APBRmetrics

Thanks for the comments.

J.E. your "boxscore metric" used as part of the prior is a form of SPM with certain other factors (height, GS, MP) also included, correct or no?

Crow wrote: J.E. your "boxscore metric" used as part of the prior is a form of SPM with certain other factors (height, GS, MP) also included, correct or no?

Here is what J.E. wrote:

viewtopic.php?f=2&t=8025&start=30#p13830

J.E. wrote:Here are the weights I found for offense and defense. Everything's scaled to "influence of height on offense"

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

I divided the boxscore totals of each player first by (team pace adjusted) minutes. League average is subtracted from a player's per minute totals, the result is being weighed by the listed weight. So if a player had more than average amount of statistic X, which also had a positive weight, that's good for his rating. And
having less than average amount of statistic X, which had a positive weight, is bad for his rating. And
having less than average amount of statistic X, which had a negative weight, is good for his rating.

When I use this new BoxScore metric to build a better prior-informed RAPM, tests seem to suggest that I should use
0.85*(0.65*last_years_RAPM+0.35*last_years_box_rating)
as a prior to compute next seasons RAPM (which then again gets combined with the BoxScore rating to form a prior for the next season, etc)

So, yes, it is an SPM derivate.

I was pretty sure it was, but thought I'd raise it to 100% confirm it and, by doing so, move to say that J.E. was now doing what talkingpractice does / suggests is the best approach.

Fantastic stuff, J.E.

I may have missed this, but how do you adjust for height and experience? I'm assuming you don't divide these by player minutes like the rest of the box score because obviously height and exp don't increase with more minutes. So I'm just curious how you factor these in with the rest of the pace/minute-adjusted box score stats.

Also, do you plan on sharing historical results for the box score only metric? I realize it won't be nearly as predictive as the prior informed RAPM hybrid results, but I'm still interested to see how the boxscore method treats historical players (yes, I saw the results of the earlier iteration you posted in this thread, but you've obviously made quite a few changes since then).

James

jbrocato23 wrote:I may have missed this, but how do you adjust for height and experience? I'm assuming you don't divide these by player minutes like the rest of the box score because obviously height and exp don't increase with more minutes. So I'm just curious how you factor these in with the rest of the pace/minute-adjusted box score stats.

Yeah some of them don't get divided by MP, all the %s and GS etc. After (not) dividing by MP all stats get converted to % in relation to league average. If you're 2.15m (~7 ft) and league average is 2.00m you should get a 2.15/2 - 1 = 1.075 - 1 = 0.075 for height.

Also, do you plan on sharing historical results for the box score only metric? I realize it won't be nearly as predictive as the prior informed RAPM hybrid results, but I'm still interested to see how the boxscore method treats historical players (yes, I saw the results of the earlier iteration you posted in this thread, but you've obviously made quite a few changes since then).

My plan was to first further improve "fake RAPM" (see viewtopic.php?f=2&t=8067 ). Once I feel it's good enough I would just compute xRAPM for the 90s in the same way as I do now for the 2000s, with the difference being that fake matchupfiles instead of real matchupfiles are used

J.E. wrote:Yeah some of them don't get divided by MP, all the %s and GS etc. After (not) dividing by MP all stats get converted to % in relation to league average. If you're 2.15m (~7 ft) and league average is 2.00m you should get a 2.15/2 - 1 = 1.075 - 1 = 0.075 for height.

Ok, that makes sense.

My plan was to first further improve "fake RAPM" (see viewtopic.php?f=2&t=8067 ). Once I feel it's good enough I would just compute xRAPM for the 90s in the same way as I do now for the 2000s, with the difference being that fake matchupfiles instead of real matchupfiles are used

Great! I'm excited to see how this turns out.

J.E. wrote:This new metric consisting of recursively informed RAPM+BoxScoreRating is now on my website as "xRAPM" and has replaced RAPM

There are full lists of players for each year, like here http://stats-for-the-nba.appspot.com/ratings/2007.html
and team pages with player ratings of '13, like this http://stats-for-the-nba.appspot.com/teams/OKC.html

Sorry for the bump. Just had a chance to read through this thread a little more carefully, but I have a confusion regarding the difference between BiRAPM and xRAPM. For xRAPM, is the BoxScore rating being used as a prior, or is xRAPM simply the average of BoxScore rating and RiRAPM (which, in an earlier post, I think you said gave the more accurate out-of-sample prediction)?

Edit:

OK, I reviewed some of the posts in this thread just now a bit more carefully, and I think I understand. xRAPM is essentially using the prior season's xRAPM (multiplied by 0.85) as a prior to compute the RAPM for the current season. And then, it takes 0.65*RAPM + 0.35*BoxScoreRating to get the current season's xRAPM. Do I have that right?

deepak wrote: Sorry for the bump. Just had a chance to read through this thread a little more carefully, but I have a confusion regarding the difference between BiRAPM and xRAPM. For xRAPM, is the BoxScore rating being used as a prior, or is xRAPM simply the average of BoxScore rating and RiRAPM (which, in an earlier post, I think you said gave the more accurate out-of-sample prediction)?

Edit:

OK, I reviewed some of the posts in this thread just now a bit more carefully, and I think I understand. xRAPM is essentially using the prior season's xRAPM (multiplied by 0.85) as a prior to compute the RAPM for the current season. And then, it takes 0.65*RAPM + 0.35*BoxScoreRating to get the current season's xRAPM. Do I have that right?

That's not exactly right but I know it's confusing.

For a season that is finished, the final rating consists of 0.65*RAPM(informed with xRAPM of the season before)+0.35*BoxScoreRating. Assuming a slight regression to league mean, this value gets multiplied with 0.85 to build a prior for next season.
For a season that's not finished I think it should be (1.0-0.35*%of season done)*RAPM+(0.35*%of season done)*BoxScoreRating.
Unfortunately I haven't computed the BoxScore part for the '13 season, so the current ratings on my site are just prior(=last years rating) informed RAPM, with no BoxScoreRating of this season.
So, ironically, the part (BoxScore) that always gave Kevin Love a rating higher than what he probably deserved has no weight in the only season in which his BoxScoreRating would actually be below average

Sorry for the late answer

J.E. wrote:

deepak wrote: Sorry for the bump. Just had a chance to read through this thread a little more carefully, but I have a confusion regarding the difference between BiRAPM and xRAPM. For xRAPM, is the BoxScore rating being used as a prior, or is xRAPM simply the average of BoxScore rating and RiRAPM (which, in an earlier post, I think you said gave the more accurate out-of-sample prediction)?

Edit:

OK, I reviewed some of the posts in this thread just now a bit more carefully, and I think I understand. xRAPM is essentially using the prior season's xRAPM (multiplied by 0.85) as a prior to compute the RAPM for the current season. And then, it takes 0.65*RAPM + 0.35*BoxScoreRating to get the current season's xRAPM. Do I have that right?

That's not exactly right but I know it's confusing.

For a season that is finished, the final rating consists of 0.65*RAPM(informed with xRAPM of the season before)+0.35*BoxScoreRating. Assuming a slight regression to league mean, this value gets multiplied with 0.85 to build a prior for next season.
For a season that's not finished I think it should be (1.0-0.35*%of season done)*RAPM+(0.35*%of season done)*BoxScoreRating.
Unfortunately I haven't computed the BoxScore part for the '13 season, so the current ratings on my site are just prior(=last years rating) informed RAPM, with no BoxScoreRating of this season.
So, ironically, the part (BoxScore) that always gave Kevin Love a rating higher than what he probably deserved has no weight in the only season in which his BoxScoreRating would actually be below average

Sorry for the late answer

Thanks. So, it sounds like I essentially had it right, except for the fact that mid-season the current BoxScoreRating won't factor in as heavily (or at all)?

If a player's xRAPM is +4 in season x and +4 in season x+1, then it seems to me that there is currently an ambiguity. Did the regression analysis really say he close to a +4 each year (case 1)? Or was the +4 for the second year almost totally due to using the previous season as a prior, because the regression analysis wasn't able to rate the player's impact in the second year for whatever reason (say, not enough possessions or colinearity) (case 2). I would be somewhat more confident that his +4 is an accurate reflection of true impact in the first case than the second case.

So, first of all, does my concern even make sense? And if it does, is there any additional information that can be provided to help us distinguish the two cases? Like 1-year RAPM?

deepak wrote:Thanks. So, it sounds like I essentially had it right, except for the fact that mid-season the current BoxScoreRating won't factor in as heavily (or at all)?

Well it *should* factor in, just not with a 0.35 weight starting with the first game.

If a player's xRAPM is +4 in season x and +4 in season x+1, then it seems to me that there is currently an ambiguity. Did the regression analysis really say he close to a +4 each year (case 1)? Or was the +4 for the second year almost totally due to using the previous season as a prior, because the regression analysis wasn't able to rate the player's impact in the second year for whatever reason (say, not enough possessions or colinearity) (case 2). I would be somewhat more confident that his +4 is an accurate reflection of true impact in the first case than the second case.

Not playing at all would obviously move the rating from 4 to 0.85*4 at the end of next season, so to stay at +4 level he'd have to play at least some minutes with positive contribution. I guess you could ballpark in which case you're in for each player by looking at his # of possessions for that season, but I'm not sure if we need special treatment for the player depending on which case you find.

Hey

'13 ratings are back up again at http://stats-for-the-nba.appspot.com/ratings/2013.html

The BoxScore rating, which has a 0.35 weight compared to end-of-season xRAPM's 0.65, was computed using the weights posted here viewtopic.php?p=13830#p13830

#1 and #2 in the BoxScore rating are Durant and LeBron.
Doron Lamb is in last place. He had a 4.8 PER, 86 ORtg, 112 DRtg, 41% eFG, 4.4% TReb, 0% BLK, 8.2% AST

Thoughts...

interestingly, not that high on lbj defensively. still best player in the league with an absurd double digit rapm. in fact, he's 20% better than the second best player (Cp3). biggest gap itl, obv.

durant fractionally behind cp3. that's the top3, with Duncan in that tier. Cool that RAPM still sees Duncan as a deserving 1st team All NBA.

Also had Duncan as the DPOY. Should have gotten more votes. Dwight, in a down year, was 2nd best there. It was his offense that was quite weak.

few other notables:
-picked up on TRG's defensive astuteness. DPOY voting may not have been correct, but it wasn't far off (4th)
-Millsap didn't get all-NBA pub and was likely, as usual, shafted

**big test of RAPM: Will Jazz get worse? Atlanta better?
On a strict RAPM basis, Atlanta profits about 2pts w/ Millsap over JSmoove (although Smoove still rates pretty well) and the Jazz, after losing Big Al (+1.2) and Millsap (+5.4) go to Favors (+1.9) and Kanter (-0.6)

-Norris Cole still WOAT
-Harden better offensive player than Durant? Marginally.
-Amir Johnson? 14th best player ITL.
-Asik->Dwight... Not much of an improvement
-Brandon Jennings: Super negative player. Jeff Teague? League average.
-Rubio was the highest rated defensive PG. Not a surprise. Conley 2nd.
-Nick Collison still clocks in as a RAPM wunderkin. His +1.5 is much better than Perk's -1.4. Perk is one of the most harmful offensive players ITL.

Other random offseason thoughts on the basis of RAPM...
Nuggets offseason is seriously horrible. Iggy is +4.1 (23rd best), Koufos +1.8, Corey Brewer only -.5. Gallo, fwiw, is +3.8. Without these guys this suggests Nuggies may be below .500. They replaced them with Hickson -1.4, Foye -3.0, and Arthur -5.0.

Think what you want but I am not so sure Denver is a playoff team after all this nonsense. Depending on minute allocations and aging (or, health, ie. 'Dre) Denver is the team likely to regress most. That's when you also discount the GK regular season effect.

The Nets gained KG (+4.1), Pierce (3.8), AK (2.5), while dumping Hump (-1.0) and G-Force (+1.8). 3 major pieces, probably should be taken that much *more* seriously.

Going from Collison and Mayo to Calderon and Ellis does help the Mavs, but only marginally. Going from Kaman to whomever else, on the other hand, could help a bit more. (Although Sammy D isn't that much better) They were close to the playoffs, and that was with Dirk injuries. As much as we could dismiss some of their moves, the Mavs are the most likely team to take Denver and the Lakers spot.

The Jazz, as noted, regress further. No playoffs for them.

Pelicans? Jrue is better than Vasquez, but barely a pt per 100 better. Swapping the 2, assuming modest gains due to age, and holding minutes constant is not worth very much. Tyreke is a league average player and apparently Aminu is better. (Aminu clocks in as a BEAST defender. Makes the Pelicans' 28th overall rating that much worse) Ryan Anderson took a beating defensively. ROLO (+2.0) was the best Pelican...
Again, depending on minute distribution as well as some slight aging modifiers (in their favor), RAPM still views this as a below .500 team. Not sure I agree but take it for what it's worth.

On that note, Portland's minor/under the radar moves should pay massive dividends. ROLO >> Hickson and Dorell Wright (+0.0) is much better than any Portland backup. Lillard (+.6) put up a very nice RAPM considering he was a rook.

Dudley (+.7) and Redick (-1.3) are nice upgrades, but that's mainly because of how bad Willie Green (-4.4) and Caron (-2.4) were. Really, the Clips should have just played Bledsoe (+1.7) at the 2 and promoted Barnes (+.9) as the full-time starter. Probably smarter, better, cheaper, and still younger.

Earl Clark (-4.1) should be so thankful for the LA publicity. It made him literal millions. According to RAPM, Artest is very much *not* done. +1.2, his offense is bad but he is still a very good wing stopper. Wonder who will pick him up via amnesty.

As noted, the Hawks scooped up Millsap. DMC though, as a +0.9, is an under-the-radar, good player. Korver, in heavy minutes, was +0.7. Underrated wing combo that may get laughed at, but will still produce. I have the Hawks as a very solid playoff team. Nothing special, but a fantastic cap situation with some decent long-term assets. Well-built franchise ATM, imo.

According to nearly every other metric, Warriors will greatly miss Landry (-1.6) and Jack (-0.7). RAPM simply doesn't agree as it's almost addition-by-subtraction. Iguodala is massive, but Barnes (-1.3, not bad for a rook), Ezeli (-1.7) and Green (-1.8) will all improve as sophomores. Scary to think, but in a super weakened Pacific (Kings maybe worse, LAL obv much worse, and PHX mega tanking) esp. with potential issue with Clips' health (CP3), they could take their division. Especially since Bogut (+2.0) missed so many games, simple health raises their bar. Their fodder to the Jazz were all ~worthless (Biedrins -1.7, Jefferson -2.3, Rush -1.0). Douglas (-1.0) is nearly on par to Jack, whereas Jermaine O'Neal (+0.3) was reborn.

A nice general "who's who" list of NBA players. I would personally rank some players higher or lower (sorry, Paul Millsap!), but hey; the numbers aren't gospel.

Gasol's DPOY was deserved. Dwight? From the tape, I thought it was an inconsistent season for him on defense, although he had to deal with his nagging back and shoulder injuries. He did perform better post All-Star break, but I wouldn't put him at 2nd overall on defense. I think some of his impact had to do with the rest of the Lakers being a subpar defensive team (rather than him being the same uber-dominant defensive player he was in Orlando), so he carried a lot of the defensive burden. Houston will enjoy a healthy Dwight in the lineup however, and he'll round back into form.

LeBron is one of the finest defensive players in the league. He's not as good of a defender as the Gasols, Duncans, etc. of the world simply because he's a perimeter defender as opposed to a rim-blocking big. But his versatility to guard multiple postions when Spolestra calls for it and ability to cover ground from wing to wing is unmatched. It is an important reason why the Heat could scrap traditional lineups and play small ball, and I certainly wouldn't favor him over someone like Ricky Rubio for example (higher DxRAPM than LeBron) when constructing a defensive team.