APBRmetrics

EvanZ wrote:Can you explain how the ratings are "empirically determined"? Also, what does a rating of 30 mean in terms of wins/losses?

As I said before, I chose 1997/98 - 2011/12 game box-score data for the regression. I don't want to get into the details of how I ran the regression here. 30 is simply a rating like PER. It does not directly relate to team wins and losses but should give an idea. It's main purpose is to show how good a given player's box score stats are.

For all other things I would use RAPM personally instead of going with a metric which uses assisted FG estimation, rebounding percentage estimation, opponent efficiency estimation etc. If someone has a metric like that I would like to test it's accuracy.

permaximum wrote:
As I said before, I chose 1997/98 - 2011/12 game box-score data for the regression. I don't want to get into the details of how I ran the regression here.

Why should anyone be interested in this if the details are a black box?

It does not directly relate to team wins and losses but should give an idea. It's main purpose is to show how good a given player's box score stats are.

Without the details, I just don't see how someone can place this statement in proper context. Are the ratings supposed to be predictive of future wins or future box score stats or something else?

EvanZ wrote:

permaximum wrote:
As I said before, I chose 1997/98 - 2011/12 game box-score data for the regression. I don't want to get into the details of how I ran the regression here.

Why should anyone be interested in this if the details are a black box?

It does not directly relate to team wins and losses but should give an idea. It's main purpose is to show how good a given player's box score stats are.

Without the details, I just don't see how someone can place this statement in proper context. Are the ratings supposed to be predictive of future wins or future box score stats or something else?

It has a chance to take attention of someone because it's a more accurate box-score rating than PER. I don't have any commercial purpose or something like that with this rating. I just wanted to make an empirical box-score rating since there is not any.

PTR explains the past performance. They don't try to predict anything. In the same time I don't think any rating can be good at predicting the future. What's RAPM's prediction accuracy? 15% or less? IMO prediction area should belong to simulations. AFAIK Accuscore simulates games but their simulations are not accurate enough.

However, if you say there are ratings or something like that 55%+ accurate at predicting the outcome of games, I would gladly pay for it.

permaximum wrote:
It has a chance to take attention of someone because it's a more accurate box-score rating than PER. I don't have any commercial purpose or something like that with this rating. I just wanted to make an empirical box-score rating since there is not any.

More "accurate" in what sense exactly? Honestly, I'm not trying to give you a hard time, but I just can't put this into the context of other rating systems unless I understand it better.

Does anyone else get it? Maybe I'm just missing something really obvious.

Check the previous posts and OP...

How can I check PTR and PER accuracy? OFC team MOV and position. So, team sum of player PTRs were found, they were listed and compared to MOV positions. PTR was more accurate at placing teams closer to their real MOV positions. In summary, PER misplaced all teams by 114 positions totally. Or we can call it" PER placed teams 3.8 positions differently than their actual MOV position". PTR misplaced teams by 106. That means PTR is 7.55% more accurate than PER atm. I will update the ratings shortly.

permaximum wrote:Check the previous posts and OP...

How can I check PTR and PER accuracy? OFC team MOV and position. So, team sum of player PTRs were found, they were listed and compared to MOV positions. PTR was more accurate at placing teams closer to their real MOV positions. In summary, PER misplaced all teams by 114 positions totally. Or we can call it" PER placed teams 3.8 positions differently than their actual MOV position". PTR misplaced teams by 106. That means PTR is 7.55% more accurate than PER atm. I will update the ratings shortly.

So this was generated by a regression of player stats on team MOV?

For what it's worth, I agree with Evan--if you would like additional feedback, please provide technical details. Both Evan and myself generate player rating systems, which are completely transparent and both significantly better than PER.

My complete methodology (regressing advanced player stats on the player's 12-year-composite RAPM) may be found at http://godismyjudgeok.com/DStats/aspm-and-vorp/ .

Evan's work (tracking all events during a game via PbP and awarding credit/blame directly based on those) may be found at http://www.d3coder.com/thecity/ezpm-player-ratings/ .

I hadn't seen this til now and don't want to beat a dead horse, but what's the scale here, and what does a value of 30 "mean"?

There are some substantial face validity issues with that top players list...

Beyond formulation, scale is the crux of the matter here, as talkingpractice has pointed out. Without some sort of scale how can we say how much better players are than each other when we don't know whether being ten better than somebody is equivalent to one better in points-speak, or one thousand better? It's a problem I also have with PER (which is the clear inspiration for this) so you aren't being singled out for mean treatment.

As for how you should measure the value of a stat? Meaningful stats will be predictive (and not necessarily retrodictive), which earlier you threw out as a fool's errand... I suggest you reconsider that stance, because it's very easy to build a box score metric to explain wins at the team level, but it's much harder to build one that does a good job predicting wins when combining player ratings. The kicker is simply that until you know how well your model makes predictions you can really say nothing about how good it is at finding the true strength of individual players.

v-zero wrote:Meaningful stats will be predictive (and not necessarily retrodictive), which earlier you threw out as a fool's errand... I suggest you reconsider that stance, because it's very easy to build a box score metric to explain wins at the team level, but it's much harder to build one that does a good job predicting wins when combining player ratings. The kicker is simply that until you know how well your model makes predictions you can really say nothing about how good it is at finding the true strength of individual players.

Hear, hear to this. It's actually very easy to prove that a metric is predictive... use a set of player ratings, turn it into a more developed model of some sort, and go out there and win 54% of your bets. I'm shamefully oversimplifying this to make a point here. But if you do that with PER (or likely with the stat presented in this thread), then you'd be just flipping coins.

First and foremost, it's a very simple metric if I wasn't clear enough. I just gave weights to box score stats based on a regression. It's that simple. If you have the same data you can build it in a short time. So I don't want to call it a model as it would be shameful.

I say it again. Only simulations will be accurate enough for the prediction of complicated sports. If it's easy to accomplish 55%+ accuracy at predicting the final score margin (I think it's called spread or something like that. English is not my native language) of NBA games, I just ask you where's that metric? There are lots of money for us to earn..

You can find approx PER weights here. I wasn't the author of article btw. http://bleacherreport.com/articles/1131 ... l-the-mess

I thought weights were not good enough and making a PER-like pure box-score stat metric based on regression would be more accurate and here it is. In fact, I searched for that kind of metric (without the implementation of +/- in any format - SPM is one of them) and I couldn't find one. So here it is.

Finally, scaling. I strongly resisted to comment on that but I'll shoot it then. It's related to how I ran the regression. As you all know box score stats do not represent a lot of things especially at the defensive end and players don't get credited for their mistakes. So that box-score ratings should be theoritically higher than +/- stats and metrics. But the margin should stay reasonably accurate since opponents' ratings will be higher too.

Then you'll point out that player-combined rating of LAC per 100 poss is 95 while WAS' rating is 64 and 31 differental definetely doesn't represent MOV. Here you should just take the half of PTR and then you'll get your MOV.

Why is it "half"? My answer is regression again. I think you should understand why I don't give the details of regression here. All I can do is reitarating the fact that Point-Opponent Points=MOV. Some people will probably guess why it's "exactly half" now.

Edit: While I was reading EZPM details (thanks DSMonK), I thought using PBP data is simply beyond awesome (I'm too lazy for such a thing) and I really liked the model. Just I spotted some minor things. For example:

Blocks: In my model, the +0.7 coefficient for blocks has real meaning, since it gives the full value of the opponent's missed field goal to the player getting the block. To me, this makes perfect sense, since the guy getting his shot blocked missed the field goal and loses 0.7 pts. Shouldn't the blocker be credited with exactly the same amount?

Simply no. This is why theoritical approach is not good enough. If your FG miss is -0.7, block should be something like 0.68. Because if Player A wouldn't block that FG and Player B missed that FG, it would be better for Player A's team. Check offensive rebound percentage of blocked shots vs unblocked field goal misses and out of bounds situation. Especially since EZPM uses PBP data, it should seperate blocked shots from field goal misses and non-steal TOs from steal-TOs (here comes fastbreak possibility. Almost 9% worse than normal turnover).

permaximum wrote:

Blocks: In my model, the +0.7 coefficient for blocks has real meaning, since it gives the full value of the opponent's missed field goal to the player getting the block. To me, this makes perfect sense, since the guy getting his shot blocked missed the field goal and loses 0.7 pts. Shouldn't the blocker be credited with exactly the same amount?

Simply no. This is why theoritical approach is not good enough. If your FG miss is -0.7, block should be something like 0.68. Because if Player A wouldn't block that FG and Player B missed that FG, it would be better for Player A's team. Check offensive rebound percentage of blocked shots vs unblocked field goal misses and out of bounds situation. Especially since EZPM uses PBP data, it should seperate blocked shots from field goal misses and non-steal TOs from steal-TOs (here comes fastbreak possibility. Almost 9% worse than normal turnover).

Don't worry, this is coming.

permaximum wrote: Simply no. This is why theoritical approach is not good enough. If your FG miss is -0.7, block should be something like 0.68.

This isn't really that large of a criticism...

I wasn't criticizing anything or other parts of the model. It was just a minor example since I remember him, he seperated blocked shots and non-steal tos before when I asked for them here. This time I'll really update the ratings shortly

I hope people who asked for the explanation of PTR satisfied now.

Updated ratings are below. Kobe's rising steadily.

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