APBRmetrics

Jerry, this is really cool!

But, amidst all this noise, I see that the PBP for 2012 is updated? Am I reading that right?
Will we finally get to see last year's end-of-the-year RAPM???


Edit: Also, I'm with D(aniel)sMok1 in that the 8-year RAPM is going to provide more significant results. Smoother data, at the least...not sure how to prove that that makes it more significant

bbstats wrote:But, amidst all this noise, I see that the PBP for 2012 is updated? Am I reading that right?

bbv's matchupfiles is what I need, and they weren't updated. So, no

not sure how to prove that that makes it more significant

Out-of-sample test should prove it

One more thing: A quick and dirty test for "what set of priors is the best" should be this:

For each set of priors, compute RAPM player values using those priors. For each player add up the distance between prior and final value. (adding up all distance values into a sum, probably square them first)

The set of priors with the smaller sum of distances should be the best set of priors. Perfect priors should require RAPM to create no distance between prior and final player value (and if RAPM creates a difference, it would be noise)

So, in a way, the main goal when designing a (BoxScore) player metric might be to create player values which, after using them for informing RAPM, underwent the least change due to the RAPM-algorithm.


This is probably equivalent to creating a metric that predicts best, but another way to quickly test different sets of priors

J. E., that would be equivalent to computing RAPM independently of the prior and finding which prior would have the highest correlation to it, wouldn't it?

DSMok1 wrote:J. E., that would be equivalent to computing RAPM independently of the prior and finding which prior would have the highest correlation to it, wouldn't it?

When you exclude low minute players and use mean squared error as measure as opposed to correlation I'd say it's equivalent, yes.
As I already said, I don't expect the results of this approach to differ much from your ASPM

J.E. wrote:

DSMok1 wrote:J. E., that would be equivalent to computing RAPM independently of the prior and finding which prior would have the highest correlation to it, wouldn't it?

When you exclude low minute players and use mean squared error as measure as opposed to correlation I'd say it's equivalent, yes.
As I already said, I don't expect the results of this approach to differ much from your ASPM

You're welcome to use/evaluate ASPM as a prior for RAPM any way you want to. I'd be glad to see what you come up with.

J.E. wrote:

bbstats wrote:But, amidst all this noise, I see that the PBP for 2012 is updated? Am I reading that right?

bbv's matchupfiles is what I need, and they weren't updated. So, no

Yeah, it's bumming out missing so many games.

Here are some historical player ratings. Weights for the BoxScore stats were optimized using 12 years of matchupdata. Player height, years of experience, and playoff totals (not per minute) were included. I'll post the weights soon, but can't right now
For some reason it's really in love with blocks/height, at least for defense. The top ranks are usually dominated by big men because small(er) players almost never get a good defensive rating. In the 90s, Jordan rates pretty well overall but never above #3 I believe; his offensive rating is usually in the top 5, but his defensive rating cannot keep up with the Robinsons/Olajuwons of the world.
The 90s look pretty much like this
Tier1: Robinson, Olajuwon, then later Shaq
Tier2: Kemp, Malone, Barkley, Stockton, Jordan, Drexler, Mutombo, Ewing, Pippen
(basically the Dream Team plus Mutombo(Congo), Olajuwon(Nigeria), Kemp and minus Laettner, Bird, Mullin, Magic. Magic is rated #2 in '91, Bird #10, Mullin ~#35)

Rodman rates well, but his offense is usually at around -2, which gets him ranked at around #30 overall in most years

The biggest (subjective) problem I have with the way it is now is that it rates McGee so well in '11 and '12. Hopefully this will become less of a problem once I work in goaltends and seperate blocks into two different categories (A: the team of the player that blocked the ball got possession, B: all other blocks)

Regression to the mean from season to season was assumed at a fixed level of 75%

The best statistical season, going by this, are Shaq, Robinson in '95, Robinson in '96 and Shaq in 2000. For Shaq, those are the two seasons he turned the ball over the least (going by TOV%)

http://stats-for-the-nba.appspot.com/PBP/1991_box.html
http://stats-for-the-nba.appspot.com/PBP/1992_box.html
http://stats-for-the-nba.appspot.com/PBP/1993_box.html
http://stats-for-the-nba.appspot.com/PBP/1994_box.html
http://stats-for-the-nba.appspot.com/PBP/1995_box.html
http://stats-for-the-nba.appspot.com/PBP/1996_box.html
http://stats-for-the-nba.appspot.com/PBP/1997_box.html
http://stats-for-the-nba.appspot.com/PBP/1998_box.html
http://stats-for-the-nba.appspot.com/PBP/1999_box.html
http://stats-for-the-nba.appspot.com/PBP/2000_box.html
http://stats-for-the-nba.appspot.com/PBP/2001_box.html
http://stats-for-the-nba.appspot.com/PBP/2002_box.html
http://stats-for-the-nba.appspot.com/PBP/2003_box.html
http://stats-for-the-nba.appspot.com/PBP/2004_box.html
http://stats-for-the-nba.appspot.com/PBP/2005_box.html
http://stats-for-the-nba.appspot.com/PBP/2006_box.html
http://stats-for-the-nba.appspot.com/PBP/2007_box.html
http://stats-for-the-nba.appspot.com/PBP/2008_box.html
http://stats-for-the-nba.appspot.com/PBP/2009_box.html
http://stats-for-the-nba.appspot.com/PBP/2010_box.html
http://stats-for-the-nba.appspot.com/PBP/2011_box.html
http://stats-for-the-nba.appspot.com/PBP/2012_box.html

In the 90s, Jordan rates pretty well overall but never above #3 I believe; his offensive rating is usually in the top 5, but his defensive rating cannot keep up with the Robinsons/Olajuwons of the world.

your 1990-91 player ratings lists michael adams higher (#3 overall) than michael jordan (#4). what is the reasoning behind this?...

denver that year went 20-62 with the league's worst defense and 6th worst offense, chicago that year went 61-21 with the league's best offense and 8th best defense...

your 1990-91 player ratings lists michael adams higher (#3 overall) than michael jordan (#4)

Looks like the numbers need to be scaled to opponent per game totals.

bchaikin wrote:In the 90s, Jordan rates pretty well overall but never above #3 I believe; his offensive rating is usually in the top 5, but his defensive rating cannot keep up with the Robinsons/Olajuwons of the world.

your 1990-91 player ratings lists michael adams higher (#3 overall) than michael jordan (#4). what is the reasoning behind this?...

denver that year went 20-62 with the league's worst defense and 6th worst offense, chicago that year went 61-21 with the league's best offense and 8th best defense...

First of all, since this is not RAPM there's a higher chance that we see players on bad teams rated well, and vice versa.
I obviously don't agree 100% with they way the players are ranked, and I don't think Adams was better than Jordan that year
The reason for Adams being rated that highly is this: the data from 2001 through 2012 suggests that a) having lots of assists and b) making and just attempting lots of 3s is a strong indicator of good (future) offensive performance. Jordan shot lots of 3s in '90 (245), but not in '91(93). Adams shot 564 3s in '91, plus he had a AST% of 39 and a TOV% of 13, which doesn't happen too often

It doesn't happen too often, but neither does a team often allow 131 ppg as Enver did in 1991.
You don't have to use any APM to prevent players on bad teams being overrated.
Just scale all players' offensive production to 100/OppPPG, or to LgPPG/OppPPG, for example.

In 1991, the Pistons allowed just 97 PPG. Joe Dumars averaged 20.4
Playing for the Nuggets, he might have averaged 20.4* 131/97 = 27.5

Those Nuggets had the 2nd-best TO/Poss rate in the league. Presumably because they didn't hold the ball very long.
Still, they were slightly worse than avg in TO per game.

The best statistical season, going by this, are Shaq, Robinson in '05, Robinson in '06 and Shaq in 2000.

You mean Robinson in '95 and '96?

You mean Robinson in '95 and '96?

Yes, thanks

Just scale all players' offensive production to 100/OppPPG, or to LgPPG/OppPPG, for example.

Right now there are no team stats involved. That makes it simple and applicable to other leagues, where people might not have access to pace-rating etc.
So far I'm scaling most stats to MP, but I could just scale by MP*(team pace). Seems more intuitive to me than scaling offensive production to opponents PPG, no? Your Dumars example works just as well when using team pace.
Seems like Denver wasn't aware of the fact that you have better chances of winning a game as the worse team when you slow the game down. Or they were just tanking

If you use it for multiple years, what is the R^2 of the current regression on multi-year RAPM? (Like I did for ASPM?).

Something doesn't sound quite right.

Actually, I think I can explain Jordan. If you are not looking at a benefit for high usage or usage*AST%, then some of the upper end effects are not captured.

Also, you might look into a team adjustment? For defense, that can be pretty significant, because box-score stats without a team adjustment tell you very, very little about a player's impact.

Right now there are no team stats involved. That makes it simple and applicable to other leagues, where people might not have access to pace-rating etc.

You may not have pace, or ORtg or DRtg, but you almost always have team and opponent PPG.
PER (as an example) adjusts for pace, but not for point-differential, so players from good and bad teams are not distinguished.

We know the NBA at large scored 131 ppg in games involving Denver, in 1991. Scoring less than 1/5 of that rate is sub-par for a Nuggets player.
In 36 minutes, a player scoring less than 19.6 points was generally being outscored by his counterpart.