Permaximum Ratings and Rankings (Updated Often)

[permaximum]

Isn't there already a rating system that loves rebounders and is "better than PER" in correlations at the team level?

If there is, I would like to compare it to PTR.

But just for you Mike, I give the weights of defensive rebound and assist. BTW people here spend a lot more time with this work than myself and I appreciate and praise all your efforts. I have no place to criticize anyone's work here.

Point: 1
Defensive Rebound: 0.507649115024634
Assist: 0.372197563457339

[v-zero]

Isn't there already a rating system that loves rebounders and is "better than PER" in correlations at the team level?

If there is, I would like to compare it to PTR.

But just for you Mike, I give the weights of defensive rebound and assist. BTW people here spend a lot more time with this work than myself and I appreciate and praise all your efforts. I have no place to criticize anyone's work here.

Point: 1
Defensive Rebound: 0.507649115024634
Assist: 0.372197563457339

Mike's response was, I assume, a nod in the direction of the generally discounted (around here anyway) Wins Produced.

Also you'll find that with the right analysis that assists are hugely more indicative of player ability than defensive rebounds.

[permaximum]

Isn't there already a rating system that loves rebounders and is "better than PER" in correlations at the team level?

If there is, I would like to compare it to PTR.

But just for you Mike, I give the weights of defensive rebound and assist. BTW people here spend a lot more time with this work than myself and I appreciate and praise all your efforts. I have no place to criticize anyone's work here.

Point: 1
Defensive Rebound: 0.507649115024634
Assist: 0.372197563457339

Mike's response was, I assume, a nod in the direction of the generally discounted (around here anyway) Wins Produced.

Also you'll find that with the right analysis that assists are hugely more indicative of player ability than defensive rebounds.

Wins Produced is fundemetally flawed imo because of adjusting for positions played. I would never ever use that rating for anything.

Although assists may be more indicative of player ability than defensive rebounds, I don't think league average assist weight is greater than league average defensive rebound weight. To correctly predict the weight of different types of assists, I would need the player positions of all players on the floor when the ball was passed and the shot was taken. You can't get that with PBP data afaik.

Additionally, I assume there are no perfect standards for crediting players with assists. From what I've seen there's a considerable human error in crediting assists... Simple pass to a close player has a strong potential to be an assist in the end.

Overall, average drb/average ast should be very close to 0.507/0.372. This is what "regression" is for. I know that assist weight is going to punish and reward some players who gets many assists but if we think there is one type of assist, it's weight is accurate.

Edit: Small note. Box score stats do not represent guards' ability to bring the ball and run plays for the team, so it's very logical that point guards are not at top (Still, CP3 and Wade got into top5 in previous seasons). Other ratings try to cover this up by weighting assists more than their worth but theoritical approaches, estimates bring their own error margins which could be seen by PER's accuracy.

This is why I would only use pure RAPM besides looking at box-score stats. Unfortunately RAPM is not pure now I think.

[v-zero]

The real problem with Wins Produced is the weightings, the positional adjustment isn't a bad idea as long as you can assume that all teams will play with two guards and three forwards (more or less), which they do. In fact a positional adjustment can make up for some of the deficiencies of the box score.

Raw assists and drbs might have a weights as you have found, but raw assists aren't meaningful really. To identify great assisters you will find that the USG*AST interaction is very valuable. Good guards in particular get a boost from this.

[permaximum]

The real problem with Wins Produced is the weightings, the positional adjustment isn't a bad idea as long as you can assume that all teams will play with two guards and three forwards (more or less), which they do. In fact a positional adjustment can make up for some of the deficiencies of the box score.

Raw assists and drbs might have a weights as you have found, but raw assists aren't meaningful really. To identify great assisters you will find that the USG*AST interaction is very valuable. Good guards in particular get a boost from this.

V-zero I don't want to use advanced stats especially if it involves player possession formula but I agree that good guards will get a boost from that interaction. Just I don't want to estimate anything... This is why more stats should be covered in box score as we all probably agree. I leave advanced stats area to experts (I really mean it. No sarcasm) like you here.

As for WP, it does that adjustment to rank guards at top too. Simple as that. How did the maker of WP come with that adjustment coefficient? Also some players play different positions in the same game and on the both ends of the floor. How's it going to estimate that?

All of these type of metrics try to reach that specific goal to be public-friendly. I think we should focus more on +/- stats than making estimations for things box-score does not cover. For example we can think more on how we can improve RAPM...Does the runner of it use "leave-one-out cross validation"? to find the lambda or is "ridge regression" the best method really or not? Things like that... OFC all of these are my opinion.

[EvanZ]

I use positions to debit players for "missed" rebounds.

[v-zero]

The positional adjustment in WP works on the basis of calculating each player's raw production, then averaging over each position and then taking this average from each player at that position, so it is at least sensible. I agree that it is impossible to fully account for the fact that player positions are somewhat fluid, but that won't introduce a systematic bias in the data assuming player positions are decently well defined.

As for concentrating on +/- type stats... well, we are at the limitation with RAPM of what we can do without introducing priors through box-score/pbp data, which is what is being done with xRAPM. That being the case I believe it is as important now as it ever was to create more predictive box-score/pbp based metrics.

[permaximum]

The positional adjustment in WP works on the basis of calculating each player's raw production, then averaging over each position and then taking this average from each player at that position, so it is at least sensible. I agree that it is impossible to fully account for the fact that player positions are somewhat fluid, but that won't introduce a systematic bias in the data assuming player positions are decently well defined.

As for concentrating on +/- type stats... well, we are at the limitation with RAPM of what we can do without introducing priors through box-score/pbp data, which is what is being done with xRAPM. That being the case I believe it is as important now as it ever was to create more predictive box-score/pbp based metrics.

That means all position values should equal eachother. Is it really true? This is an another problem along with player position fluidity. But I think if I implement that method for my rating, nobody will be able to beat Chris Paul.

Introducing priors sound interesting... Unfortunately I have zero information about it. Could you summarize the process for me? Especially I'm interested in chosen box-score/pbp stats. Thanks for this information.

[v-zero]

On priors I suggest you read this thread as it's where the 'x' part of xRAPM originates: viewtopic.php?f=2&t=8025

Regarding the problem of assuming all positions have the same value, the question you really need to ask is: Is the guard-guard-forward-forward-forward setup optimal? If it is not optimal then why does it continue to be universally used? If it is optimal then each position must be filled and hence each position is intrinsically equal in value, because what really matters is how good your guards are compared to the opposition guards, and how good your forwards are compared to the opposition forwards.

[Crow]

Pemaximum's defensive rebound and assist weights are within the range of what I find reasonable to at least look at and consider and seem close to what I'd consider "best" or "right".

[permaximum]

Pemaximum's defensive rebound and assist weights are within the range of what I find reasonable to at least look at and consider and seem close to what I'd consider "best" or "right".

It's nice to hear that

On priors I suggest you read this thread as it's where the 'x' part of xRAPM originates: viewtopic.php?f=2&t=8025

Regarding the problem of assuming all positions have the same value, the question you really need to ask is: Is the guard-guard-forward-forward-forward setup optimal? If it is not optimal then why does it continue to be universally used? If it is optimal then each position must be filled and hence each position is intrinsically equal in value, because what really matters is how good your guards are compared to the opposition guards, and how good your forwards are compared to the opposition forwards.

I'm not sure if it's the optimal setup but as you pointed out it's universally used. A forward's effect could be greater than an additional guard as long as there's a guard on the floor. Too much estimation and you probably figured out by now that I don't like it . Add constant player position changes in games into the mix for lots of different reasons such as mismatchs, different plays, defensive-offensive adjustment etc. and suddenly you have a big problem imo. Although this position adjustment sounds sensible it brings serious drawbacks.

Thanks for the link btw. I tried to read most of the posts and although I found his tone harsh, I completely agree with a poster named mystic on the discussion. Those chosen variables, coefficients and regression are all out of place. I see why J.E. (I assume the runner of RAPM) tried to include everything from box score but there should be something wrong with that regression. Prediction is probably more accurate for wrong reasons. In the end it's going to predict worse the next seasons. I tihnk I won't use this version of RAPM.

e.g: Defensive weight of FGM is -0.3614711184. OFC it was going to be negative since making a field goal rewards the opponent with a new possession. But should a player be penalized on the defensive end for making a field goal? According to the better prediction accuracy, yes. To me? No. This is not an estimate but a simple fact.

As a side note, I don't like seperating the weights of all box-score stats for offense and defense. It's simply not valid.

[Mike G]

... team rankings. MOV adjusted by SOS, HCA, B2B and Pace. HCA is different for each team. Pace adjustment made for 100 possessions.

Aren't there some "estimates" in these adjustments? Are estimates OK for teams and not for players?
Team boxscore stats are still boxscore stats. What if you multiply player rates by (TmPts/OppPts), for example? I guarantee your teams' total rates would then be ranked closer to their rankings by 'adjusted' MOV.

[v-zero]

Aren't there some "estimates" in these adjustments? Are estimates OK for teams and not for players?
Team boxscore stats are still boxscore stats. What if you multiply player points by (TmPts/OppPts), for example? I guarantee your teams' total rates would then be ranked closer to their rankings by 'adjusted' MOV.

More or less exactly my feeling on this issue. All you actually do by avoiding advancing your analysis is making your estimates of player strengths that much fuzzier because you've got all your box score numbers in their own little boxes and won't let them out to play together... Everything here is an estimate, from assists numbers to RAPM values.

[permaximum]

... team rankings. MOV adjusted by SOS, HCA, B2B and Pace. HCA is different for each team. Pace adjustment made for 100 possessions.

Aren't there some "estimates" in these adjustments? Are estimates OK for teams and not for players?
Team boxscore stats are still boxscore stats. What if you multiply player rates by (TmPts/OppPts), for example? I guarantee your teams' total rates would then be ranked closer to their rankings by 'adjusted' MOV.

They're not really estimates. It's like coming up with the weights of box-score stats. I used chosen data to find HCA and B2B effect. I mean it's empirical. As for pace, I agree that it's a theoritical estimate but I believe it's a very accurate estimate because it's at team level and it takes opponent stats into account.

e.g: Rebounding percentage at player-level. It assumes each minute there is the same amount of available rebounds. If you use PBP data, this is not a problem. But if you use box-score data adding advanced stats into mix brings more error margin with each stat.

Agreed that teams' total rates would be ranked closer to MOV if I multiply player rates by TmPts/OppPts. But I'm not interested in finding team strengths with PTR. Again, it was just a test to compare PTR and PER's accuracy...

Do you honestly think each player is equally worth at generating TmPts and OppPts? That would be another wrong way to accomplish better accuracy...

[v-zero]

Weights for box score stats are estimates. Not all assists, shot attempts, blocks, steals, fouls, etc etc are made equal, and in the case of assisting there are certain players whose assists are of high value, and then there are those whose assists are really just passes that happen to be the last pass before a shot.