Appr. 5.x year reg. adj. +/- (J.E., 2010)

[J.E.]

Did you just essentially develop a means for estimating standard errors for RAPM?



That would be a big deal, if that is valid.

I suppose I did not. I'm assuming a couple of things that might not actually be true, like normal distribution and a specific standard deviation.

[EvanZ]

The fact that certain players had less error(?) than others (say Udoh), does that mean something? Or would this change every time you do a new simulation?

[J.E.]

The fact that certain players had less error(?) than others (say Udoh), does that mean something? Or would this change every time you do a new simulation?

It does change for every new simulation, but RAPM seems to capture some player's rating better than others.
What I listed above is RMSE for each player over 100 simulations.

I suppose if I did another 100 simulations it would paint a similar picture on terms of who has large/tiny RMSEs

[bbstats]

...yes, but the level of RMSE would not be determined by player inconsistency but by RAPM bias for their specific lineups/minutes. Yes?

[EvanZ]

Jerry, if you can figure out some kind of pattern (i.e. for player substitutions) that produces more (less) reliable ratings, obviously that would be a great help. Kudos.

[DSMok1]

Did you just essentially develop a means for estimating standard errors for RAPM?



That would be a big deal, if that is valid.

I suppose I did not. I'm assuming a couple of things that might not actually be true, like normal distribution and a specific standard deviation.

Yes, I see that it isn't really standard errors, but at the same time, somehow quantifying how the distributions of substitutions is manipulating the error range is very useful.

If you assume a distribution of players (from, say, ASPM's distribution historically for an average season), and then assign randomly the APM, but use actual minutes and substitutions, and then repeat, say, 100 times,.... wouldn't you have just tallied enough info to infer a rough standard error for each player?

[J.E.]

...yes, but the level of RMSE would not be determined by player inconsistency but by RAPM bias for their specific lineups/minutes. Yes?

Player skill is consistent over the entire simulation but what a player actually produces is still very much influenced by chance, due to me simulating the games.
I could, alternatively, not simulate if and how much gets scored but simply make it so that a +5 offensive unit scores (averagePPP+0.05) on *every* possession, while still assigning random player ratings.

If you assume a distribution of players (from, say, ASPM's distribution historically for an average season), and then assign randomly the APM, but use actual minutes and substitutions, and then repeat, say, 100 times

If by "assigning APM" you meant assigning player ratings, then that's what I did.

wouldn't you have just tallied enough info to infer a rough standard error for each player?

I suppose we get a rough confidence measure, but can we call it "standard error"?

[J.E.]

New euroleague RAPM

http://stats-for-the-nba.appspot.com/PB ... gue3y.html (3year)
http://stats-for-the-nba.appspot.com/PB ... gue2y.html (2year)
http://stats-for-the-nba.appspot.com/PB ... gue12.html (only this year)

Data provided by http://www.in-the-game.org/

Rubio still stands out in terms of being one of few that are rated well despite not being >27 years old. Sasha Kaun is 26, OKC has his rights. Sergio Rodriguez is also rated high.
Mirotic, rated at +2, was drafted by HOU but then traded to CHI via MIN

I count ~8 (ex-)NBA players in the top 25