APBRmetrics

If SWagR wins are based on average RAPM points, are the positive SWagRs overstating impact given that some games players deliver more impact (more than enough) and other games they are lesser or negative rather than metronome average positive? I can't see the SWagR definition at the moment on my phone to see if this is accounted for or an issue.

If it is an issue, then guys who are mild positive or even negative might help win more games than their average indicates by being better say 30% of the games. Maybe there is a simplistic assumption that the game to game impact curves of all players are pretty similar even across quality levels. Possible but I doubt it. Some are likely more consistent than others, some more volatile one way or both. It would be interesting to measure consistency and volatility for teams / teammates across team opponent types. Possibly home and away too. Team consistency might depend on lucky or thoughtful blending of player consistency / volatility. Probably should study teammate interactions too.

If you had the time and motivation this should be taken to the RAPM factor level to look at interactions and also just to see the RAPM factors as distribution curves of at least several cluster points instead of as a single point average. Players of similar average impact could have very different curves and differing impacts on different teams depending on teammate mix and team quality. So when a player shifts teams their RAPM or RPM could change or even if it doesn't average pointwise, it could in terms of win impact.

There has been talk of RAPM based on change in win probability play to play instead of points. That is another technique worth using.

2015-16 RAPM https://docs.google.com/spreadsheets/d/ ... eMNZI/edit
I don't think playoffs data gets used in any of these, but I could be mistaken.

Hi JE, I'm worried that I haven't been clear enough about what exactly I'm doing and why I need pure APM, so let me try again.

My goal is to create a statistical plus/minus (SPM) model by using box score stats to predict APM. As I envision it, my model will assume that a player's APM is a normally distributed random variate with mean equal to their true plus/minus value (and uncertainty a function of minutes played). To the extent that this is an accurate assumption, a model that predicts APM is a model that predicts true plus/minus value, which is exactly what I want.

I cannot use this simple assumption if I use RAPM instead of APM because due to the zero prior, a player's RAPM cannot be described as a random variate with mean equal to their true plus/minus. Instead, the mean is some nontrivial function of lambda, the exact value of the zero prior, minutes played, and the player's true plus/minus value (and the uncertainty would also be a function of lambda and the zero prior as well as minutes played). Therefore a model that predicts RAPM is not a model that predicts true plus/minus alone, instead if predicts the value of a complicated function of lambda, the zero prior, minutes played, and true plus/minus. Without knowing values for lambda and the zero prior, and more importantly the nature of this function, I cannot use these predictions to derive predictions for true plus/minus.

This is why I'm looking for a source of pure APM, with no zero prior. Could you provide this, or given that you have experience in making SPM, could you give some advice on how to arrive at SPM starting from RAPM?

Thanks as always!

Nathan wrote:Hi JE, I'm worried that I haven't been clear enough about what exactly I'm doing and why I need pure APM, so let me try again.

My goal is to create a statistical plus/minus (SPM) model by using box score stats to predict APM. As I envision it, my model will assume that a player's APM is a normally distributed random variate with mean equal to their true plus/minus value (and uncertainty a function of minutes played). To the extent that this is an accurate assumption, a model that predicts APM is a model that predicts true plus/minus value, which is exactly what I want.

I cannot use this simple assumption if I use RAPM instead of APM because due to the zero prior, a player's RAPM cannot be described as a random variate with mean equal to their true plus/minus. Instead, the mean is some nontrivial function of lambda, the exact value of the zero prior, minutes played, and the player's true plus/minus value (and the uncertainty would also be a function of lambda and the zero prior as well as minutes played). Therefore a model that predicts RAPM is not a model that predicts true plus/minus alone, instead if predicts the value of a complicated function of lambda, the zero prior, minutes played, and true plus/minus. Without knowing values for lambda and the zero prior, and more importantly the nature of this function, I cannot use these predictions to derive predictions for true plus/minus.

This is why I'm looking for a source of pure APM, with no zero prior. Could you provide this, or given that you have experience in making SPM, could you give some advice on how to arrive at SPM starting from RAPM?

Thanks as always!

Nathan, I understand you need an unbiased APM for your SPM regression. But given that single season APMs are extremely noisy (very nearly all noise), have you considered using a longer term APM? That's the course I chose with Box Plus/Minus. Have you read my methodology? www.basketball-reference.com/about/bpm.html Is what you want to do similar?

You make a good point, but I do have reasons for preferring single-season APM. Firstly, though the data will be less noisy if I use multi-year APM, I'll also have fewer data points (assuming the samples don't overlap, and if they do overlap, then my data points are not independent). Therefore I'm not sure I actually gain any "signal" by using multi-year APM instead of single year APM. Secondly, I'm very interested in looking at season-to-season player "evolution," and for this purpose it would be nice to have single year APM.

EDIT: I will be sure to read your methodology though; I imagine my approach will be fairly similar.

J.E. - Whenever you have a chance, would love to see final RAPM (NPI and multi-year) through the end of the finals. Thanks for all you do.

fpliii wrote:J.E. - Whenever you have a chance, would love to see final RAPM (NPI and multi-year) through the end of the finals. Thanks for all you do.

Was just about to tweet him

We're shameless

https://t.co/PANZmsWNpq
https://t.co/C3sHzpJokl

Thanks J.E.!

zzz

May I make a request for a 20-year RAPM (playoffs included) age adjusted for the regression and reverted back after. If you still don't have PBP data for 1997-2000, then 16-year should be sufficient enough. Thanks in advance.

Is there any level of RAPM from the MJ era? I noticed awhile back JE you said you didn't have the data to get this done from say 89. What is the earliest year RAPM we have?

Arman_tanzarian wrote:Is there any level of RAPM from the MJ era? I noticed awhile back JE you said you didn't have the data to get this done from say 89. What is the earliest year RAPM we have?

The earliest I've ever seen PlayByPlay data for is '97-'98. NBA teams have it, and I believe it was/is on NBA.com. The only results I've seen for that time frame are here
http://ascreamingcomesacrossthecourt.bl ... ormed.html
Look at adjacent posts for more years

Arman_tanzarian wrote:Is there any level of RAPM from the MJ era? I noticed awhile back JE you said you didn't have the data to get this done from say 89. What is the earliest year RAPM we have?

It goes back to 1996-97.
Here's 1996-97.

permaximum wrote:

Arman_tanzarian wrote:Is there any level of RAPM from the MJ era? I noticed awhile back JE you said you didn't have the data to get this done from say 89. What is the earliest year RAPM we have?

It goes back to 1996-97.
Here's 1996-97.

Reading through this thread over at realgm I was surprised to find that we have NPI RAPM for 94-96 as well. Albeit with a few difference such as not included playoff data or players who switched teams. Credit to colts18 for doing the regressions. That said, I don't know where the data to fit the models was found.

94
95
96