Appr. 5.x year reg. adj. +/- (J.E., 2010)

[bbstats]

Jerry,

What's the timeframe for this "better" version with the priors? Just 06-present?

EDIT: Derrick Rose is +0.5 ?? Ouch.

[J.E.]

What's the timeframe for this "better" version with the priors? Just 06-present?

"Just"? We don't have any more matchupdata. Until I, or someone else, converts older PBP files that's all there is.

Why the ""better""? It's definitely better

[DSMok1]

Wonderful!

Did you settle on a specific lambda to use for each year-to-year translation?

I'd still like to see a couple of things: aging curve preprocessing (that may be hard), and perhaps using some sort of statistics (maybe just MPG) as a prior.

Can you quantify how much better this latest version is for OOS in this year?

EDIT: would the ratings be improved if you adjusted w/ aging curve between seasons? That might help get you better priors, particularly for the older players (Nash, KG). In fact, I strongly recommend trying this. It wouldn't be hard at all to do! This curve would be a good starting point: http://sonicscentral.com/apbrmetrics/vi ... .php?t=225

[Crow]

Amir Johnson continues to put up very good numbers on RAPM and a number of other metrics.

Not sure what caused the insiders with Detroit (and Milwaukee) to deal him. Maybe his low usage was considered a problem, a fit issue. Or maybe they didn't like his attitude or body language. He was headed for a raise but it still appears he is a good or more than good value using RAPM and / other stuff.

Kinda surprised somebody else didn't outbid Bryan Colangelo when Johnson became a free-agent. I've always thought he was a desirable young guy and a desirable component of a rotation.


Top 20 guys provide about 43% of the cumulative total positive RAPM impact. Top 50 out of 462 guys with almost 75% of the cumulative positive impact. Only 30% of guys are estimated above neutral impact (less than 5 per team if they were evenly distributed, which they probably aren't).

40% of the players in the league are estimated at -2 or worse. The bottom 50 would be as bad as the top 50 is good if they played as much, which they don't.

If you fill a rotation spot with a guy worse than -2 on RAPM, I'd think you'd want to look around at your options, even for someone with just a less bad net impact, not necessarily good.

The list includes in order of worsening estimated impact: Monta Ellis, Cousins, A Brooks, Dragic, Favors, Pietrus, J Green, G Davis, DeRozan, A Jefferson, Bargnani, Villanueva, Mo Williams, S Brown, Gooden, Warrick and Hickson. Only 5 of these guys haven't been moved or dropped yet. For most or all of Ellis, DeRozan, Davis, Bargnani and Hickson it is probably coming. In a couple cases their current teams might still believe in them or maybe they are asking too much for them.

Is the RAPM rating notably "wrong" for any of these guys? Possibly a few, more likely the young ones, but I'd guess that it is generally right and will generally stay that way for these guys. Changing the team context gets tried and maybe some leave this basement that way but other options would be to change the role, placement in a rotation and / or minutes significantly. Most of these guys aren't starter grade though about half of them do.

[bbstats]

Jerry,

I wasn't trying to be negative. I'm actually overly excited about these numbers!

By "just" I meant "am I describing it entirely?" I don't know if more years of PBP will improve the numbers, I was just wondering what the timeframe was (i.e. if it was 2-years, 3-years, or with all of your data).

By "better" I just meant that it is theoretically "better" than the current 3.x year...scientifically I am just remaining at least a little skeptical about any one method describing a player's causal value

In terms of out-of-sample prediction, yes I do agree it is better as you have explained!

I took a look at regular season adv. stats, Net +/- & 1 year of prior RAPM predicting this rating (500 MP or more)

Code: Select all

Regression Statistics		
Multiple R        0.887167081	
R Square          0.787065429	
Adjusted R Square 0.779518332	
Standard Error    1.207795337	
Observations      338
	
                CoefficientsP-value
Intercept       0           #N/A
TOV%            0.042074334 0.009341427
SUMteammateNET -0.032481061 0.006919675
ORtg            0.027312218 0.002305021
TeamSRS*Min%    0.179031782 0.0001058
DRtg           -0.050497366 1.26284E-07
MPG             0.055993617 1.08698E-08
2010rapm        0.586084237 3.48866E-29
2011ReNet       0.481836941 1.58951E-35

Edit: SumteammateNET is team(OnCourtMinusOffCourt per 100 from BBV * Min%) - player(OnCourtMinusOffCourt per 100 * Min%)

ReNet is basically player NET (OnCourtMinusOffCourt per 100) which I tuned to 3-year RAPM for minimizing prediction error based on possessions played.

Rookies get "-1" for 2010 RAPM, and traded players get "0" for teammates & SRS.

[J.E.]

Did you settle on a specific lambda to use for each year-to-year translation?

I settled on 5000. The differences between anything from 3000 to 6000 were tiny

I'd still like to see a couple of things: aging curve preprocessing (that may be hard)

I thought about doing this the same way I did things with coaches: If you have a 31 year old on the court for team A you add P_31_years to their lineup, and so on. You'd probably have to split for offense and defense because I think players start to regress earlier on offense.

One other thing that really needs to be tested is to simply do a multiyear analysis, but give lower weights to older seasons. Joe Sill did this and saw improvements

Can you quantify how much better this latest version is for OOS in this year?

Total out of sample error gets reduced by 0.2%. Doesn't seem like much but this is not an area where gains seem easy to come by.

EDIT: would the ratings be improved if you adjusted w/ aging curve between seasons? That might help get you better priors, particularly for the older players (Nash, KG). In fact, I strongly recommend trying this. It wouldn't be hard at all to do! This curve would be a good starting point: http://sonicscentral.com/apbrmetrics/vi ... .php?t=225

I'll do it when I have more time

The list includes in order of worsening estimated impact: Monta Ellis, Cousins, A Brooks, Dragic, Favors, Pietrus, J Green, G Davis, DeRozan, A Jefferson, Bargnani, Villanueva, Mo Williams, S Brown, Gooden, Warrick and Hickson. Only 5 of these guys haven't been moved or dropped yet. For most or all of Ellis, DeRozan, Davis, Bargnani and Hickson it is probably coming

Favors is not even 20 years old yet, I'd give him two more years at least. Similar thing with Cousins. Ellis might thrive if put in the right situation.
For A Jefferson, Bargnani, Charlie V, M Williams, Gooden, Warrick and Hickson I believe the ship has sailed. Although some could definitely benefit by being put in a different role, maybe on a different team.

I don't know if more years of PBP will improve the numbers

I'm not sure either. I think I can find that out by starting with 0 as prior in '07 (instead of '06) and see if that changes things in '11

[J.E.]

It just occured to me that I should probably use a lower lambda for rookies. High lambda for non rookies probably does alright because the prior was pretty good. The rookie prior is likely a bit worse and a lower lambda would allow for more distance from the (not so good) prior.

That's how it makes sense in my head. Whether this is an correct assumption will be told by out of sample error

[DSMok1]

It just occured to me that I should probably use a lower lambda for rookies. High lambda for non rookies probably does alright because the prior was pretty good. The rookie prior is likely a bit worse and a lower lambda would allow for more distance from the (not so good) prior.

That's how it makes sense in my head. Whether this is an correct assumption will be told by out of sample error

Sounds right to me.

[Crow]

The newest overall offensive and defensive RAPM ratings and the 3 component parts (adjusted turnovers, adjusted PPS and adjusted rebounding) can be compiled for players and then the sum of the parts can be computed and compared with the newest overall ratings.

(To convert adjusted defensive rebounding (which I believe in stated in terms of number of rebounds and not the point value of rebounds- correct me if I am wrong) to point value, I assumed they were worth .3 points. I assumed offensive rebounds were worth at least .7 points.)

I did the compilation for 10 of the low ranking guys I mentioned earlier in this thread, 5 bigs and 5 guards. Curiously the overall offensive and defensive RAPMs for the 5 bigs (Hickson, Davis, Jefferson, Gooden and Villanueva) estimated they had a worse overall impact than the sum of the component parts in every case on offense and defense.

(I used multi-season for all of these. Now the overall RAPM is 5 year and 2 of the parts are 4 year and one is 3 year but I'll note that and put it aside as a modest difference.)

For the 5 guards (Ellis, Dragic, Brooks, Williams and Brown) the overall offensive rating estimated them as being worse than the sum of the parts in 3 cases while on defensive the overall estimate was worse than the sum of the parts in 4 cases.

Obviously this is very small sample. But it may be that the overall ratings could need regularized priors to produce scores than conform more closely with the scores from the sum of the regularized parts. Not that was necessarily the author's goal (to date), but it is an observation coming from the brief attempt to compare the overall and component databases. If that were a goal, then maybe your planned next step will move your datasets to closer agreement than right now. Potentially regularized priors could also be used at component or factor level, if you wanted to do so and push for presumably even greater accuracy at component level and even greater conformity between overall and component level data.

Ideally all this data could be compiled by player into a summary file with player ID and team indicator and minutes for simple review, to avoid checking multiple files and manual assembly. If that was done and the basketballvalue position estimates were imported and listed as well then it was be easier to do additional systematic study, including by position and team level and by other criteria.

These are the observations and options that come to mind at this point for your consideration. But of course, thanks again for all the work to produce the data and for making it available.

[J.E.]

(To convert adjusted defensive rebounding (which I believe in stated in terms of number of rebounds and not the point value of rebounds- correct me if I am wrong) to point value, I assumed they were worth .3 points. I assumed offensive rebounds were worth at least .7 points.)

This is wrong. In "adjusted world" +1 on defensive rebounding is the same as +1 on offensive rebounding
Just summing all the component parts is also wrong, you need weights.
Also, anything that has to do with Free Throw shooting is not in the components. Another reason why you can't expect components and overall ratings to have a high correlation.
Add the fact that this new version and all the component versions use different time frames...

Ideally all this data could be compiled by player into a summary file with player ID and team indicator and minutes for simple review

This is certainly coming at some point

[J.E.]

Year to year correlation with this new approach is around 0.89 (r). Not really surprising, since we penalize for moving away from last years rating

[Crow]

"In "adjusted world" +1 on defensive rebounding is the same as +1 on offensive rebounding."

Maybe, but that doesn't change what I found much.


"Just summing all the component parts is also wrong, you need weights."

Not if everything is / was truly stated in points per 100 or 200 possession scoreboard impact.


"Also, anything that has to do with Free Throw shooting is not in the components."

Are you saying that Adjusted PPS doesn't have free throws incorporated? I thought it rolled the shooting and scoring factors together into one component for simplicity. If not, then why call it Adjusted PPS instead of just "Adjusted Shooting"?

[J.E.]

Maybe, but that doesn't change what I found much.

Not "maybe". That's how it is

Not if everything is / was truly stated in points per 100 or 200 possession scoreboard impact.

Well.. One is stated in impact per (100) shots, the next is impact per "rebounding opportunities", and the last is in impact per possession ..

Are you saying that Adjusted PPS doesn't have free throws incorporated?

Yes. I said so in the original thread, but I should probably also put it on the site to avoid confusion

[Crow]

"Not if everything is / was truly stated in points per 100 or 200 possession scoreboard impact."

I meant and should have said more directly that I personally preferred a simple common form of impact input, i.e. per 100 team possessions or 200 game possessions for all forms of Adjusted +/- for easy and proper comparison, summation or other action.

I knew about the per rebounding opportunities varied from that basis and roughly converted it before in other work but I forgot this time. I may not have noticed that PPS is by 100 shots instead of possessions. It makes things more complicated but it is possible to roughly work around it.

But put aside my improper summing of components above, and the general discussion about the desirability of the overall offensive and defensive RAPM matching up with the sum of the components properly summed and easily summed still stands as does the talk about the possible assistance of regressed priors for each.

[EvanZ]

Jerry, I can't remember if this is on your list, but if not...adjusted assists? That could be really interesting.