I am an avid basketball fan from Germany who is not going to become a contributor on this site due to the fact that I have neglected math at school, but I’m often visiting your forum because of the intriguing models, ideas and discussions – even if I cannot fully grasp most of the advanced math stuff.
I am writing this because I need your help. The basketball discussion in Germany is generally not very educated because it does not rely much on statistics as it does on eye test and points per game. Just to give you an idea of how bad it is – recently, the official German contractor for NBA content took a poll on the 'most underpaid players by position' with Monta Ellis leading the way at the Shooting Guard position....
Well, I am a contributing writer on a German basketball blog that has a nice followership, and tries to go against the stream of the NBA ‘narrative’ by promoting educated discussion and the usage of advanced statistics like USG%, TS% an so on.
I have become interested in Advanced Plus/Minus Statistics, especially RAPM because of its defensive potential. I have been thinking about realizing an introductory article about (R)APM for our website that comments on the models strengths, field of operation, limitations, future potential, etc. I have googled a number of posts regarding the ridge regression (here, sportskeptic, godismyjudgeok.com, realgmforums, ..) that is used to erase the problem of co-linearity and think that I have a decent idea of how it works (for somebody who is just plain dumb in the mathematical field) but I still have a few questions. I would be very grateful if somebody could answer those and I will put references to this forum, and several stat-databases like the one of J.E., Daniel Myers or Evan Zamir (if that’s okay for them – if you have more databases that you think should be featuredd just drop a link).
This is how I understand the ridge regression:
Usual APM includes the problem of co-linearity. Some players work in pairs or trios all the time and it’s hard to say which player is 'more' responsible for those lineups results. The problem with APM is that it rates all players equal. If Roy Hibbert, Paul George and D.J. Augustin play in a lineup, APM gives them a neutral ‘zero’ before they start logging any minutes. The trick with RAPM is that it gives each player a certain range of expected performance. Thus, Hibbert and George would ‘go’ into those lineups with a bigger defensive potential right away. This range (or prior?) depends on past offensive and defensive performance.
Is that more or less correct?
I know that the use of advanced statistics, especially one-number-stats like RAPM is a controversial issue with some conservative NBA folks. I want to be precise in my article but also be as clear and comprehensible as possible – the average German basketball fan is less ‘prepared’ for such an approach as the average US-American fan, and I believe that making it too complicated would just alienate our readership for this topic. I know that this might be a little offensive towards the scientific method/significance behind the approach, but as I said - this is supposed to be an introductory article that hopefully will open a few people eyes towards more detalined and informed basketball evaluation.
Oh man, I'm sorry for writing so much.
Here are some questions that I have:#1 – Why exactly is the ridge regression a useful method for basketball? Is it because the regression has proven to be successful with similar non-basketball problems where you have to resolve which element is mainly responsible for a certain effect?
#2 – While usual Plus Minus says “With player A on the court, team 1 is X points better/worse than with another player of team 1 on the field”, (R)APM says “With player A in his role on the court, team 1 is X points better/worse than with an average NBA player (0) on the field”. Is this assumption correct? Is that why it is a regression instead of just a point differential tracker? This would imply that it is also not a strict ranking metric like WP claims to be, but more a efficiency ranking based on role (at least on offense)?
#3 – As I understood, RAPM seems to be very flexible in its scope (primer?). You can include all sorts of data – box score numbers, advanced stats, shot location data, and so on. Is it possible to build RAPM-models for all kinds of different, specific purposes? For example one model for 3-and-D players, another model for bigs who are exceptional at passing, et cetera. Do you know whether the coach evaluation system which the Mavs build to find Rick Carlisle was APM-based?
#4 – What do you think can the future bring for a statistic like RAPM? I remember Kirk Goldsberry writing in his paper about post defense that his approach and findings might be a first, small step towards a bigger development in rating big men defense. Do you see something similar with RAPM now that so much of the chaotic, fast paced NBA action can be tracked more detailed by playbyplay-data and camera tracking? Any idea how that could look like?
#5 - Do you know of anybody working for an NBA team using specifically RAPM?
Phew, I hope some of you find the time to help me out a little.