APBRmetrics

v-zero wrote: There's nothing wrong with the method, it was never intended to take everything into account, only to look at the league as a whole, using nothing other than data available from previous box scores prior to each game that is the simple linear combination of box score per 48 stats which best predicts the average minute allocations.

Well, there is obviously everything wrong with your method, because you are making an assumption before the fact, the assumption that coaches would assign minutes by using some sort of cumulated stats. I have a method which is much better than your method in order to predict minutes: average minutes played in previous games. It is way simplier than your method, it is not biased towards offense or defense (the boxscore is actually biased towards offense) and reflects coaching decisions pretty good. Why? Because coaches tend to use the same players in the same roles throughout the season unless they are forced by injuries to change something.

v-zero wrote: I really don't care that you think coaches aren't that naive, I believe that the game has a few great coaches and a large number of mediocre coaches,

Assuming a Gauss distribution for coaches is reasonable, and has actually nothing to do with the overall quality of the coaches. Whether coaches are naive or not is not something we can determine by the distribution.

v-zero wrote: but I didn't create this stat with any bias in mind, I just wanted a useful, simple equation.

Well, you might not have that bias in your mind, but you should have taken care of the bias of the sample. And, matter of fact, the equation is rather complicated given the fact that the average minutes played in previous games is giving you a better prediction.

v-zero wrote: If you use it throughout the season to predict minutes for each game (with knowledge of player availability) it has around a 0.7 coefficient of determination, and that is in predicting, not retrodicting.

If you take "availability" into account (meaning, you adjust the predicted minutes to sum up to 240 min for the whole team), you will easily get a R²>0.9 by using mpg.

v-zero wrote: I'm not looking for you to agree, but you can't accuse the numbers of lying because there is a place in your heart for the eye test.

I never "accused the numbers to lie", it is just a case of failed interpretation, because you have no clue where your R²=0.7 is coming from. Also, you shouldn't engage into the illusion of transparency, because I can assure you that I have no place in my heart for the eye test.

Your arguments fall down on the fact that you have willingly ignored what I have said I was trying to do in creating this stat, you assume things in your argument that are in opposition to what I've actually said, and go on to explain a solution to a problem I'm not trying to solve.

By the way, taking care of availability actually meant not including injured players in games for which they weren't eligible.

v-zero wrote:Your arguments fall down on the fact that you have willingly ignored what I have said I was trying to do in creating this stat, you assume things in your argument that are in opposition to what I've actually said, and go on to explain a solution to a problem I'm not trying to solve.

I'm pointing out that your method is insuffiecent to explain the minute distribution and thus not able to let you conclude that coaches don't have a clue. It is simply a matter of a lack of information and misinterpretation on your part. Just think about why you are so far away in terms of accuracy from the mpg-method.

The main reason for your R² is not that you found a method which can predict minute distribution very well, but that the minutes are given out consistently by the coaches to the same players. That is an inherent character of the system you are working with, and not something you discovered and then gives you the ability to evaluate coaches.

v-zero wrote: By the way, taking care of availability actually meant not including injured players in games for which they weren't eligible.

That much should be clear, but when you have a player with 40 mpg missing a game, the remaining players would not add up to 240 minutes for that specific game. That's what I meant with the adjustment.

mystic wrote:

v-zero wrote:Your arguments fall down on the fact that you have willingly ignored what I have said I was trying to do in creating this stat, you assume things in your argument that are in opposition to what I've actually said, and go on to explain a solution to a problem I'm not trying to solve.

I'm pointing out that your method is insuffiecent to explain the minute distribution and thus not able to let you conclude that coaches don't have a clue. It is simply a matter of a lack of information and misinterpretation on your part. Just think about why you are so far away in terms of accuracy from the mpg-method.

The main reason for your R² is not that you found a method which can predict minute distribution very well, but that the minutes are given out consistently by the coaches to the same players. That is an inherent character of the system you are working with, and not something you discovered and then gives you the ability to evaluate coaches.

v-zero wrote: By the way, taking care of availability actually meant not including injured players in games for which they weren't eligible.

That much should be clear, but when you have a player with 40 mpg missing a game, the remaining players would not add up to 240 minutes for that specific game. That's what I meant with the adjustment.

It was never forced to add up to 240, but that doesn't mean it couldn't. It does predict minutes pretty well, it uses no information about who plays for what length of time as it merely uses per-48 values for the box score stats, so the 'you're just seeing the players coaches like playing' argument is wrong - rather the equation is fitting itself to the kinds of players coaches really love to have on court.

v-zero wrote: rather the equation is fitting itself to the kinds of players coaches really love to have on court.

Just the simple truth that the boxscore is biased towards offense makes that conclusion wrong. You could say that, if you had an unbiased sample and would have the same ability to predict minutes as the mpg-method. Neither of that is true, and you really need to understand that, otherwise everything you conclude is basically useless.
What you found is some sort of statistical relationship between a supposed to be dependent variable and some independent variables, but you haven't shown any kind of causality here.

I know that it is convenient to think of oneself to have superior knowledge than people working in a specific field, but that is most times not even close to be true. Take your idea of Kenneth Faried as an example: He has yet to show that his presence can actually improve the performance level of the Nuggets on a consistent basis. The matter of fact is that the Nuggets played nearly 5 points worse with him on the court than without him for his career. So, what exactly is he really bringing to the table in terms of winning? Is such a player a good fit in order to build lineups which can outscore the other team's lineups? Is that something which came to your mind at least once?

For the first time in several years, perhaps, a new top dog in the field. Season at 12.4%

mystic wrote:
I know that it is convenient to think of oneself to have superior knowledge than people working in a specific field, but that is most times not even close to be true. Take your idea of Kenneth Faried as an example: He has yet to show that his presence can actually improve the performance level of the Nuggets on a consistent basis. The matter of fact is that the Nuggets played nearly 5 points worse with him on the court than without him for his career. So, what exactly is he really bringing to the table in terms of winning? Is such a player a good fit in order to build lineups which can outscore the other team's lineups? Is that something which came to your mind at least once?

This is sort of interesting in that Faried appears to be above average, both this season and last, by every measure I can find (Mike's post below yours, Win Shares, WP, PER, the RAPM currently on Jerry's site). Why do you say the Nuggets fail to benefit from his play?

mystic wrote:

v-zero wrote: rather the equation is fitting itself to the kinds of players coaches really love to have on court.

Just the simple truth that the boxscore is biased towards offense makes that conclusion wrong. You could say that, if you had an unbiased sample and would have the same ability to predict minutes as the mpg-method. Neither of that is true, and you really need to understand that, otherwise everything you conclude is basically useless.
What you found is some sort of statistical relationship between a supposed to be dependent variable and some independent variables, but you haven't shown any kind of causality here.

I know that it is convenient to think of oneself to have superior knowledge than people working in a specific field, but that is most times not even close to be true. Take your idea of Kenneth Faried as an example: He has yet to show that his presence can actually improve the performance level of the Nuggets on a consistent basis. The matter of fact is that the Nuggets played nearly 5 points worse with him on the court than without him for his career. So, what exactly is he really bringing to the table in terms of winning? Is such a player a good fit in order to build lineups which can outscore the other team's lineups? Is that something which came to your mind at least once?

Bias doesn't invalidate conclusions, it simply makes them incomplete. I'm really sick of asking you to realise that I do understand the limitations of the metric. There is no magic number in the box-score that can improve it, and all I wanted to use was the box score. You wouldn't have the same ability to predict minutes as mpg even if you had a perfect sample, because there is no way to model coaching decisions perfectly, because people and their decisions aren't that simple - it would be better, it would not be as good as minutes per game.

You see, you can make your statements about Faried and I can make mine. His RAPM may be poor, but that doesn't mean he is not adding to his team, because guess what? RAPM is both biased and imprecise, especially for young players with few minutes and a quickly evolving game. RAPM is not a panacea, not even close, it's a nice idea and might be fantastic if the season was one hundred thousand games long, but it isn't, and I'm bored of it being used as a crutch. I believe in the stats that are predictive, and build them as such.

I have no interest in continuing this discussion because neither of us is going to admit we are wrong, so we are wasting each other's time. I believe you have failed to understand this, you believe I have failed to frame things properly, at this point we should stop.

Mike G wrote:For the first time in several years, perhaps, a new top dog in the field. Season at 12.4%

Code: Select all

e82     per36 rates   tm   Min   Eff%   Sco   Reb   Ast   Stl   TO   Blk   e484
14.6   Durant,Kevin   Okl   39  .611   26.8  10.8   5.0   1.5   3.6  1.3   2.22
14.1   James,Lebron   Mia   36  .571   26.4  10.1   7.0    .9   2.1  1.0   2.28
13.3   Paul,Chris     LAC   33  .584   21.5   4.1  12.4   2.5   2.4   .0   2.40
12.9   Bryant,Kobe    LAL   36  .626   31.1   6.0   6.0   1.5   3.4   .0   2.10
11.9   Howard,Dwight  LAL   36  .576   22.9  12.2   2.0   1.0   3.5  2.8   1.97

11.8   Duncan,Tim     SAS   31  .553   24.0  12.8   3.0   1.2   1.6  3.2   2.27
11.2   Gasol,Marc     Mem   37  .583   17.4   7.7   5.1    .7   1.0  1.3   1.81
11.2 Kirilenko,Andrei Min   36  .655   18.0   8.9   3.4   1.5   2.9  2.3   1.82
10.6 Westbrook,Russel Okl   36  .485   21.8   5.2   9.9   1.7   2.9   .1   1.74
10.5   Noah,Joakim    Chi   39  .550   15.6  10.2   3.8   1.2   2.6  1.9   1.60

e82     per36 rates   tm   Min   Eff%   Sco   Reb   Ast   Stl   TO   Blk   e484
10.4   Rondo,Rajon    Bos   38  .555   13.8   4.8  12.8   1.8   3.5   .0   1.78
10.2 Varejao,Anderson Cle   36  .563   15.0  14.7   3.2   1.4   1.3   .5   1.88
9.6   Griffin,Blake   LAC   32  .521   21.8  11.9   4.3   1.5   3.7   .8   1.77
9.6   Batum,Nicolas   Por   40  .606   20.6   6.1   2.8   1.9   2.5  1.1   1.44
9.6   Randolph,Zach   Mem   37  .517   17.7  14.4   1.1    .7   2.3   .7   1.52

9.5   Bosh,Chris      Mia   33  .614   24.2   9.1   1.7    .9   2.1  1.5   1.72
9.5  Jennings,Brandon Mil   35  .503   17.9   3.5   7.9   3.3   2.4   .3   1.59
9.5   Anthony,Carmelo NYK   36  .519   27.0   8.1   1.8    .9   2.3   .6   1.56
9.3   Monroe,Greg     Det   34  .524   18.9  11.5   3.5   1.8   3.1  1.1   1.61
9.3   Holiday,Jrue    Phi   38  .534   20.7   3.9   8.1   1.4   4.7   .5   1.44
 
e82    per36 rates    tm   Min   Eff%   Sco   Reb   Ast   Stl   TO   Blk   e484
9.2   Conley,Mike     Mem   34  .585   18.5   3.6   8.0   2.2   3.3   .1   1.62
9.2   Williams,Deron  Brk   35  .552   21.6   3.0   9.0    .9   3.9   .4   1.55
8.8   Harden,James    Hou   38  .546   24.6   4.2   4.5   1.3   3.9   .6   1.38
8.8   Jefferson,Al    Uta   32  .479   17.9  13.6   2.3   1.2   1.8  1.1   1.60
8.7   Millsap,Paul    Uta   31  .534   19.6  11.1   2.9   1.1   2.4  1.6   1.64

8.5   Dragic,Goran    Phx   34  .563   18.2   3.2   7.1   2.2   2.4   .2   1.49
8.5   Lopez,Brook     Brk   30  .550   24.9   8.8   1.0    .6   2.0  3.5   1.69
8.3   Walker,Kemba    Cha   36  .508   19.3   3.8   5.2   2.6   2.2   .3   1.34
8.2   Smith,J.R.      NYK   34  .575   19.8   5.8   3.5   1.7   1.3   .6   1.44
8.2   Pierce,Paul     Bos   33  .534   22.4   7.5   3.2   1.6   2.1   .3   1.45

e82    per36 rates    tm   Min   Eff%   Sco   Reb   Ast   Stl   TO   Blk   e484
8.2  Aldridge,Lamarcu Por   39  .480   19.2   7.6   3.5    .8   1.9  1.3   1.25
8.1   Faried,Kenneth  Den   30  .544   17.9  14.0    .4   1.1   1.8  1.1   1.60
8.0   Teague,Jeff     Atl   27  .580   21.5   3.8   8.6   2.1   4.2   .8   1.73
8.0   Mayo,O.J.       Dal   35  .638   26.0   3.7   3.2    .7   2.7   .2   1.35
8.0   Ellis,Monta     Mil   36  .489   22.3   3.7   5.9   1.4   3.1   .7   1.32

8.0   Felton,Raymond  NYK   33  .481   19.2   3.4   8.6   1.6   2.5   .0   1.44
7.9   Kidd,Jason      NYK   25  .768   16.1   3.5   5.1   2.6    .9   .6   1.87
7.8   Gasol,Pau       LAL   37  .483   14.3   9.6   3.5    .7   1.7  1.4   1.24
7.8   Irving,Kyrie    Cle   35  .557   24.4   4.2   5.3   1.2   4.3   .3   1.31
7.8   Lee,David       GSW   37  .479   15.3  11.1   3.5   1.1   2.4   .2   1.24
7.8   Gay,Rudy        Mem   35  .496   22.2   6.1   2.6   1.6   2.4  1.3   1.30

Has the formulation of eWins been made public?

xkonk wrote: This is sort of interesting in that Faried appears to be above average, both this season and last, by every measure I can find (Mike's post below yours, Win Shares, WP, PER, the RAPM currently on Jerry's site). Why do you say the Nuggets fail to benefit from his play?

The xRAPM is biased by the boxscore, thus, it is not really a good way to estimate the effect (biased towards height for example). WP is obviously thinking that Faried just used up 89 possessions to score 154 points, which makes him an offensive juggernaut, but in reality he isn't. Faried can't create offense for himself or others in halfcourt sets, that limits the possibility to put his teammates into better positions for themselves. He can't shoot from the outside, which does not help in terms of overall floor spacing. Those are some of the reasons that his offensive rebounding does not lead to a heavily increased offense. He is also a pretty poor 1on1 defender, he constantly allows his direct opponent to be in a better position to score. Crashing the offensive glass at every opportunity will also lead to worse transition defense. Overall, Faried's skillset is not something which helps a team to be above average with him on the court. Overall that leads to the fact that the Nuggets in average for the last two seasons got outscored during the time Faried was on the court, while being able to outscore their opponents when he was on the bench.

Btw, he is also above average in my metric.

v-zero wrote: Bias doesn't invalidate conclusions, it simply makes them incomplete.

Actually, having an incomplete conclusion is as good as having an invalid conclusion. You have yet to show that your conclusion and equation is worth anything, because you haven't showed any kind of causality here. Can you imagine that you have it backwards here? Can you prove that putting players into different roles and with different minutes would actually lead to more success?

v-zero wrote: I believe you have failed to understand this, you believe I have failed to frame things properly, at this point we should stop.

See, that is the difference, I know that you are wrong when you conclude that coaches have no clue in average, what gives them wins. As I said, it is convienent to believe you have superior knowledge, but as it seems right now, you don't even grasp the overall framework of the game and the decision making process.
I understand that you wanted to use boxscore entries in order to predict minute distribution, while assuming that this would give you a clue about the coaching decisions in terms of playing time. You wanted to know why certain players are getting more minutes than others. It is easy to understand what you wanted to do, really, what is not easy to understand is why you failed. And you are right now struggling to understand your mistake.

Btw, which boxscore based metric are using which is actually a better predictor than RAPM?

mystic wrote:...

v-zero wrote:I have no interest in continuing this discussion.

My best box score metric currently has a one-step forward RMSE of 11.57 for the score difference over a sample of 7230 games assuming we are able to perfectly predict minute allocation, which is a good assumption because we are trying to measure the value of the metric, not the quality of the algorithm one might use to predict minute allocation. I have yet to see RAPM used properly to do this sort of prediction, so I can't say with certainty that in building a properly smoothed RAPM metric you might find a better solution, but I am certain that using the RAPM values currently available you cannot do so.

v-zero wrote:I have no interest in continuing this discussion.

Well ...

v-zero wrote: My best box score metric currently has a one-step forward RMSE of 11.57 for the score difference over a sample of 7230 games assuming we are able to perfectly predict minute allocation, which is a good assumption because we are trying to measure the value of the metric, not the quality of the algorithm one might use to predict minute allocation.

Out of sample or retrodiction? Do you have the sample available for a download?

v-zero wrote: I have yet to see RAPM used properly to do this sort of prediction, so I can't say with certainty that in building a properly smoothed RAPM metric you might find a better solution, but I am certain that using the RAPM values currently available you cannot do so.

You haven't seen that yet, but you are certain it cannot do it? Sounds a bit weird, don't you think?

mystic wrote:

v-zero wrote:I have no interest in continuing this discussion.

Well ...

v-zero wrote: My best box score metric currently has a one-step forward RMSE of 11.57 for the score difference over a sample of 7230 games assuming we are able to perfectly predict minute allocation, which is a good assumption because we are trying to measure the value of the metric, not the quality of the algorithm one might use to predict minute allocation.

Out of sample or retrodiction? Do you have the sample available for a download?

v-zero wrote: I have yet to see RAPM used properly to do this sort of prediction, so I can't say with certainty that in building a properly smoothed RAPM metric you might find a better solution, but I am certain that using the RAPM values currently available you cannot do so.

You haven't seen that yet, but you are certain it cannot do it? Sounds a bit weird, don't you think?

Out of sample, prediction not retrodiction. I don't have it available for download and have no plans to make it available at this time, so I understand I haven't provided any proof and that you are welcome to not believe me. I haven't seen anybody build a proper framework for prediction using RAPM (that is a framework properly weighting recent results and measurements rather than using simple averages), and if they do and it is better then fine, but what data is currently available in RAPM form is not better, from everything I have come across. I actually don't hate RAPM, but some around here (you included) seem to use it as a panacea and a crutch. "It doesn't regress well onto RAPM" seems to be a common statement, but regressing one uncertain model onto another to prove a point is pretty f*****g dumb.

v-zero above: "I haven't seen anybody build a proper framework for prediction using RAPM (that is a framework properly weighting recent results and measurements rather than using simple averages)..."

Are you using /suggesting using results and measurements of less than a full season? Say monthly? I'd be interested in a bit more explanation if you are willing.

Crow wrote:v-zero above: "I haven't seen anybody build a proper framework for prediction using RAPM (that is a framework properly weighting recent results and measurements rather than using simple averages)..."

Are you using /suggesting using results and measurements of less than a full season? Say monthly? I'd be interested in a bit more explanation if you are willing.

I'm suggesting using some form of smoothing and as much data as you have. If you're solving a big Bayesian problem properly then constructing a decent pseudolikelihood function and then using MCMC to solve it would be ideal. In my case I use simple exponential smoothing on game-by-game (not aggregate) results and box score data, though my smoothing does make exceptions for rookies, and it also increases its update parameter when a player hasn't played for a long period (e.g. between seasons). If anybody is doing anything like this with RAPM then I would love to see it, genuinely.

v-zero wrote: I don't have it available for download and have no plans to make it available at this time, so I understand I haven't provided any proof and that you are welcome to not believe me.

Well, my own boxscore metric showed a similar performance when tested last season (for 610 games (all games until March 11) with a RMSE of 11.19), even though I did not take the availability of the players into account. Thus, I actually believe you, I just would have liked to test the same sample.

v-zero wrote:I actually don't hate RAPM, but some around here (you included) seem to use it as a panacea and a crutch. "It doesn't regress well onto RAPM" seems to be a common statement, but regressing one uncertain model onto another to prove a point is pretty f*****g dumb.

How can you say "you included"? Where did I actually made a statement even remotely close to "It doesn't regress well onto RAPM"? You are pretty quick with a judgement, while again having no real proof for anything. It is rather sad, isn't it?