as the rest of rebounds, a gain of 5 means a gain of 1.35 offensive rebounds and ~3.65 d-boards. (and a course the opposition’s totals would drop as much so the denominator for rebounding percentage wouldn’t change. This means that the change would result in an offensive rebounding percentage of 12.45/41 or ~30%. That’s a good rebound rate for certain, but it’s not unheard of.
Now the change from a 5 rebound player to a 10 rebound player at the same position is a *huge* change. Realistically, the only way we’ll see that is to have a guy who’s close to inept on the boards being replaced with someone who is pretty good. Do you think that upgrading a 5th of the lineup from poor to good can change a team from being an average rebounding squad to one that’s one of the better rebounding squads? That doesn’t seem unrealistic to me at all.
Jason
January 10, 2008
Mountain, what does a player rating system that *isn’t* related to how a player affects team wins mean? If a player is ‘better’ but doesn’t help a team win more, can we really say that he’s better in any meaningful sense related to the game?
antonio
January 10, 2008
“Do you think that upgrading a 5th of the lineup from poor to good can change a team from being an average rebounding squad to one that’s one of the better rebounding squads? That doesn’t seem unrealistic to me at all.”
In almost every post you write, I see the phrase *empirical evidence*. Can you show the correlation between a players own rebounds and their team rebounds. Wouldn’t this let us know whether 1 personal additional rebound add 1 team rebound. And if numerous anecdotes were to be said, wouldn’t that be enough proof that its not true. I understand one anecdote doesn’t break a model, but at some point if their are many anecdotes, can’t that refute a point?
antonio
January 10, 2008
“Dave has said that he used regression to derive his coefficients. Rosenbaum asserts otherwise. I’m inclined to believe the person who made the model used the method he claimed rather than a counter-assertion.”
SHC has shown that isn’t what Dave is saying, and also it is not just Dave who is saying that. Justin Kubatko, Kevin Pelton, and Dean Oliver also say that and Dave claims it to be peer reviewed. Dave does not stand alone in this argument he is making
Mountain
January 10, 2008
Jason, a player rating system that isn’t set up and driven by how a player affects team wins is an individual performance comnarison system and I think dated and lesser than one that reconciles with team wins, Given the ability to check predictive power and the example and challenge posed by Dave Berri’s work and the later discussion of it I agree it is worth aiming higher than in the past. Earlier after saying “I am interested in both player comparisons and team win predictions.”, I said “Doing both from the same method would make sense.” This really isn’t a point of disagreement about what the future should be. I was describing the past and present that we can move beyond.
Mountain
January 10, 2008
(Sorry for the typos.)
On substance I should go back in and say …
a player rating system that isn’t set up and driven by how a player affects team wins is an individual performance comnarison system and I think dated and lesser “in original ambition”. But old methods can still be tested and scored relative to newer methods. I won’t debate the scores on one test right now. I’ll move on to say that these old methods can be reviewed, critiqued, dispensed with in lieu of better alternatives, used with caution, used in parallel for the different insights they raise for consideration or tweaked for improvement or blended or substantially superceded with new elements, frameworks, etc.
In additon to the tools offered here.
I try to use many tools to sharpen my perception and opinions.
Mountain
January 10, 2008
Or if you were explaining / defending past formula you could suggest that they were implicitly assuming that their best player rankings would align fairly well with team strength measures- at least on offense, but didn’t (often or at all) show how well they did. Past formula (at least the simple forms) lacked shot defense and couldn’t reasonably claim to explain team wins fully without that.
Mark
January 10, 2008
Thankyou very much Jason.
It seems like this is the key contention:
““You’ve said many many times (please forgive me if i’m mis-paraphrasing you) that players reboudn averages are highly correlated year to year. But isn’t the key factory how much it affects total TEAM rebounds?”
Yes, that’s the key. If a player adds to his total but the team total doesn’t change, then the additional ‘wins’ don’t materialize. It would give the illusion that other players got worse as the rebounds the new player got must be coming from some deficit by them. Does this happen regularly? Is there a way that player rebound rates can be highly correlated *without* a player adding nearly arithmetically to rebound totals? ”
Like you I’ve only seen anecdotal evidence of team rebound changes vs individual player rebound changes.
Is is possible for some stathead to crunch the numbers and compare for players who change team between seasons, how much his personal rebound rate changes, what the player he’s replacing produced, and what effect this has had on total team rebounds?
SHC
January 10, 2008
So I was reading about Rosenbaum’s explanation of how wins-produced is computed.
It seems that the main thing that wins-produced does is find the relation ship between the number of wins and points differential. – This is extremely simple to compute and I think that everyone does this already.
Now this was nothing new. However, dberri then computed the relationship of possessions to the points. It turns out in the average, each possession produces approximately 1.06 points or something. So you could approximately say that points == number of possessions.
Then he derived coefficients for the other things from possessions. Let’s say that # of possessions for the own team is ‘OwnFGattempts – OwnOffensiveRebounds + OwnTurnovers’ and the # of possessions for the other team is ‘OppFGmade + OwnDefensiveRebounds + OppTurnovers’ (simplified to exclude FT attempts). Then, if you subtract # of opponents possessions from # of own team possessions, you should approximately get points differentials. You could simply use the coefficient that relates point differentials to # of wins you learned here!
What’s funny here is that if you compute coefficients this way, you end up excluding very important items in winning, namely failure to stop the other team from scoring – there’s no individual stat number for this – you can only find team-wide defensive statistics. So this goes into the ‘team defense adjustment’ that dberri did, and I don’t think that he has published numbers for team defense adjustments.
Rosenbaum argues that without the team defense adjustments, wins-produced would be very inaccurate and this makes a lot of sense, because you’re essentially taking out opponents made field goals from the equation and the whole equation doesn’t make a lot of sense.
So if the defense is so important and there’s no way to compute this into each player’s wins-produced, wins-produced becomes a highly inaccurate measure of an individual performance.
TG Randini
January 10, 2008
Watch out SHC! Your making way too much sense! ‘Jason’ might think your “gray matter needs an upgrade”.
How childish, Jason. Your petty insults always remind me of an eighth grader’s last defense when they feel they are losing an argument.
I’ll go IQ against IQ with you – independent testing – for $100k winner take all. Then we’ll see whose “gray matter” needs a lift.
SHC
January 10, 2008
antonio’s link also provides an excellent argument to the ambiguity of wins-produced.
Rosenbaum shows that you could compute the correct number of possessions in almost an infinitely different number of ways by allocating different weights (using the same alpha) among missed shots, offensive rebounds, and defensive rebounds. However, if we use a symmetric equation for both own and opponents’ possessions, it’s not just team-defense adjustment that’s needed – we also need numbers for preventing the other team from grabbing rebounds (that again don’t exist).
So in essence, wins-produced is highly arbitrary, because you could’ve come up with wins-produced that could predict exactly the same way the current wins-produced does – however, you’d need some team-defense, team-boxing out adjustment, and not just ‘team-defense adjustment’. Rosenbaum makes perfect sense IMO.
Owen
January 10, 2008
SHC – . Look, no offense, but it really isn’t necessary for you to recapitulate Rosenbaum’s arguments. He has come on this site numerous times under many different aliases and made his case, marginally better than you are able to. We can also read APBR and see what he has written there.
If you have original thoughts, please share, but if not, be aware you aren’t really going to be adding anything worthwhile to the conversation.
SHC
January 10, 2008
Well, I didn’t know that he already made his case on this site numerous times. My apologies.
However, I just wanted to point out that his arguments are mathematically true, and that wins-produced is just one manifestation of numbers you can get by substituting some arbitrary division of credits among missed shots, defense rebounds, and offense rebounds.
I had a feeling that wins-produced was at least very scientific and provable (at least in terms of numerical stats), but it’s shocking to find out that there are just too many arbitrary heuristics used in the computations.
dberri
January 10, 2008
SHC,
Despite what you think you have learned, what you say is incorrect. Wins Produced is not an arbitrary model. In a few months the paper that explains Wins Produced will be published and you can read the details then.
SHC
January 10, 2008
Prof. Berri,
I am looking forward to the paper (I’m surprised that there has not been any yet).
However, unless my fundamental understanding of wins-produced is falwed (i.e. coefficient is mainly derived from points differentials and wins/losses and then possessions are used to derive the rest), this model ‘is’ arbitrary in evaluating relative importance of statistical categories (and of course in evaluating individual players), no matter what kind of justification you can provide for the choices.
If I’m wrong, it’d help if you could even briefly provide some counter arguments to Rosenbaum’s.
Thank you
Owen
January 10, 2008
SHC – Dave is writing a book. So give him a break. From what you have written, it very much seems like you haven’t picked up a copy of The Wages of Wins. Great book, great read. Why don’t you get it, read it, and see if it answers your questions. It might be helpful to look through the old posts on this site also. “A Guide to Evaluating Models” might be very helpful, as well as “Incorporating Team Defense.”
dberri
January 10, 2008
SHC,
Your understanding of the model is flawed. What you describe is not how this model was put together. Again, the paper will be out in a few months. This paper fleshes out the theory behind the model (and also re-hashes some stuff that has already been published).
SHC
January 10, 2008
Owen, I’ve read those two posts, but they are not very explicit about how things are computed. I guess that I’m spending too much time here, and should just wait.
Looking at ‘Incorporating Team Defense’, I get the feeling that # of possessions were not used as a constraint, but rather the coefficient for every single category was learned from data.
My take on this is that the equation is like this:
a * ownFGM + b * ownFGMissed + c * ownOffensiveRebounds + d * ownDefensiveRebounds + …… + a’ * oppFGM + b’ * oppFGMissed ….. = points-differential
Then you could simply use least squares to find a, b, c…, a’, b’, c’…. However, the problem is still combining opponents’ statistics to be reflected in the own team’s statistics to compute wins-produced. This will have to involve some heuristics.
TG Randini
January 11, 2008
A haiku about why credit for rebounding should be apportioned across the team:
Basketball rebounds.
Last stage of defensive stop.
Is so beautiful.
SHC
January 11, 2008
I just wanted to correct my previous post.
FGMissed, OffensiveRebounds, OppDefensiveRebounds are most likely related by the equation FGMissed – OffensiveRebounds ~= OppDefensiveRebounds. So solving the equation above will not give you a unique solution, but there will be an infinitely many solutions.
So even if you start from pure linear regression for all categories, you reach the same conclusion as Rosenbaum… You need to constrain this (with reasoned or arbitrary heuristics) if you want to get a unique solution.
The Real TG Randini
January 11, 2008
Nice poem. I can’t take credit for it, but thanks to the one who posted it. I couldn’t have said it better.
Christopher
January 11, 2008
Given the thread hijackcing and some of tone here, I am hardly surprised dberri does not go to APRB. Have some of you listened to yourselves rant? That’s the problem with blogging, everyone has some pet theory they think is so special but not the data analysis skills to flesh it out. So now we get trivial haikus and an overblown sense of self-importance. The SNR of the comments is just a waste. Get a life, eh…
Mountain
January 11, 2008
Christopher, rebounding played an important role in the article as it plays an important role in WP. TG was interested in that aspect and shot defense and added his perspective. I added my own thoughts on those subjects.
I personally do not consider that shift in and of itself to be threadjacking, I consider that moving to a different level of an on-going discussion. But if tone or length seem excessive to you then ok that’s your view.
I am all for civil discussion and know from being a fairly regular reader here that it it hasn’t always gone that way here, that many comments come in repeatedly rejecting the WP approach on rebounding or shooting from the jump and in the form of declaration rather than cordial discussion. I can understand the frustration with that.
But folks can declare their viewpoint in a forum. And folks can express their ideas and leanings without having to be top level statistical analysts pursuing all their ideas to completion to have a voice in the discussion. Folks are different and their role in the discussion are different. Some more influential than others, some friendlier in tone than others. But there are occasional tone issues on both sides. Everyone was to sort out the signal to noise issue themselves.
I am tried to participate on this issue a bit here as well as APBR since it came up both places at same time. This is a busy thread and I am not taking all critical comments as being about my part in it. But in general if contrary or non-committed fully to WP voices are not particularly welcome (given the history) then I or anyone can process whether to participate further here or not. I got other places I can be and plenty else to do. I just took a little time to participate on what interested me at the time.
Polar Bear
January 11, 2008
Mountain, I don’t think Christopher’s comments were directed towards you or any of the others that have composed well-written and on-topic comments (wher positive or negative towards WP) that contribute to the discussion. I believe Christopher’s comments were aimed towards those that fill up the comments with non sequiturs (like haikus) or repetitive off-topic trolling.
Mountain
January 11, 2008
Maybe so. Some things do go too far away from acceptable, as I indicated awareness of. Still felt my post added some balancing perspective.
Look forward to reading future posts and discussions and may join in if it seems fitting.
Polar Bear
January 11, 2008
Mountain, I was including your post in the category of “well-written and on-topic comments.” I didn’t meant to imply otherwise.
Polar Bear
January 11, 2008
Incidentally, that was a well-written haiku. Adheres to 17 syllables and concisely presents an argument. Even if it’s completely gratuitous.
Mountain
January 11, 2008
I agree with Dave’s presentation that Paul and Chandler are the biggest keys of the Hornets success. There probably isn’t much opposition to that, especially here, hence not much discussion of it.
But to acknowledge the topic and maybe make amends a bit for the shift, here is some.
In the yahoo article Paul calls West the “The 17-foot assassin”. His NBA hotshot chart does confirm him as very strong from straightaway to left wing on long jumpers (about 60% FG).
But overall his FG% on all jump shots is 41% eFG. Nice enough for a PF but, except for his sweetspot from near top of key,
jumpers from him are not special – back out the sweet spot data from the jumper total and it would be somewhere in 30%s and these shots from elsewhere are not bread & butter for the team to live on.
Still overall his shooting/scoring does help keep a 3rd offensive force on the court most of the time (among Paul, West, Peja, Peterson). That on paper is good strategy and team offense with him on court is among the strongest group.
West’s impact as a 3rd scoring force might be creating value beyond his individual numbers via multiple threats and floor spacing. Team eFG% is almost 4 percentage points higher with him on than off. That is a bit less than the other scorers but still quite strong.
His raw +/- and on court win % are comparable to Paul and Chandler this season and a big reason is that the big 5 play together as a unit a lot (3rd most used unit in league) and win 77% of the time. West may not that great individually but he fits very well. He has strong player pairs with all 4 other main guys.
Perhaps a good amount of credit and praise should go to GM and coach for the team construction..
Mountain
January 11, 2008
It will be interesting to see the pure adjusted +/- rating for the Hornet leaders this season and where West ranks. In 05-06 West ranked on e of th eworst. Last season (plagued by injuries and may not a good test) it was near neutral. Will this year’s ranking show continued upward movement or little impact or less favorable? Another perspective beyond WP that may be considered by those who wish to.
Owen
January 12, 2008
Mountain – Interesting analysis. David West is yet another example of the usage/efficiency problem. What is the value of having a jumpshooting power forward who isn’t very efficient offensively, at just 52.2% ts%. To me he looks like a little like Zach Randolph minus the horrible selfishness, personality disorder, max contract, and unbelievable inability to play defense.
Obviously, it’s better to have West taking shots than Pargo etc, but I find it difficult to make a case for West being a very good player. I would rather have him than Randolph, just on personality, but is he better than David Lee? More precisely, would the Hornets get worse with Lee on the court next to Chandler? It’s an interesting question, it’s too bad those kind of player experiments don’t happen all that frequently.
My suspicion these days, undoubtedly heavily influenced by Zach’s arrival in New York, is that a lot of frontcourt players take a lot of perimeter shots because they understand they will get paid a lot more money if they score 20 points per with low efficiency, while rebounding competently, than 14 points with high efficiency. That is certainly how Zach Randolph got his max contract.
I don’t know if that is the case here. West has his contract and to my eye he looks to be fairly paid, at least from reading Ryan Schwan’s piece last year on West it seems that way. He is decently efficient…
Anyway, rambling, go Giants…
http://thehornetsfan.blogspot.com/2007/09/david-west-worth-every-penny.html
Mountain
January 12, 2008
Yeah the motivation to score to get paid big is a factor. Thanks for the link. I didnt know West’s deal was one that declined over time. Given the market for scoring bigs (partly based on belief in impact of that on rest of team’s efficiency) that deal is a pretty good one.
West may be better beside Chandler than Lee would be- as we know him in NY. I don’t know Lee well but I wonder if he could give more scoring if given such a role. Maybe next contract we find out.
Eliot
January 12, 2008
Thing with Lee is he has problems with fouls and he will have trouble staying in games for significant amount of time if/when he becomes a consistent starter. He also has a very high energy style and he may become worn out with major minutes over a season. If he manages to cut down on his fouls he could be a nice piece on a good team.
Mark
January 13, 2008
Perhaps the other thing with lee, chandler, wallace etc is that they are ‘non-scoring’ bigs.
A team can support one such player, but having two on the court at the same time puts too much of a dent in team scoring.