APBRmetrics

Hmm....Lebron, Lebron, Lebron...

I don't think that ridge regularization can precisely take into account altered distribution of possession usage (i.e. Cleveland Lebron vs. Miami Lebron). Can anybody think of theory to resolve this? This most certainly decreases the predictive power of RAPM on lineup data. Even if we eyeballed Lebron's usage patterns, we could probably predict the decrease in his overall RAPM this year. (Although his efficiency has decreased this season for some reason as well...)


What about using the regularized algorithm on "Player Possessions Used per team 100" with "Possessions Used" equaling FGA+.44*FTA+TO...or something along those lines.

My original idea that I offhandedly mentioned to JE was to create a regularized "team-possessions used per 48 minutes," but this would obviously be causally related to Tempo in addition to Usage.

Anywho, just musing.

EvanZ wrote:Interesting. That had a huge effect on two rookies (Ekpe Udoh and Omer Asik) that were near the top previously.

The changes for these two players suggest further tweaking, I think. Both of DSMok1's suggestions should help here: Using MPG instead of # of possessions and, maybe, using team point differential. Udoh didn't play because he was injured, not because he was too worse of a player to crack the rotation. I only need to figure out a simple way to know if a player didn't play because of injury, or because the coach decided not to play him. It's not so smart to use # of possessions anyway because we're unfairly punishing player's from slower teams.
Asik didn't play much because he's on a very good team. He probably would have played way more minutes on Minnesota. Using point differential might help here, but that might not be perfect either. Consider the PG and SF positions in Miami. The SF backup will have a hard time getting minutes but the equally skilled PG backup might not. So maybe we need to consider the rating of players playing the same position on the team to get a better prior. I have no idea if that could work though, seems like the algorithm might run in circles.

Do you plan to do the Adjusted +/- Factor for Turnovers?

Planned, yes. Soon? no

Is adding a Baynesian prior (or several) a near term work goal or a long-term option?

Whatever gives best out of sample predictions is the long-term work goal

Any interest in adding any other game context parameter beyond home court (e.g. rest & altitude)?

Not altitude. Rest maybe, but I feel there are more important things to research before that

RAMP

Regularized adjusted minus/plus? Just kidding.

I did get slightly improved next season predictions using this method (using my data, I couldn't calculate SPM for random games to do in-season cross-validation), but the differences were not very large.

What did you predict? Each games point differential? Wins? Points scored for a 5on5 unit? Good to know that you got better results.

Regularized adjusted minus/plus?

J.E. - you're well on your way to the Grand Unified Theory of basketball! Congrats.

Out of pure curiousity I wanted to compare non regularized Adj +/- to regularized Adj +/- and found, for the 4 year ratings, an correlation coefficient of r=0.91. I sure didn't expect it to be that high. Players with less than 2000 possessions were omitted from the non regularized Adj +/- ratings and from the computation of the correl. coefficient

Non regularized ratings are here http://stats-for-the-nba.appspot.com/non_ridge

Omer Asik, 2nd best defender in basketball. I'll definitely run this with coaches sometime


X axis is non regularized, Y is regularized

You get a cookie if you can guess who is in the upper right corner without looking at the numbers

There's my man, Ekpe Udoh!

I assume it is LeBron?

Totally voted for Asik on the IBBA's.

Estimated +/- impact varies on absolute value difference between RAPM and otherwise comparable non-ridge APM by 3 points or more in about 14% of cases.
By 2 or more in about 28% of cases. 1 or more in a bit over 50% of cases. So almost 50% of the values are within one point. The number of variances on the offensive and defensive splits very slightly from the overall numbers depending on the cutoff.

Crow wrote: I assume it is LeBron?

What kind of cookies do you prefer? Chocolate chip? (I'm serious)

Another idea for a prior: Why not use RAPM rankings from 08-10 as a prior for 2011 RAPM. Would that make sense? I like the idea right now but I haven't thought much about it, yet

RAPM rankings from 08-10 as a prior for 2011 RAPM sounds at least worth trying to me.

As would using a prior that was based partly on previous multi-season RAPM and partly on a statistical model for previous multi-seasons.

Both metrics for multiple seasons as a prior might get closer to "overall prior perfomance / "true skill" average estimate" and help steer the 1 year RAPM better than with no prior guidance at all or just minutes or just prior RAPM.

And using the same type of prior to find a 2011 RAPM and then blend that with a statistical model to produce a hybrid (in the same general vein as Rosenbaum's "overall +/-" blend back near the beginning of the public Adjusted +/- research) would seem to be worth trying to me. If a RAPM approach is going to achieve a much higher R2 I think it probably will have to be a RAPM / statistical model hybrid with priors and perhaps other adjustments.

Now with single year RAPM going back to 05-06. Everything comes with playoffs except 06

http://stats-for-the-nba.appspot.com/ranking10
http://stats-for-the-nba.appspot.com/ranking09
http://stats-for-the-nba.appspot.com/ranking08
http://stats-for-the-nba.appspot.com/ranking07
http://stats-for-the-nba.appspot.com/ranking06

Comments:
06: From eye balling it, it seems Sacramento had a nice run after Artest (#1) came to town, Wade (#2) won finals MVP. Sheed (#3) and Ginobil i(#5) played for the SRS leaders of their respective conferences.
07: #1, #2 and #3 meet in the finals, Garnett(#5) looks good even though he plays for a team with a -3 SRS, the TWolves. Baron Davis(#4) with his best year leading Golden State over #1 seed Dallas
08: Exact same top 4 as in 2011. Can they do it again in 2012?
09: Lakers get an extremely good year from Odom. Rashard Lewis looks good. Unfortunately it was the year he used steroids

Players you see at the top almost every year: Ginobili, Duncan, Nash, Carter, LeBron, Dirk, Kobe

Thanks, Jerry. Glad you put these up.

Beautiful stuff.

Very cool!

Really glad to see the '05-06 data, and encouraged it looks much better than the '05-06 APM data over on RealGM. The Pistons tended to both dominate the list, but with poor Rip Hamilton way at the bottom. I'd hoped RAPM could improve on that, and it definitely did. Still ain't sold enough to make 1 year RAPM my end all be all, but it's a step in the right direction.

Man, I remember that was a year I started seriously looking at what could be done with +/- and became discouraged because the results seemed so crazy. Didn't realize how much of an outlier the year was at the time.