Questions and Comments on ASPM

[DSMok1]

Split off from topic "More with RAPM"
viewtopic.php?p=11521#p11521

Kemba Walker with the 10th worst mark in the league on this factor and in the bottom 25 on overall A4PM. But almost everybody in the league and outside of it is going to continue to use boxscore metrics exclusively or almost exclusively and think somewhat or a lot better of him and almostly complete or completely ignore APM and his estimates on it. Call me when K Walker is actually leading his team to a good result. If that ever happens I am pretty confident that his RAPM won't be bottom 25 in the league as it is now.

ASPM is a box score metric, and I have him at -1.2 so far this year--below average, but above replacement level. That puts him in the company of Sessions, Duhon, Isaiah Thomas, and Earl Watson. http://godismyjudgeok.com/DStats/aspm-a ... 2012-aspm/

We'll see if his RAPM continues to have him below replacement level.

[Crow]

ASPM is a boxscore metric that uses stat weights established by underlying regressions which are conceptually similar to the ones being used in RAPM though right? So it is not a simple or pure boxscore metric IMO and is a cousin of APM in that they are both using regression based weights to determine the value of boxscore production elements.

[DSMok1]

Crow wrote:
ASPM is a boxscore metric that uses underlying regressions similar to the ones used in RAPM though right? So it is not a simple or pure boxscore metric and is a cousin of APM.


ASPM is a pure boxscore metric, like Win Shares or Wins Produced or PER. No play-by-play data or APM-style regression used, except as the baseline for figuring out the weights for the boxscore data. The full derivation is here: http://godismyjudgeok.com/DStats/aspm-and-vorp/

[Crow]

Sorry Daniel, I accidentally and temporarily posted over your last comment (a minor hazard of moderator privileges) but was able to restore it, though with slightly different formatting.

[Crow]

Wins Produced also has regression based weighting of the team impact of individual boxscore performance. So does PER, based on Hollinger's report of using such regression based weighting. I don't know what Oliver used for his weights for Offensive Rating that get used as the basis for calculating win Shares.

I probably shouldn't have attacked all boxscore metrics as bluntly. And I may or may not be making things clearer by now calling them cousins or near-cousins. There is a fundamental difference between them in that APM estimates the specific team impact of a player based on his specific boxscore and nonboxscore performance while these cousins or near cousins found weights that estimate average player boxscore impact across all league data and then use the average impact as an estimate of specific player production impact for all players alike. And miss or mis-assign player impact on shot defense. So in the end the differences are significant even if there is also some similarity.

[Mike G]

... The full derivation is here: http://godismyjudgeok.com/DStats/aspm-and-vorp/

"Advanced" stats are used in lieu of raw totals or rates: Reb% is rebounds per available rebound, Ast% is assists per teammate FG, etc.
But the 'scoring' component isn't standard to anything? Why not % of points scored?
Even if it isn't in b-r.com's Advanced section, it's still an advancement in measurement.
Scoring is the biggest part of many players' stat profile, and scoring 20 in 90-85 games is surely bigger than scoring 20 in 115-110 games.

[Crow]

Mike G: "Why not % of points scored?"

Why not points scored over points expected for that shot level or usage level instead? A la some of the stuff Evan fairly recently computed (just boxscore, not adjusted shooting and scoring impact though I also value taking it to that even broader perspective). I prefer much more emphasis on scoring efficiency impact than simply or heavily scoring rate.

[DSMok1]

"Advanced" stats are used in lieu of raw totals or rates: Reb% is rebounds per available rebound, Ast% is assists per teammate FG, etc.
But the 'scoring' component isn't standard to anything? Why not % of points scored?
Even if it isn't in b-r.com's Advanced section, it's still an advancement in measurement.
Scoring is the biggest part of many players' stat profile, and scoring 20 in 90-85 games is surely bigger than scoring 20 in 115-110 games.

The scoring component is the most complex part of the regression. I'm using TS% and USG%, which are both per possession, which is better than % of points scored.

[Mike G]

But on the team level, % of points scored is everything -- it determines who wins the game, the margin of victory...
You might shoot .600 one night and lose, shoot .400 the next night and win.
If on those nights, your opponent shot .650 and .350, respectively, would you say your scoring efficiency was better in the loss?

Because basketballers play both offense and defense, it seems the value of [points per possession] is always relative to that of the opponents.
Giving up more than you score is inefficient, except in the narrow definition (ignoring defense) used by PER and similar metrics.

We generally assume constant rates of rebounds (for a given team) when we calculate Reb%, even though it may change when the lineup changes.
We can just as well assume a player's scoring rate yields a % of the points scored while he's in the game.

How can the most important team statistic not be one of the most important individual statistics?

[DSMok1]

But on the team level, % of points scored is everything -- it determines who wins the game, the margin of victory...
You might shoot .600 one night and lose, shoot .400 the next night and win.
If on those nights, your opponent shot .650 and .350, respectively, would you say your scoring efficiency was better in the loss?

Because basketballers play both offense and defense, it seems the value of [points per possession] is always relative to that of the opponents.
Giving up more than you score is inefficient, except in the narrow definition (ignoring defense) used by PER and similar metrics.

We generally assume constant rates of rebounds (for a given team) when we calculate Reb%, even though it may change when the lineup changes.
We can just as well assume a player's scoring rate yields a % of the points scored while he's in the game.

How can the most important team statistic not be one of the most important individual statistics?

So you are arguing that "percentage of total points scored" is a useful individual statistic?

At their core, APM, RAPM, SPM, and ASPM are "possession" metrics. The objective is to predict the team's offensive rating (points per possession) and defensive rating. Points, per se, are not all the same--the value of points scored depends on how many possessions were used to generate them, and the defense on the other end is not related to the offense. That is what ASPM attempts to capture. This variation in the value of points is not easy to capture with "percentage of total points scored".

[Mike G]

[A thread will get quite bloated if we quote the entire preceding post.]

... the value of points scored depends on how many possessions were used to generate them, and the defense on the other end is not related to the offense. ..

The ratio of points for/against is the same per game, per possession, per100, per season.
Players go essentially equal numbers of offensive and defensive possessions. The value of their scoring is only measured vs the scoring of their opponents.

[deepak]

[A thread will get quite bloated if we quote the entire preceding post.]

... the value of points scored depends on how many possessions were used to generate them, and the defense on the other end is not related to the offense. ..

The ratio of points for/against is the same per game, per possession, per100, per season.
Players go essentially equal numbers of offensive and defensive possessions. The value of their scoring is only measured vs the scoring of their opponents.

But how many points the other team is scoring has very little to do with a player's scoring ability. That falls under defense, and by my understanding that should be accounted for by the other factors in the ASPM formula.

[DSMok1]

But how many points the other team is scoring has very little to do with a player's scoring ability. That falls under defense, and by my understanding that should be accounted for by the other factors in the ASPM formula.

Precisely. Offense and Defense are decoupled in the regression--unrelated. How well you score, per possession, is not related to how well you defend.

[schtevie]

But how many points the other team is scoring has very little to do with a player's scoring ability. That falls under defense, and by my understanding that should be accounted for by the other factors in the ASPM formula.

Precisely. Offense and Defense are decoupled in the regression--unrelated. How well you score, per possession, is not related to how well you defend.

Theoretically, O and D are not unrelated. My understanding is that offenses are less productive after opponents' made baskets than otherwise (for the obvious reasons). Should we believe that this effect is small enough to be ignored? Or does ASPM take it into account?

[DSMok1]

Theoretically, O and D are not unrelated. My understanding is that offenses are less productive after opponents' made baskets than otherwise (for the obvious reasons). Should we believe that this effect is small enough to be ignored? Or does ASPM take it into account?

True--I am assuming it is small enough to be ignored--the team adjustments somewhat take it into account, though, since the offensive & defensive efficiencies must sum to the observed efficiencies.