APBRmetrics

Though I find it really ironic that in a aforementioned link your own first words were "I’m sure you agree that this discussion can only be productive if everyone engages the actual arguments made by the other side" )

In what possible way is that "ironic?"

As people from WoW here! I mean the guy who started this topic got offended literally in the first reply... and he started with a premise of making changes in WP! Not surprisingly he quickly left.

Didn't leave, wilq, just got busy. Plus I've been working a bit on refining what I originally introduced in this topic.

Jon Nichols wrote:Just curious: do any of the writers of wagesofwins.com or its various related sites ever attend the MIT Sloan conference? I know most of them rarely participate on this board. I for one would be curious to meet one of those folks and talk about these issues.

Jon: I noticed your exchange with Arturo over at WOW. Just FYI, the correlation between wins and WP does decline if you omit the team adjustment. Berri finally reported this in a discussion thread, after resisting a long time. I don't recall the result, but it was well below .95 -- in the .80-.90 range I think.

Ironically, a high correlation w/o the team adjustment is not something they should brag about. Since opponent eFG% accounts for about 25% of wins, non-adjusted WP should max out at a correlation of c. .70. The fact that the correlation is much higher is a problem, not a vindication.

Guy wrote: Ironically, a high correlation w/o the team adjustment is not something they should brag about. Since opponent eFG% accounts for about 25% of wins, non-adjusted WP should max out at a correlation of c. .70. The fact that the correlation is much higher is a problem, not a vindication.

Doesn't that depend on what all is in the model? If something else in there correlates with opponent eFG%, you could go over .7 even leaving it out.

xkonk wrote:Doesn't that depend on what all is in the model? If something else in there correlates with opponent eFG%, you could go over .7 even leaving it out.

WP isn't supposed to work that way. The coefficient on each variable is supposed to represent the value of that item (e.g. a rebound), holding everything else constant.

What's really happening, of course, is that players who get defensive rebounds are essentially being credited for causing the missed shot. That's how you can get a high correlation without accounting (openly) for opponent eFG%.

mystic wrote:wilQ, I know it is always convienent to think of oneself as being objective and thus have an overview about everything, but you are wrong. First, we have people HERE who are arguing against WP who use boxscore based metrics. Take me for one, I have a metric completely based on the boxscore, and yet, I argue against WP. You just want to see a pattern, you want to put people into a certain box and then apply your ideas to it, because it just fits your personal feeling.

A debate about WP happens like years ago, right now the people can just repeat the argumentation, because in essence WP has the same flaws despite making that adjustment regarding the defensive rebounds. And while you might not see it, we had a discussion with pro WP people here just a couple of month ago, you are just not paying attention. Usually, the discussion gets less, when one side has the better arguments. Your conclusion upon this is: here are only APM people who despise WP, while the reality is that here on APBR is much more diversity. Mike is basically only posting the results of his boxscore metric, Evan has a play-by-play based boxscore metric, myself has a boxscore metric, DSMok is using a boxscore metric, which is just based on regression on RAPM values. Seriously, if you think here are only APM people around trying to trash WP, you are obviously not paying attention at all.

And indeed, wilQ, the writers on that blog are getting probably paid by Berri for providing a content which suits his way of thinking and his books. People, questioning WP and the arguments are getting blocked from the comment section, you may haven't notice that, but I can't comment over there, Evan is blocked and I guess others as well. The site is censored, when you ask question and explain things in too much detail, you get thrown out. There is a reason Berri set a limit on the characters and that limit is not there to improve the debate about WP.

So, do you see now why I said that you haven't looked at the argumentations?

I'm blocked from posting on his blog as well. I had to use some pseudonyms to post, but got banned shortly thereafter. Now I don't care, but it's mind boggling how much attention is paid to WOW.

Guy wrote:

xkonk wrote:Doesn't that depend on what all is in the model? If something else in there correlates with opponent eFG%, you could go over .7 even leaving it out.

WP isn't supposed to work that way. The coefficient on each variable is supposed to represent the value of that item (e.g. a rebound), holding everything else constant.

What's really happening, of course, is that players who get defensive rebounds are essentially being credited for causing the missed shot. That's how you can get a high correlation without accounting (openly) for opponent eFG%.

I think the way that WP is supposed to 'work' is within the framework of what is specifically in the model. If you start changing the variables, given that there are correlations between the predictors (like defensive rebounds and opponent eFG), obviously things like the correlation with wins or the weights on the predictors are going to change. But you could have alternatively taken defensive rebounds out and presumably opponent eFG would then start soaking up some of that variance, and we could have a discussion about how WP correlates with wins higher than it should given however much defensive rebounds correlates with wins. Presumably, if I started playing with Daniel's ASPM and taking out a variable here or there, we would see something similar with its relationship to RAPM.

xkonk wrote:I think the way that WP is supposed to 'work' is within the framework of what is specifically in the model. If you start changing the variables, given that there are correlations between the predictors (like defensive rebounds and opponent eFG), obviously things like the correlation with wins or the weights on the predictors are going to change. But you could have alternatively taken defensive rebounds out and presumably opponent eFG would then start soaking up some of that variance, and we could have a discussion about how WP correlates with wins higher than it should given however much defensive rebounds correlates with wins. Presumably, if I started playing with Daniel's ASPM and taking out a variable here or there, we would see something similar with its relationship to RAPM.

I'm really not sure what you're talking about. No one has "changed the variables" in WP. This is pretty simple: prior to the final step of adding the team defense adjustment, WP shouldn't be accounting for opponent efficiency (except in a minor way via blocks). And yet, somehow it does this to a significant degree (or completely, if you believe Arturo). This means that opponent efficiency is being captured elsewhere, which is in dreb's. Since I don't believe there is anyone -- even among WP supporters -- who would argue the rebounds-cause-misses theory (rather than misses-cause-rebounds), this means WP is misallocating credit for opponent efficiency.

The claims they make about accounting for 95% of wins, etc etc, depends on the entire framework. If you take something away, like the defense adjustment, then it becomes difficult to talk about correlations/R squareds because of the correlations between the predictors. That's all I was saying. Similarly, if I were unhappy with ASPM, I could try adding a defense component or removing an interaction or what have you. Since the predictors are correlated, the R squared with RAPM will not go up or down as much as you would expect given the simple correlation between RAPM and what you changed.

Guy wrote: What I don't understand is why you seem to think this site should provide such an exchange to you. You're very late to the party, and most/all of the issues have been hashed over at length. Are half of the posters here supposed to pretend to like WP for your entertainment/edification? Why? Would you go to a biology website and expect a "balanced" discussion of intelligent design?

A good analogy. Similarly, we don't see discussion of the earth being flat on geology and physics sites, and someone who goes to such sites and bewails the lack of opposing viewpoints is ... not understanding what those websites are about.

Guy wrote: The pages of this site recently recovered by Daniel include at least 2 long threads on WP from 2006 and 2007. There are probably others.

Yes, indeed the discussion goes way farther back than that. Dave Berri first presented the regression results that would eventually lead to WP on the old apbr email list (I don't know if we even had a separate statistical email list in those days) ... I can't even begin to remember how many years ago it was. At that point it was a new result for him and for us in APBR. At that point there was the back-and-forth and open discussion one would expect (although even then I recall Berri as being not a very good listener). His main innovation was bringing in a functional form from econometrics, namely the Cobb-Douglas production function. The main weakness of his model, which has persisted to this day, is that he was taking regression results from team stats and assuming they could be applied directly to individual players' stats -- leading to results that didn't make sense, most famously Dennis Rodman being the best player in the NBA according to his results.

Thus the back-and-forth discussion came to the conclusion that Berri's model had severe weaknesses, including the theoretical weakness of directly applying team stats to individuals, and the empirical weakness that nobody believed that Dennis Rodman was the NBA's best player.

The way that research is supposed to progress is that when a model has severe flaws, the modeller is supposed to go back and tinker with the model or discard it in favor of another one, or if he is going to continue to use the model, to acknowledge its weaknesses.

IIRC this was in the late, or maybe even mid, 1990s. We had the discussion. Berri AFAICT ignored the commentary, and proceeded to push his model even harder with his book and website. We do get some discussions here that come from or are provoked by the WOW website, probably 2 or 3 times per year. But they're typically not about Berri's model -- what else is there to say? We had the discussion years ago. Instead they are on subsidiary but sometimes still interesting topics such as how to evaluate models (accuracy of prediction is an important factor, but it has to be out-of-sample prediction, and no single test is sufficient for us to conclude which model is best -- it simply tells us which model had the best prediction for that set of data).

Apparently they sell t-shirts.

http://wagesofwins.com/2012/09/08/wages-of-wins-store/

Maybe APBRmetrics should start selling T-shirts. Get some enterprising artist to make a logo of a shivering ape holding a basketball in one hand and a ruler in the other.

kjb wrote:Maybe APBRmetrics should start selling T-shirts. Get some enterprising artist to make a logo of a shivering ape holding a basketball in one hand and a ruler in the other.

I like it! Maybe I'll design a shirt like that...

kjb wrote:Maybe APBRmetrics should start selling T-shirts. Get some enterprising artist to make a logo of a shivering ape holding a basketball in one hand and a ruler in the other.

Or some dude watching a game with a pair of eyes in one hand and a calculator in the other.

EvanZ wrote:

kjb wrote:Maybe APBRmetrics should start selling T-shirts. Get some enterprising artist to make a logo of a shivering ape holding a basketball in one hand and a ruler in the other.

Or some dude watching a game with a pair of eyes in one hand and a calculator in the other.

Make it a shivering ape and you've got a deal.