APBRmetrics

I thought I'd address some of the comments regarding ASPM in this thread.

First of all: I am in the process of revising ASPM, improving it in several ways--both in accuracy and in robustness, particularly at the extremes--where Iverson resided.

The revised spreadsheet including historical data is located at https://drive.google.com/file/d/0Bx1NfC ... sp=sharing

This should handle Iverson better, particularly since I ditched the USG^2 term. It was not highly significant and did not improve the fit much at all, within the new model specification.

That said, this new model does lower Iverson's numbers significantly. He now never had a season much above 200th in the last 40 years--which is still good, but not in the 60s like before.

His Hall Rank drops from 19th to 36th in the era.

---

schtevie is correct--for extreme outliers, of which Iverson was certainly one, box score stats in any flavor are going to be questionable. If there is any nonlinearity modeled for the usage/efficiency tradeoff, then that will add a lot of potential error to Iverson's estimate--this was the case for my older version of ASPM, and was legitimately a reason to question the validity of his ASPM estimate. The newer model should be more robust (and conservative) at the extremes.

Therefore, RAPM (NOT xRAPM) should be the tool to use here, and the best version available for that is the 14 year RAPM average. I made a Viz of it here: http://public.tableausoftware.com/profi ... 14YearRAPM Sure, it won't capture the peak, but it gives us a good idea--and indicates that Iverson may not have been the greatest.

I would dispute the contention that 14 year dataset is not enough for RAPM--on the contrary, I believe that, when properly aging adjusted (as Jerry has done) it's quite accurate. There are some flaws, still, due to players that have unusual aging/progression (Hi, KD!), but on the whole it's really solid.

---

As for the questions about ASPM's coefficients and flaws--MPG is a highly significant term, and effectively a proxy for "what coaches value." I have debated at times over whether I should not include the term, but ultimately decided to go ahead and leave it. In the older regression, the coefficient was 0.08, so a 40 MPG player was getting a boost of +3.2 to his per minute rating. The newer regression has an even higher coefficient. It's certainly a significant part of the rating system, and also open to debate.

Let me know if you all have any other questions on ASPM; I would be happy to answer them.

---

Here is Iverson's career, with my new ASPM, PER, and Win Shares. Note he is now a below average defender, which jives with the RAPM numbers. His offense has also declined, but is still considerably above the RAPM estimate. Which is more right, RAPM or ASPM? That's hard to know. Long term RAPM should certainly not be dismissed summarily.

The context is pretty crucial in judging Iverson's career. He played with no spacing in an era already less efficient than now. The Sixers were built to win on defense but everyone needs passable offense to win and Iverson's contributions were important to their survival on that end. Iverson created good shots at the rim and free throw line to counteract the bad midrange ones. A case can be made his midrange shots would have likely been "replaced" by just as bad midrange shots by other players if he hadn't forced them, especially considering playing on a team and in an era were his teammates still thought 2pt jump-shots were good too. It's not as if Iverson was chucking midrange shots when the alternative was a Steve Nash and Amare pick and roll.

That doesn't mean he was a top 5 player or deserved to get near MVP and a case can be made he doesn't deserve HOF. I would put him in the same type of tier as players like Deron Williams and Elton Brand.

Thanks for the update and data.

So for his career on aspm, a modest rookie season, then 5 of 8 seasons over plus 3 and 3 just under then 5 falling off pretty fast and then really far.

If RAPM is the "right tool" why isn't it in the chart (I know the 1 yrs are different than the 14 yr but still seems worth listing when you go out of way to list so many others). How does aspm and rapm look side by side? I'd like to see and will but ideally it could be here already.

Iverson, again?

It could have been bad, boring or just not useful.

It hasn't been, imo.

DSMok1 wrote: This should handle Iverson better, particularly since I ditched the USG^2 term. It was not highly significant and did not improve the fit much at all, within the new model specification.

That said, this new model does lower Iverson's numbers significantly. He now never had a season much above 200th in the last 40 years--which is still good, but not in the 60s like before.

His Hall Rank drops from 19th to 36th in the era.

I would like to know why did you remove USG^2? Has this imporoved ASPM's explanation power or prediction power? Or both?

Dr Positivity wrote:That doesn't mean he was a top 5 player or deserved to get near MVP and a case can be made he doesn't deserve HOF. I would put him in the same type of tier as players like Deron Williams and Elton Brand.

I'm sorry but this career ensures first-ballot HOF without looking at further statistical evidence.

1st pick in 1996 draft
1996 - Rookie of the year
1996 - Rookie Game MVP
11-time All-Star
2-time All-Star game MVP
2001 NBA MVP
4-time scoring champion
3-time steals leader
7-time minutes leader
7-time ALL-NBA selections (3-time ALL-NBA First)
3-time ALL-Defensive Team nomination

Crow wrote: If RAPM is the "right tool" why isn't it in the chart (I know the 1 yrs are different than the 14 yr but still seems worth listing when you go out of way to list so many others). How does aspm and rapm look side by side? I'd like to see and will but ideally it could be here already.

1 or even 2 year RAPM is worthless, xRAPM has box score biases like ASPM does (though to a lesser extent), and the best RAPM is the 14 year average--which has already been posted and discussed. The main reason I didn't list xRAPM, though, is it's not in my database and the others were. Please pardon the omission! In general, I feel xRAPM/RPM is probably the best single stat out there.

permaximum wrote:

DSMok1 wrote: This should handle Iverson better, particularly since I ditched the USG^2 term. It was not highly significant and did not improve the fit much at all, within the new model specification.

That said, this new model does lower Iverson's numbers significantly. He now never had a season much above 200th in the last 40 years--which is still good, but not in the 60s like before.

His Hall Rank drops from 19th to 36th in the era.

I would like to know why did you remove USG^2? Has this imporoved ASPM's explanation power or prediction power? Or both?

Both. Basically, once I made other revisions (and maybe even before to some extent) it was a totally unnecessary term--only bumped the R^2 by a small fraction of a percent, at the expense of wackiness at the extremes (like AI).

DSMok1 wrote:Both. Basically, once I made other revisions (and maybe even before to some extent) it was a totally unnecessary term--only bumped the R^2 by a small fraction of a percent, at the expense of wackiness at the extremes (like AI).

So you removed it despite the fact that it lessened R-squared because it sounded more right to you since there were players like Iverson who you thought should have not been up there?

If it's what it is, I will drop using ASPM

permaximum wrote: So you removed it despite the fact that it lessened R-squared because it sounded more right to you since there were players like Iverson who you thought should have not been up there?

If it's what it is, I will drop using ASPM

I removed it because it was not statistically significant, pure and simple. It was not only not statistically significant, it hardly effected the R^2 of the regression (which isn't always the case), so there was no way to defend having it in at all. I don't care who is ranked high or low, I only care about the quality of the model.

By the way, any additional variable that's not completely random will by definition increase the R^2, no matter if it's significant or not. If you use those, though, you are overspecifying the model and lose out of sample accuracy. The fewer degrees of freedom, the better, and that's particularly true of nonlinear terms. Be very wary of them. I was not cautious enough with that when I originally developed ASPM a few years back.

DSMok1 wrote:

permaximum wrote: So you removed it despite the fact that it lessened R-squared because it sounded more right to you since there were players like Iverson who you thought should have not been up there?

If it's what it is, I will drop using ASPM

By the way, any additional variable that's not completely random will by definition increase the R^2, no matter if it's significant or not. If you use those, though, you are overspecifying the model and lose out of sample accuracy. The fewer degrees of freedom, the better, and that's particularly true of nonlinear terms. Be very wary of them. I was not cautious enough with that when I originally developed ASPM a few years back.

What do you think J.E is doing with his xRAPM then?

I believe xRAPM is the best overall stat available right now; J.E. knows what he's doing statistically, and cross-validates everything for maximum OOS accuracy. He's not overspecifying anything. That said, Iverson is an outlier, and the box score weights for xRAPM or ASPM or any box score are going to be most accurate toward average stats, away from outliers. RAPM, on the other hand, will have the same accuracy on outliers and players with normal box score stats. Since this thread is about Iverson, I think pure RAPM is probably the best stat to avoid those potential issues.

DSMok1 wrote:I believe xRAPM is the best overall stat available right now; J.E. knows what he's doing statistically, and cross-validates everything for maximum OOS accuracy. He's not overspecifying anything. That said, Iverson is an outlier, and the box score weights for xRAPM or ASPM or any box score are going to be most accurate toward average stats, away from outliers. RAPM, on the other hand, will have the same accuracy on outliers and players with normal box score stats. Since this thread is about Iverson, I think pure RAPM is probably the best stat to avoid those potential issues.

height 1 0.8931422196
exp -0.065287672 0.0537292936
GS -0.0185391812 0.1017509227
MP -0.04221383 0.1703430321
FGM 1.0019454696 -0.3614711184
FGA -0.7713214889 -0.0802363173
FG% 0.2848825566 -0.0243612852
3FG 0.2288358653 -0.0252052757
3FGA 0.0349326238 -0.0660744428
3-% 0.3220324436 0.1719594885
FT 0.3366949904 0.1357251166
FTA 0.0736131376 0.0098560925
FT% -0.0252195806 -0.8179984551
OREB 0.296998827 -0.0366778245
DREB -0.1021371556 0.4586301605
ASS 0.4790862015 -0.0081967213
ST 0.0581352101 0.3665636712
BLO 0.0205847853 0.2253025491
TO -0.4038566074 -0.0454896575
FOULS -0.2794180757 0.016693846

If you were overspecifying with USG^2, I can't really say J.E. is not overspecifying with all these unless I miss something here.

permaximum wrote: If you were overspecifying with USG^2, I can't really say J.E. is not overspecifying with all these unless I miss something here.

J.E. uses K-fold cross validation to develop coefficients, rather than a standard in-sample regression model. So he won't have overspecifying problems. J.E. is really diligent with the statistical/theoretical foundation for his models--he really knows what he's doing.

DSMok1 wrote:I thought I'd address some of the comments regarding ASPM in this thread.

First of all: I am in the process of revising ASPM, improving it in several ways--both in accuracy and in robustness, particularly at the extremes--where Iverson resided.

The revised spreadsheet including historical data is located at https://drive.google.com/file/d/0Bx1NfC ... sp=sharing
.

It wont let me download that spreadsheet


same here on different tries

I know that he's diligent with statistics. I think he uses 3 or 5 folds. I thought you too use k-fold or loocv to get your coefficient values so removal of USG^2 didn't make sense to me before. Thanks for the info.

Still, cross validation cannot solve overspecifying issues. I mean look at those values for offense and defense.

I suggest J.E. to use leave-one-out cross validation on 1997(the year NBA reverted back to old 3pt line)-2014 box-score data playoffs included (all regular season and playoff games). He'll find out that new values are significantly different... after some computation time if he has a regular computer.

Overall ASPM estimate higher than overall RAPM estimate in all years except last 3.