Possible Steps to improve SPM

[DSMok1]

dsmok1, if asst% helps on offense, would it help on defense? if you reference player and team ts% on offense, have you tried doing so on defense? Is the confidence interval on steals and blocks (among the few defensive variables) comparatively high? Can that be reduced with more defensive variables? Have you considered defensive versions of most or all of your interactive variables? Overall, why isn't the defensive side as much as possible the flipside of what you've with the offensive side? Even if you dont do exactly the same, why not at least do something to simulate that part of player impact? Why are you comfortable or resigned to leaving the defense without shot defense, spacing defense and assist defense beyond steals?

If, as I expect, the main answer is that counterpart defensive data are not perfect because of team defense and imperfect matchup, I'll go back to is it better to be right 60-70% and off somewhat on the rest or just leave shot defense and other nuance out 100%? Would it be worth even a trial look or simultaneous existence? If you still opt to leave a lot out, do you think you should have bold notes or footnotes stating this so the lesser informed on the details know about these major absences in ASPM?

Counterpart stats haven't been tracked traditionally, and ASPM is for use back to the 1970s. In addition, I have not seen any good information that showed counterpart data to be useful--I have seen quite a bit indicating it is deceptive. ASPM will stick with this basic approach to defense, and should in general be accompanied by a caveat about box scores and defense.

I still think that this can get you pretty close; the R^2 for ASPM defense is 0.54 onto long term RAPM.

Eyeball test of last season:

Code: Select all

╔══════╦══════════════════╦══════════════╗
║ Rank ║     Player       ║ DASPM Rating ║
╠══════╬══════════════════╬══════════════╣
║    1 ║ Joakim Noah      ║ 5.1          ║
║    2 ║ Andrew Bogut     ║ 5.0          ║
║    3 ║ Draymond Green   ║ 4.0          ║
║    4 ║ DeAndre Jordan   ║ 3.8          ║
║    5 ║ Tim Duncan       ║ 3.7          ║
║    6 ║ Ian Mahinmi      ║ 3.5          ║
║    7 ║ Roy Hibbert      ║ 3.3          ║
║    8 ║ Kawhi Leonard    ║ 3.2          ║
║    9 ║ Kyle O'Quinn     ║ 3.1          ║
║   10 ║ Elton Brand      ║ 2.7          ║
║   11 ║ Anderson Varejao ║ 2.7          ║
║   12 ║ Marc Gasol       ║ 2.6          ║
║   13 ║ Greg Stiemsma    ║ 2.5          ║
║   14 ║ David West       ║ 2.5          ║
║   15 ║ Steven Adams     ║ 2.4          ║
║   16 ║ DeMarcus Cousins ║ 2.4          ║
║   17 ║ Jimmy Butler     ║ 2.4          ║
║   18 ║ Tony Allen       ║ 2.4          ║
║   19 ║ Andre Iguodala   ║ 2.4          ║
║   20 ║ Kevin Garnett    ║ 2.3          ║
╚══════╩══════════════════╩══════════════╝

[Crow]

Just one player below 6-7. Not surprising. Defensive rebounds and blocks included
(though not heavily weighted), shot defense not.

[DSMok1]

Just one player below 6-7. Not surprising. Defensive rebounds and blocks included
(though not heavily weighted), shot defense not.

That's not much of a surprise, though--look at the top players in the 14 year Defensive RAPM: http://public.tableausoftware.com/profi ... 14YearRAPM
(Sort by Defense per 100)

Almost no guards in the top 50. Just Tony Allen, Bruce Bowen, and a couple others.

[Crow]

Good point.

Further info: 11 of top 50 for efg% allowed are under 6-7 here
http://www.gotbuckets.com/statistics/ff ... m-defense/



All but maybe 8 of the lowest on ft allowed (or ft.fga, not sure what they did) are under 6-7. Is that best or worst?

[mtamada]

I'm not sure which variables you already are using but in addition to the counterpart stats already mentioned, how about stats relative to team and/or league averages, and stats which account for the quality of the team? Stats such as Usg% and Rebd% do already account for the team context in significant ways, but not completely. A high Usg% on a very good offensive team may be different from an equal Usg% on a team which is poor offensively.

[J.E.]

he's 7th on the "actual points scored against - expected points scored against" list

out of curiosity just what is this list?...

As I described in the 1st post I'm creating a list of players that my SPM has under-overrated since 2001. I already posted the list for offense. The top/bottom 10 for defense are

Code: Select all

╔══════════════════════╦════════════════════════════════════════════════════╗
║        Player        ║ Actual Points scored against – Exp. Points sc. ag. ║
╠══════════════════════╬════════════════════════════════════════════════════╣
║ Tony Parker          ║                                              -1064 ║
║ Tim Duncan           ║                                               -851 ║
║ Bruce Bowen          ║                                               -649 ║
║ Paul Pierce          ║                                               -594 ║
║ Eddie Jones          ║                                               -572 ║
║ Kirk Hinrich         ║                                               -554 ║
║ Jim Jackson          ║                                               -537 ║
║ Kenyon Martin        ║                                               -520 ║
║ Manu Ginobili        ║                                               -517 ║
║ Shawn Marion         ║                                               -502 ║
║ ..                   ║                                                    ║
║ Vladimir Radmanovic  ║                                                379 ║
║ Josh Smith           ║                                                425 ║
║ Jamal Crawford       ║                                                439 ║
║ Pau Gasol            ║                                                483 ║
║ Michael Redd         ║                                                490 ║
║ Al Jefferson         ║                                                546 ║
║ Jason Richardson     ║                                                567 ║
║ Troy Murphy          ║                                                593 ║
║ Antawn Jamison       ║                                                628 ║
║ Kobe Bryant          ║                                                793 ║
╚══════════════════════╩════════════════════════════════════════════════════╝

less is better. Ben Wallace has dropped to 20th after I made some changes

D is defined as total - O.

This is interesting. I just tried this and the difference in defensive rating when
a) fitting coefficients for D-RAPM only, and
b) fitting coefficients for both total and offensive RAPM and then defining D = TOT - OFF
is very small

Note--this was an unweighted regression

I think we should definitely use weights here. Getting the SPM of a 10 MP/season player "right" can't be as important as getting a superstars SPM right - the superstar will play way more minutes in the future; being "off" with the superstars' SPM will lead to worse team/lineup-predictions, while being off with the 10MP/season-player will almost not hurt at all.

An easy way to do this is figure out average career MP over the sample, then, for each player
weight = career_MP / average_career_MP
and then multiply each row by sqrt(weight)

Thus far I've usually done my SPM by fitting BoxScore stats onto PPP for every possession since 2002, but I'll fool around more with fitting BoxScore stats onto long term RAPM like DSMok1 does it. For one thing it's less expensive computationally

[DSMok1]

D is defined as total - O.

This is interesting. I just tried this and the difference in defensive rating when
a) fitting coefficients for D-RAPM only, and
b) fitting coefficients for both total and offensive RAPM and then defining D = TOT - OFF
is very small

Right, that's what I found also. I just get to do one less regression, and it helps keep the focus on the overall number.

Note--this was an unweighted regression

I think we should definitely use weights here. Getting the SPM of a 10 MP/season player "right" can't be as important as getting a superstars SPM right - the superstar will play way more minutes in the future; being "off" with the superstars' SPM will lead to worse team/lineup-predictions, while being off with the 10MP/season-player will almost not hurt at all.

An easy way to do this is figure out average career MP over the sample, then, for each player
weight = career_MP / average_career_MP
and then multiply each row by sqrt(weight)

Thus far I've usually done my SPM by fitting BoxScore stats onto PPP for every possession since 2002, but I'll fool around more with fitting BoxScore stats onto long term RAPM like DSMok1 does it. For one thing it's less expensive computationally

Right, for the actual final coefficients I did the full weighted methodology; I only used unweighted for the variable selection (but I used only players with a significant number of minutes played for that portion).

[Crow]

82games counterpart stat quality may have been surpassed by one or more of synergy, sportVu and vantage sports. Is the quality high enough? It may never be high enough for some. But it might be high enough for others. As Mike G points out, the existing boxscores are probably not perfect representations of reality, even beyond what they lack entirely.

[talkingpractice]

i think we need 1-2 more years of sportvu to have something useful for prediction.

i can already vouch for Vantage data.

[Crow]

Including the individual shot defense component? Are you using it to improve your spm or IPV?

[Crow]

JE, assuming your spm also has lacked to date shot defense of any kind, do you think that distorts rpm / xrapm? Do have any interest in using 82games or vantagesports data to try to plug most or all of this gap?

[talkingpractice]

Including the individual shot defense component? Are you using it to improve your spm or IPV?

we did find some benefit to using opponents fg% at the rim, but nothing else (for defense). offense doesnt have anything useful yet. but a lot of that is due to having just one year of sport vu data (i think).

heres an article about that from gotbuckets: http://www.gotbuckets.com/2014/06/17/an ... rtvu-data/

we wont use it this year tho, because we're clients of Vantage now, and their data subsumes sportvu.

[Crow]

Thanks for the link. I don't believe I had seen it. Not surprised that at the rim defense (rate and fg% allowed) was most significant. Was a bit surprised that they were the only shot defense variables tested in the group of seven. Would another shot defense variable such as 3pt efg% show as significant or close?
Decent chance I'd guess. An overall shot defense might too.

[permaximum]

That chart about most overrated defenders made me wonder, was Kobe a good defender or average at best in his prime? In 14-year RAPM (2001-2014) Kobe's D-RAPM is -0.9. Even worse than Iverson's who had at least 2 years of strong decline.

[J.E.]

JE, assuming your spm also has lacked to date shot defense of any kind, do you think that distorts rpm / xrapm? Do have any interest in using 82games or vantagesports data to try to plug most or all of this gap?

I'm wary with 82games counterpart data because we don't really know who a player is really guarding. And even then, teams play lots of P&Rs, which forces defenders to switch etc.

SportVU and Vantage data will be helpful at some point, moreso for defense than offense, but we need way more than just one season of data. Also, the SportVU data I've come in contact with had extremely severe quality issues, so I have some doubts