Inferred playoff RPM?

[Mike G]

Team RPM for playoffs.
It's (see previous post) player inferred RPM * player minutes; summed; divided by team minutes; *4 -- I don't recall the accepted multiplier, but it's less than the intuitive 5. In any case, ranks will be the same and proportionate.

Code: Select all

iRPM   tm     iRPM   tm
16.6  GSW      2.4   Mil
13.9  Cle      2.2   Ind
6.3   SAS      1.6   Atl
5.0   OkC       .4   Mem
4.7   Uta      -.1   Tor
4.5   Hou      -.4   Bos
3.0   LAC     -5.7   Chi
2.8   Was     -6.7   Por

Wizards the 2nd best team in the East.
Without the Finals input -- GS by 7 ppg over Cle -- it seems they should have been pretty evenly matched.

[Crow]

Thanks to Shadow and Mike G for stepping in and carry the idea further. I don't always have the time or ability to max it out. This is what I hope for.

[shadow]

I usually multiply by 5 and divide by 1.2 to add the 'up by X affect' back in to team ratings. So my effective multiplier is 4.167. I forgot where I found that, but it usually works pretty well.

Here are my final playoff only ridge regressed team ratings:

Code: Select all

GSW	+11.82
CLE	+8.39
SAS	+2.1
IND	+1.88
HOU	+0.24
UTA	-0.64
WAS	-1.08
ATL	-1.19
LAC	-1.25
MIL	-1.46
TOR	-2.1
BOS	-2.24
POR	-2.64
MEM	-3.29
OKC	-3.75
CHI	-4.79

[Crow]

OKC is very different by the playoff implied RPM and playoff only RAPM.

[Mike G]

And is it possible the espn RPM weighs playoff minutes more heavily? That might produce the exaggerated effect in the team RPM I recreated.

[Mike G]

At this point I'm interchangeably using per-game, per-48-min, and per-100-poss; and I don't much care.
Here are postseason leaders in plus-minus points by way of inferred RPM*Min/48

Code: Select all

+pts   playoffs       tm   min  iRPM    BPM   WS/48   PER
195  LeBron James    Cle   744  12.6   11.4   .273   30.0
126  Draymond Green  GSW   593  10.2    8.1   .196   17.9
112  Kawhi Leonard   SAS   429  12.5   11.3   .314   31.5
110  Stephen Curry   GSW   601   8.8   10.8   .272   27.1
85   Kevin Love      Cle   578   7.1    3.2   .178   19.7

68   Kevin Durant    GSW   533   6.2    8.7   .281   27.5
68   Otto Porter Jr. Was   428   7.6    3.1   .166   16.5
56   Jae Crowder     Bos   596   4.5     .7   .108   14.5
50   Markieff Morris Was   373   6.4   -1.4   .052   12.3
44   Kyle Lowry      Tor   300   7.1    2.5   .116   15.7

+pts   playoffs       tm   min  iRPM    BPM   WS/48   PER
43 Russell Westbrook OkC   194  10.6   13.1   .103   27.7
41   James Harden    Hou   407   4.9    7.9   .157   23.0
41   Klay Thompson   GSW   596   3.3   -1.5   .049    9.3
40   George Hill     Uta   281   6.8     .5   .075   13.4
38   Mike Conley     Mem   224   8.1    8.5   .204   27.4

37   Joe Ingles      Uta   334   5.4    3.4   .064   10.2
36   Chris Paul      LAC   260   6.7   12.1   .246   27.8
36   Kyrie Irving    Cle   653   2.7    1.2   .147   21.3
28   Andre Iguodala  GSW   419   3.2    1.6   .064   10.5
24   Pau Gasol       SAS   365   3.1    2.6   .093   15.1

edited to add BPM, WS/48, and PER

[shadow]

And is it possible the espn RPM weighs playoff minutes more heavily? That might produce the exaggerated effect in the team RPM I recreated.

All games have equal weight according to J.E.
https://twitter.com/JerryEngelmann/stat ... 8204209152

[shadow]

Added off/def splits to playoff only RAPM and adjustments for HCA and scoring margin at start of each matchup stint.

https://docs.google.com/spreadsheets/d/ ... ng/pubhtml

[Mike G]

That's impressive work. Due to small sample, everyone tends to zero; Curry on top with a modest +3.80, LeBron and Draymond the only others > 2.5

You also show only 5 teams above zero. That's about the same as the # which won more than half their games or had positive pt-diff. And this would be due to zero priors for everyone?

So it seems that if this is calculated before the Finals, the East and West would both sum to zero -- for players or for teams -- and with the Finals, everyone in the West gets a big boost, esp. relative to the East.

Is there really any advantage to doing it this way, rather than the inferred method? Crow and I used bigger (full season) samples, subtracted from even bigger samples, and got somewhat irrefutable results. Like A + B = C

LeBron's (2794 min) RS RPM was 7.32; now it's 8.42 after 744 PO minutes. The PO RPM that gets him there is 12.57, not 3.57

[jgoldstein34]

Well Shadow's using RAPM, not RPM. That explains a bit of the difference. And small, playoff only, sample size brings that all to zero for RAPM.

[shadow]

Just to clarify, the team ratings I posted are not built from player RAPM. They're simply team point differential adjusted for SoS and HCA. It is true that the because of the ridge regression teams who didn't advance far in the playoffs will not deviate far from the zero prior.

I allowed the software to select the best lambda via 10 fold cv for the playoff only RAPM and it still selected a reasonably high value of 1880 compared to the 3000 value typically used for full season RAPM. Combined with the relatively small sample size of 79 games that's going to result in a pretty small range of ratings. This probably isn't fair to players on teams like CLE or GSW who advanced deep into the playoffs.

If I prorate the typical 3000 lambda value based on the number of games played by GSW/CLE, I get a value of around 640. Re-running the playoff only RAPM with that lambda results in LeBron having a rating of 9.54. That seems like a more reasonable estimate to me.

LeBron wouldn't have to necessarily play at a +12.57 in the playoffs for his rating to increase from 7.32 to 8.42. Hypothetically, if he played at a +8.42 level every game, his rating would gradually increase as he accumulates more minutes, since the ridge regression would become more confident he is a 8.42 as the sample size becomes larger. There's also the factor of the boxscore prior becoming less important as the number of minutes played increases for a player, which would also allow his rating to increase even if his RAPM component stayed constant (assuming it was greater than his box score prior).

[Mike G]

Thanks for the clarifications.
I still like the idea of RPM, RAPM, BPM, WS, PER, etc for playoff series, as they're in progress.

[Mike G]

Players from the bottom and top of "more favored by RAPM (shadow) vs RPM (ESPN) inferred".

Code: Select all

playoffs    tm   min  RAPM  iRPM   PER   ws/48   BPM
Chris Paul LAC   260   -.2   6.7   27.8   .246  12.1
JR Smith   Cle   487   1.5   -.5    9.9   .095   2.5

This may just demonstrate how much of RPM is boxscore influenced, while playoff-only NPI RAPM is purely team performance, with and without a player.
Not many players are positive in one and negative in the other.

JR Smith did very little that appeared in the boxscores. But he played a lot of minutes; partly because his backups were a serious downgrade; but apparently he was actually pretty good at something that isn't a boxscore stat.

Here's a list of correlation between player minutes (not per game) and these 2 things claiming to be plus-minus.

Code: Select all

tm    W   iRPM   RAPM
GSW  16    .83    .90
Cle  13    .71    .81
Bos   9    .37   -.35
SAS   8    .13   -.40
Was   7    .62    .67
Hou   6    .54    .02
Uta   4    .25   -.36
Tor   4    .47    .11
LAC   3   -.05   -.42
Mem   2    .31   -.27
Mil   2    .23   -.14
Atl   2    .04   -.39
Chi   2   -.23   -.68
OkC   1    .63    .47
Ind   0    .38    .25
Por   0    .34   -.78

Chi-RR     .55   -.62
Hou-N      .76    .32
LAC-BG     .42   -.16
Bos-IT     .29   -.05

The correlations are for the top 9 players (minutes) in each lineup. At bottom, I've manually removed Rondo, Griffin, Nene`, and Isaiah from their teams' top 9. This helps all their correlations.

Outside of the 2 finalists, there are few instances in which coach seems to believe in the numbers?
RAPM of course pushes everyone with low minutes toward zero. The poor Blazers were overwhelmed, nobody could be found who could outplay the opposition; and their 'worst' players in RAPM also went the most minutes.

Is there a case to be made for assuming bench scrubs are not in fact 'average' -- vs the Warriors, that means very, very good -- but in fact are the marginal players / replacement types their coaches deem them to be?
If they are mathematically pushed toward (say) -5 instead of zero, surely that helps the RAPM of Lillard and McCollum. They weren't really the worst players on team, were they?

Code: Select all

. Blazers       Min   iRPM   rapm    BPM
Damian Lillard  151   -4.0   -1.5    2.1
CJ McCollum     140    -.3    -.9   -4.2
Evan Turner     124    -.3    -.4    2.2
Al-Farouq Aminu 113   -1.4    -.6    -.4
Noah Vonleh     100   -6.8    -.7   -1.0

Maurice Harkless 99    3.3    -.6    -.1
Allen Crabbe     92   -5.3    -.5   -4.8
Shabazz Napier   47   -2.9     .1    2.9
Meyers Leonard   31   -6.0    -.4  -10.2

Well there is little agreement.
It's a "small sample". But a coach has to make decisions based on much smaller samples -- one or 2 games or quarters, if he means to survive.