APBRmetrics

I have been working to develop a revised RAPM prior to use for long-term (i.e. 3+ years) RAPM datasets.

I demonstrated some years ago that MPG was a very good prior. I use regressed MPG, using 4 games of 0 MPG, to adjust players who played very few minutes.

What I found is that MPG has a consistent positive and linear relationship with offensive production. However, that relationship breaks down for stars, who are "capped" on minutes but their production can be far above the linear relationship.

On defense, I found that there is almost no relationship between MPG and defensive production. This indicates that playing time is selected primarily based on offensive production, which makes sense because the spread in offensive production appears to be significantly wider than the spread in defensive production.

In addition--the "bench"/0 MPG player intercept is well below zero for offensive players, in the range of -4 pts/100 poss, while the defensive intercept is very close to 0.

Recently I've explored a 2D position spectrum, with results here: https://public.tableau.com/views/ofNBAS ... zHome=no#1

I applied those positions to a revision of this RAPM prior. I found that the offensive creation role had a mild positive effect on offensive production, only +1.5 from a minimum creation role to a max creation role. I found the "size" position had a +2.5 spread on defensive production from minimum size to maximum size.

Finally, to handle the "star" issues with the MPG linear relationship with production, I pulled in NBA awards data. Yes, I know this could be faulty and could overrate Kobe, but it proved useful nonetheless. Here are the values I found for the awards:
Note: the DPOY and MVP use the actual award shares rather than a boolean True/False. So a unanimous MVP or DPOY gets that full credit. Also, these are additive. The MVP usually also is All NBA 1st team and All Star.

The end result is a prior that has a 0.87 offensive correlation with a long term RAPM dataset and a 0.59 defensive correlation with the RAPM dataset.

Here are the full values I'm using: The coefficient on ReMPG for each team is calculated separately after applying all of the above values, such that the total team production sums to the team's adjusted offensive or defensive efficiency.

EDIT: See post below for updated values.

Interesting fitting information.


So BPM, which has always been built on NBA exclusive RAPM, has the RAPM prior modified by NBA minutes behavior and NBA exclusive award information, but won't use NBA on the court shot defense data or NBA tracking data, because "it is being used for college and other leagues"?...

There is a contradiction there.

Crow wrote: ↑Thu Jan 25, 2024 2:48 pm Interesting fitting information.


So BPM, which has always been built on NBA exclusive RAPM, has the RAPM prior modified by NBA exclusive award information, but won't use NBA on the court shot defense data or NBA tracking data, because "it is being used for college and other leagues"?...

The goal of an RAPM basis is to be as accurate as possible to what the NBA player has actually accomplished, with the caveat that we want to limit box-score data in the prior as much as possible to avoid data leakage (overrating box score data because the prior is biased towards box score data.) The RAPM is still dominated by the lineup data because we are using multi year RAPM--at least 3 years and typically 4 years.

BPM itself is then regressed onto this "best possible RAPM basis". BPM only uses box-score data, as previously discussed, with the goal of building as robust and flexible a metric as possible, to best capture the actual value of box score statistics.

NBA data based BPM used for college players... an advantage for assessing NBA potential or not?

Accuracy for assessing individual college performance (vs. some truly non-context based model or one exclusive to college data)?

Accuracy of current BPM in matching team college performance (vs. alternative models such as PER, WinShares and PRPG!)?

Anybody want to do neutral / independent studies?

DARKO coefficients give different values to at rim fga2 and fgm2 and nonrim2s.

Elsewhere there is consternation about that.

Is that an element of BPM model, current or future? I didnt look it up right now. Regardless of whether it is or not, anyone have comments on finding different values for these? I see reasons for and against.

Crow wrote: ↑Thu Jan 25, 2024 11:34 pm DARKO coefficients give different values to at rim fga2 and fgm2 and nonrim2s.

Elsewhere there is consternation about that.

Is that an element of BPM model, current or future? I didnt look it up right now. Regardless of whether it is or not, anyone have comments on finding different values for these? I see reasons for and against.

BPM will not because it's not in the box score, but I absolutely support different values for different shot attempts.

A shot attempt at the rim, self-created, has a lot different value to the team then a mid-range field goal attempt, off a pass. Who gets credit for generating the attempt? What is the expected value (expected field goal percentage) of the attempt? And then, the shooter themselves gets credit for the making or not of the shot.

I was working with PBPstats.com data recently looking at a lot of these things.

[edited]

I was looking at some old threads from 2009, and found some discussion about RAPM priors there. Crow, you had some good ideas then...

This was when Joe Sill first unveiled RAPM.
Page 1: http://godismyjudgeok.com/DStats/APBRme ... =2386.html
Page 2: http://godismyjudgeok.com/DStats/APBRme ... 97%20.html

I didn't remember the specific comments from nearly 15 years ago but it was worth re-seeing the comment that an average impact profile of a player may not get at their range of impact based on certain lineup context and that understanding a player as a small set of impact profiles might be desirable.

It could possibly be standardized (as a start) as with 3 or 4 starters vs. less (and maybe with none separately).


These pages show high effort analysis by more than 1-2 persons. Those were the days.

It would be worth re-visiting more old threads to remember and maybe try to extend. I don't know if I will.

After using the initial prior from the first post for a collection of 3-year RAPM calculations, I have refit the values to better fit the output from those analyses.

Here are the revised coefficients for this prior: The coefficient on ReMPG for each team is calculated separately after applying all of the above values, such that the total team production sums to the team's adjusted offensive or defensive efficiency.

These adjustments slightly increase the team role/position adjustments, a small adjustment to make role/position adjustment increase with more playing time, a decrease to all-star value, increase MVP value, and reduce 0 MPG player value somewhat. Overall it's a minor adjustment, with the prior going up a maximum of around 2.0 for a few MVP level players and no player being reduced more than around -0.7.

The highest 3 year average priors in the dataset belong to Duncan, Giannis, and LeBron, and are in the +11.5 range... which is approximately what their prior-informed RAPM values are for those periods as well. They have large samples so the data is generally confirming.

The star player that appears the most overrated by the prior is 1997-99 Karl Malone, who RAPM thinks was around a +5 player but gets a +10 prior. The most underrated stars by the prior are Manu Ginobili (few minutes, few accolades) and early career Nikola Jokic, who had great RAPM values.