Why does anyone like RAPM?

[udfa]

Among the several advanced stats (RAPM, PIPM, RPM, BPM/VORP, WS) that are commonly used in argumentation on this board, RAPM seems to have particular currency for a lot of posters. I'd like to know what about RAPM gives those who are inclined to use it above other advanced stats confidence in its particular validity? Far more than RPM, BPM or WS, RAPM appears to return anomalous ratings for many of the league's most acclaimed players. Consider these from the 2018-19 regular season RAPM:

Giannis #14 (behind #6 Derrick White, #7 Tatum, #11 Bev and #13 Danuel House)
Butler #27 (#15 Bertans, #16 Neto, #24 Ego, #25 Nene, #26 Patty Mills)
Harden #35 (#28 Lamb, #29 Baynes, #33 Looney)
Jokic #40 (#38 Theis)
Mitchell #65 (#42 Kleber, #45 Ish Smith, #49 Broekhoff, #55 Aaron Holliday, #61 Ojeleye)
LMA #71 (#67 Carroll, #69 Hayward, #70 Royce O'Neal)
LeBron #77 (#75 Gerald Green, #76 Iwundu)
KAT #81 (#79 Treveon Graham, #80 Darius Miller)
AD #86 (#83 Nwaba, #84 JR Smith, #85 Pinson)
Kawhi #97 (#87 Bjelica, #89 Korver, #91 DiVincenzo, #93 Caruso, #94 Yogi Farrell)
Gobert #101 (#98 Olynyk, #100 Dotson)
DeRozan #177 (#102 Wannamaker, #142 Norman Powell, #150 Rivers, #176 Svi)
Westbrook #187 (#184 Kaminski, #185 Jabari Parker)
Beal #260 (#189 Temple, #204 Crabbe, #218 Wes Matthews, #240 Ian Clark, #241 McDermott)
DLo #302 (#266 DSJ, #267 Bullock, #271 Korkmaz, #279 Tim Frazier, #298 Evan Turner)
Doncic #314 (#304 Wiggins, #305 Joe Harris, #311 Courtney Lee)
Blake #324 (#316 Marvin Williams, #319 Saric, #321 Tristan Thompson)
Booker #331 (#325 Gary Clark, #326 Shumpert)
Simmons #340 (#332 Mahinmi, #336 DJ Wilson, #339 Jevon Carter)

Note: RAPM ranks from the Github source list the same player once for each team he appears on plus the "total" all as separate entries. I have removed all of the duplicate entries so each player is only listed once.

Shouldn't such a wealth of results that are wildly at variance with common sense (far, far more so than RPM for instance) be totally discrediting for the accuracy of any statistic?

Consider the top-5 by position for RAPM vs RPM. Which stat's top-5s more closely conform to reality in your mind?

PG
RAPM: #1 Steph, #2 Jrue, #3 White, #4 Lowry, #5 Bev
RPM: #1 Steph, #2 Dame, #3 CP3, #4 Lowry, #5 Jrue

SG
RAPM: #1 Green, #2 House, #3 Middleton, #4 Ego, #5 Lamb
RPM: #1 Harden, #2 Green, #3 Mitchell, #4 Richardson, #5 Lamb

SF
RAPM: #1 KD, #2 PG, #3 Tatum, #4 Butler, #5 Ingles
RPM: #1 PG, #2 LBJ, #3 KD, #4 Butler, #5 Gallo

PF
RAPM: #1 Giannis, #2 Bertans, #3 Siakam, #4 Millsap, #5 Draymond
RPM: #1 Giannis, #2 AD, #3 Millsap, #4 Siakam, #5 Looney

C
RAPM: #1 Embiid, #2 Horford, #3 Nurkic, #4 Drummond, #5 Nene
RPM: #1 Jokic, #2 Embiid, #3 Vucevic, #4 Gobert, #5 Horford

[giberish]

The big advantage of RAPM is that it doesn't have any box-score input. It's all about winning and losing (or at least team scoring more or less). Box-score stats are inherently limited by what doesn't make it into a box-score. By not including them RAPM is 'pure'.

The big disadvantage of RAPM is that it's noisy. It takes a long time for the signal to clearly show through the noise. Even a full season doesn't have great signal/noise and smaller samples There are two ways of dealing with this. One is to just take more data, either looking at multi-year samples or to use previous years data in place of just using 0's as the default value in the calculations (misleadingly called a 'prior'). The other is to mix it in some way with box score data, which is less noisy but also less accurate (at least as far as on-court team scoring/defending impact). So there's an accuracy/precision tradeoff.

[ShotCreator]

PI RAPM is the most informative to me. The methodology makes more sense than any other metric out there.


I haven’t seen it in years however. I think JE is the last one to make it.

That is obviously NPI RAPM.

Even that 5 year playoff RAPM that got released was NPI.

[Joao Saraiva]

It has some use. However I feel like some people put too much weight on it.

It's a team sport and that stat in particular is highly influenced by team results. Situations that create RAPM are not equal for every player - the rotations they're in influence it, the good and bad supporting casts do it too...

I personally don't use it. WS/48 is already a stat that is influenced by being in a winning situation or not. But I guess you can use RAPM too. Just don't use it alone. Like any stat, alone, it will produce awkward results.

[trex_8063]

gibberish put in nicely in post #2, highlighting both its strengths and weaknesses.

Another big take-away is in Joao's post (#4): ANY stat used in isolation will yield some questionable results. Anyone who is using ONE stat to the exclusion of others in evaluating players is doing it wrong. Anyone who is taking one stat or even a collection of stats at face-value without understanding their different philosophies is also in error.

The other "advanced" box-based metrics do a relatively poor [or at best mediocre] job of capturing impact; even the ones that are curved toward team results (like WS and BPM), because they still carry their box-based biases [say that five times really fast]. An example is WS or WS/48 often placing too much value on low(ish) volume/high efficiency scoring and probably not even on high-volume/mediocre efficiency scoring (in many instances, anyway).
Also, because there's a finite amount of indicators it can get from the box, it frequently ascribes value inappropriately......such as emphasis on defensive rebounds for defensive impact (because there are precious few other defensive box stats), which consequently makes guys like Charles Barkley, Carlos Boozer, and Amar'e Stoudemire look like solid defenders based on "advanced stats".

And where the well of info that can be obtained from the box runs dry, you might say these same advanced stats otherwise sort of spread the credit around the roster evenly (a general "team curving").


RAPM not only takes a more "pure" look at impact (no box inputs), it also attempts to tease out which guys on the roster are actually responsible for it (so you can get guys on winning teams with negative RAPM's, and vice versa). But as indicated above, though, this is noisy business and takes a long time to gel.

The sample you're citing is NPI (non-prior informed), which tends to be noisier. I prefer PI RAPM in most instances (though I don't know of a PI sample for '18 and '19), though there are rare instances where NPI may be just as good or better (e.g. a player who's made great strides forward in the last year or two; a player who's declined significantly in the last year or two; a rookie).

In addition to being NPI, what you're citing is also a source I've come to be skeptical of. This gitlab author has offered some very questionable results here and there.......the one that really threw me was from one of the Nash Suns teams ('10???), a season in which Nash was a very close 2nd on the team [to Channing Frye] in raw on/off [both rs and playoffs]----with no one else particularly close----while also being 2nd on the team in minutes; and yet this author had his RAPM dead last on the roster (literally like 14th/14, or whatever).....which I think is damn near a mathematical impossibility given the on/off spread.
Results like that have made me worry about occasional carelessness and/or wonky inputs.

In general, his numbers have more "anomalous" [to use your word] results than others.

fwiw, here's a source for '18 rs NPI RAPM, and here is a source for '19 rs NPI RAPM which I trust more (I think the author of these is actually Englemann [the creator of RAPM], if I'm not mistaken).

The '19 results are somewhat more in keeping with expectation. Although guys like Jimmy Butler, Nikola Jokic, and KAT fall marginally further in the ranks, Giannis is 6th instead of 14th (#1-5 are Danny Green, Kevin Durant, Paul George, Steph Curry, and Jrue Holiday), James Harden is 32nd instead of 35th, Rudy Gobert is 49th instead of 101st, Donovan Mitchell is 35th instead of 65th, Anthony Davis is 39th instead of 86th, Westbrook is 105th instead of 187th, Bradley Beal is 107th instead of 260th, etc.

[udfa]

The big advantage of RAPM is that it doesn't have any box-score input. It's all about winning and losing (or at least team scoring more or less). Box-score stats are inherently limited by what doesn't make it into a box-score. By not including them RAPM is 'pure'.

The big disadvantage of RAPM is that it's noisy. It takes a long time for the signal to clearly show through the noise. Even a full season doesn't have great signal/noise and smaller samples There are two ways of dealing with this. One is to just take more data, either looking at multi-year samples or to use previous years data in place of just using 0's as the default value in the calculations (misleadingly called a 'prior'). The other is to mix it in some way with box score data, which is less noisy but also less accurate (at least as far as on-court team scoring/defending impact). So there's an accuracy/precision tradeoff.

Given that a player's conditions can change substantially each season, doesn't the usage of past seasons to enlarge the sample size introduce its own heavily distorting effect into the rating of a player's current season performance?

[udfa]

gibberish put in nicely in post #2, highlighting both its strengths and weaknesses.

Another big take-away is in Joao's post (#4): ANY stat used in isolation will yield some questionable results. Anyone who is using ONE stat to the exclusion of others in evaluating players is doing it wrong. Anyone who is taking one stat or even a collection of stats at face-value without understanding their different philosophies is also in error.

The other "advanced" box-based metrics do a relatively poor [or at best mediocre] job of capturing impact; even the ones that are curved toward team results (like WS and BPM), because they still carry their box-based biases [say that five times really fast]. An example is WS or WS/48 often placing too much value on low(ish) volume/high efficiency scoring and probably not even on high-volume/mediocre efficiency scoring (in many instances, anyway).
Also, because there's a finite amount of indicators it can get from the box, it frequently ascribes value inappropriately......such as emphasis on defensive rebounds for defensive impact (because there are precious few other defensive box stats), which consequently makes guys like Charles Barkley, Carlos Boozer, and Amar'e Stoudemire look like solid defenders based on "advanced stats".

And where the well of info that can be obtained from the box runs dry, you might say these same advanced stats otherwise sort of spread the credit around the roster evenly (a general "team curving").


RAPM not only takes a more "pure" look at impact (no box inputs), it also attempts to tease out which guys on the roster are actually responsible for it (so you can get guys on winning teams with negative RAPM's, and vice versa). But as indicated above, though, this is noisy business and takes a long time to gel.

The sample you're citing is NPI (non-prior informed), which tends to be noisier. I prefer PI RAPM in most instances (though I don't know of a PI sample for '18 and '19), though there are rare instances where NPI may be just as good or better (e.g. a player who's made great strides forward in the last year or two; a player who's declined significantly in the last year or two; a rookie).

In addition to being NPI, what you're citing is also a source I've come to be skeptical of. This gitlab author has offered some very questionable results here and there.......the one that really threw me was from one of the Nash Suns teams ('10???), a season in which Nash was a very close 2nd on the team [to Channing Frye] in raw on/off [both rs and playoffs]----with no one else particularly close----while also being 2nd on the team in minutes; and yet this author had his RAPM dead last on the roster (literally like 14th/14, or whatever).....which I think is damn near a mathematical impossibility given the on/off spread.
Results like that have made me worry about occasional carelessness and/or wonky inputs.

In general, his numbers have more "anomalous" [to use your word] results than others.

fwiw, here's a source for '18 rs NPI RAPM, and here is a source for '19 rs NPI RAPM which I trust more (I think the author of these is actually Englemann [the creator of RAPM], if I'm not mistaken).

The '19 results are somewhat more in keeping with expectation. Although guys like Jimmy Butler, Nikola Jokic, and KAT fall marginally further in the ranks, Giannis is 6th instead of 14th (#1-5 are Danny Green, Kevin Durant, Paul George, Steph Curry, and Jrue Holiday), James Harden is 32nd instead of 35th, Rudy Gobert is 49th instead of 101st, Donovan Mitchell is 35th instead of 65th, Anthony Davis is 39th instead of 86th, Westbrook is 105th instead of 187th, Bradley Beal is 107th instead of 260th, etc.

Thanks for the better NPI RAPM link.

I agree that any single stat is bound to produce questionable results, but isn't the degree of doubtfulness the issue here? The reason I asked about RAPM is because the degree of disagreement between its ratings and the perceived rankings of players seem to radically larger than other advanced stats (RPM for instance). The NPI RAPM ratings you linked are less drastic in their divergence from consensus, but the difference is still notably large in comparison. If someone without any bonafides attempted to introduce their own stat that had LeBron, Harden, Butler, AD, Jokic and Gobert all ranked between 29th and 49th, wouldn't it be completely discreditable just on its face? When the frequency of these wild anomalies is unusually large, doesn't it indicate that the metric in question does a very poor job of evaluating the impact these players are actually having on their team's performance?

[edited to refer correctly to the link as NPI rather than PI]

[trex_8063]

Thanks for the PI RAPM link.

I agree that any single stat is bound to produce questionable results, but isn't the degree of doubtfulness the issue here? The reason I asked about RAPM is because the degree of disagreement between its ratings and the perceived rankings of players seem to radically larger than other advanced stats (RPM for instance). The PI RAPM ratings are less drastic in their divergence from consensus, but the difference is still notably large in comparison. If someone without any bonafides attempted to introduce their own stat that had LeBron, Harden, Butler, AD, Jokic and Gobert all ranked between 29th and 49th, wouldn't it be completely discreditable just on its face? When the frequency of these wild anomalies is unusually large, doesn't it indicate that the metric in question does a very poor job of evaluating the impact these players are actually having on their team's performance?

To be clear: those links I provided were not [to my knowledge] PI RAPM. They're still NPI, but from [imo] a better/more reputable source.

Another thing to be clear on when looking at RAPM---and strictly speaking, I suppose this is true of every metric, but perhaps none more than RAPM---it's not a measure of "player goodness". It's a measure of player goodness + fit/chemistry + utilization.

And another thing worth mulling over when looking at RAPM is what you might call scarcity of talents, or how replaceable someone is......
A good 3&D wing [like Danny Green] might provide a lot of positive impact (and have a high RAPM)......but he's also relatively replaceable (in that nearly every team has at least 1-2 decent 3&D role players on their roster). Not that all 3&D role players are created equally; some are better than others [and imo Danny Green is one of the best of them presently], but you get what I'm saying: there are a lot of them.

By comparison, take someone like James Harden, Giannis Antetokounmpo, Lebron James, or Steph Curry.......they might not always have measurable greater impact than each and every year than a one of the best [and best-utilized] role players, but I think it goes without saying that the special blend of [super-elite] talents they bring to the table are not so easily replaced. You can't just find one of them around every corner (obviously; most teams don't have someone like that at all), and some of the above are downright singular in the league.


One last thing to consider when looking at rs-only numbers (as we have been), in this era where stars are consistently coasting during the rs: those rs RAPM's are not fully representative of what they're capable of (like....in the playoffs).

'19 Kawhi's rs RAPM is a great example. The only team I watched more than the Raptors last year was the Jazz; and tbh, I'd barely have batted an eye if I saw Kawhi had a neutral RAPM. He coasted on defense all year until the playoffs (got All-D honors pretty much based on prior rep alone), and on offense he can score like hell, but was/is a bit of a ball-stopper who sometimes interrupts their flow.

Now in the playoffs, that latter feature didn't really matter all that much, because game-planning tough defenses tend to interrupt offensive flow regardless, and thus teams really need a reliable go-to isolation scorer they can fall back on in the playoffs......this is often what separates a good regular season team from a good playoff team, and it's what the Raptors have been lacking the last few seasons [and why they couldn't previously "get over the hump"].
But in the rs, when [usually] the offense is flowing and other players are able to get theirs, it's less necessary. And consequently, we didn't see Kawhi providing much lift offensively in the rs. Combine that with him coasting defensively [again: IN THE REGULAR SEASON], and I don't think one should be shocked to see an utterly pedestrian RAPM for the rs.

tbh, I'd have been surprised [and a little skeptical] if he'd had a monster rs RAPM.

[Dr Positivity]

You have to use multiple years to get a better picture from RAPM.

RPM basically took the concept of RAPM and made a version using box scores to help predict better. But there is some use in just a pure +/- calculation. Conceptually RAPM makes perfect sense (using other players +/- in combination with the players +/-) while the weight of +/- and boxscore in RPM is only really determined by what predicts the best

[Doctor MJ]

Good posts here, and I'll just add:

If I'm using only one stat, I'd prefer a stat in the XRAPM family (such as RPM or now, RAPTOR). It does not make sense to ignore the box score.

But if I'm going in depth, I have a better ability to understand what's driving the +/- data if I can see it without it previously being merged into a black box, a la XRAPM stats.

The best of all worlds is to have access to it all of course.

[Doctor MJ]

It has some use. However I feel like some people put too much weight on it.

It's a team sport and that stat in particular is highly influenced by team results. Situations that create RAPM are not equal for every player - the rotations they're in influence it, the good and bad supporting casts do it too...

I think the real truth of things is that because it's a flow-based field sport, impact itself is "noisy", and hence we should expect any stat tied to impact to be pretty noisy.

The box score stats - especially production/volume stats - by contrast are actually artificially stable. The same players tend to have the same primacy from game to game regardless of matchup, health, or groove, and the players themselves apply artifice to "get their numbers" (which of course comes at the cost of not acting in accordance of the opportunity presented on a given play).

None of this means we shouldn't be cautious about drawing conclusions from noisy +/- stats, but I think people tend to overrate stat reliability and not really think about validity.

[Cavsfansince84]

I feel like what we need is a metric which takes about 5-6 different metrics and combines them into one. It wouldn't be perfect of course but it would be attempting to merge all these different ways of trying to measure a player's impact on the court. For instance, take each metric and use standard deviation to measure a player's value above the mean(be it +1.2, -1.6 or +3.3) and then just add them all together.

[Jaivl]

I feel like what we need is a metric which takes about 5-6 different metrics and combines them into one. It wouldn't be perfect of course but it would be attempting to merge all these different ways of trying to measure a player's impact on the court. For instance, take each metric and use standard deviation to measure a player's value above the mean(be it +1.2, -1.6 or +3.3) and then just add them all together.

That's, in a sense, pretty much what RAPTOR does.

Since I don't have much interest in prediction, I prefer the use of (multiple) explanatory metrics, plus using RAPM in a "big picture"/range sort of way, to gauge the real impact of those multiple smaller observations.

[Doctor MJ]

I feel like what we need is a metric which takes about 5-6 different metrics and combines them into one. It wouldn't be perfect of course but it would be attempting to merge all these different ways of trying to measure a player's impact on the court. For instance, take each metric and use standard deviation to measure a player's value above the mean(be it +1.2, -1.6 or +3.3) and then just add them all together.

And I feel that while this is useful, a trained human brain is the best tool we're probably ever going to have for this. As much as I love all the progress that's been made in statistics, every single metric we have that aims to actually define a characteristic we observe always ends up falling short of the nuance of what we have in our mind. A shot made is a shot made, easy to turn into a number, but what does that say about the player that he made that shot? A number is a one-dimensional entity aiming to approximate an n-dimensional truth.

I don't think we're ever going to be able to do anything with supervised learning that will let us actually capture the true goodness of a player.

On the other hand, I have no doubt we'll be able to design algorithms that will be able to outbet human bettors at a certain point - you don't need to understand everything even close to perfectly to beat humans at a prediction game - and some would argument that this has been the true acid test basketball statisticians are aiming toward.

[eminence]

On the other hand, I have no doubt we'll be able to design algorithms that will be able to outbet human bettors at a certain point - you don't need to understand everything even close to perfectly to beat humans at a prediction game - and some would argument that this has been the true acid test basketball statisticians are aiming toward.

Still hanging with the fancy computers for the last couple of seasons

[Cavsfansince84]

And I feel that while this is useful, a trained human brain is the best tool we're probably ever going to have for this. As much as I love all the progress that's been made in statistics, every single metric we have that aims to actually define a characteristic we observe always ends up falling short of the nuance of what we have in our mind. A shot made is a shot made, easy to turn into a number, but what does that say about the player that he made that shot? A number is a one-dimensional entity aiming to approximate an n-dimensional truth.

I don't think we're ever going to be able to do anything with supervised learning that will let us actually capture the true goodness of a player.

On the other hand, I have no doubt we'll be able to design algorithms that will be able to outbet human bettors at a certain point - you don't need to understand everything even close to perfectly to beat humans at a prediction game - and some would argument that this has been the true acid test basketball statisticians are aiming toward.

I don't think what I suggested would do that either. Just as far as using numbers to try and measure players I think it would do more than using any of the single metrics are doing by themselves. Numbers can shed a light on some things while obviously watching games is another important element to being a fan or self styled expert on nba history and the skills of its best players over the years.

[udfa]

Thanks for the PI RAPM link.

I agree that any single stat is bound to produce questionable results, but isn't the degree of doubtfulness the issue here? The reason I asked about RAPM is because the degree of disagreement between its ratings and the perceived rankings of players seem to radically larger than other advanced stats (RPM for instance). The PI RAPM ratings are less drastic in their divergence from consensus, but the difference is still notably large in comparison. If someone without any bonafides attempted to introduce their own stat that had LeBron, Harden, Butler, AD, Jokic and Gobert all ranked between 29th and 49th, wouldn't it be completely discreditable just on its face? When the frequency of these wild anomalies is unusually large, doesn't it indicate that the metric in question does a very poor job of evaluating the impact these players are actually having on their team's performance?

To be clear: those links I provided were not [to my knowledge] PI RAPM. They're still NPI, but from [imo] a better/more reputable source.

Another thing to be clear on when looking at RAPM---and strictly speaking, I suppose this is true of every metric, but perhaps none more than RAPM---it's not a measure of "player goodness". It's a measure of player goodness + fit/chemistry + utilization.

And another thing worth mulling over when looking at RAPM is what you might call scarcity of talents, or how replaceable someone is......
A good 3&D wing [like Danny Green] might provide a lot of positive impact (and have a high RAPM)......but he's also relatively replaceable (in that nearly every team has at least 1-2 decent 3&D role players on their roster). Not that all 3&D role players are created equally; some are better than others [and imo Danny Green is one of the best of them presently], but you get what I'm saying: there are a lot of them.

By comparison, take someone like James Harden, Giannis Antetokounmpo, Lebron James, or Steph Curry.......they might not always have measurable greater impact than each and every year than a one of the best [and best-utilized] role players, but I think it goes without saying that the special blend of [super-elite] talents they bring to the table are not so easily replaced. You can't just find one of them around every corner (obviously; most teams don't have someone like that at all), and some of the above are downright singular in the league.


One last thing to consider when looking at rs-only numbers (as we have been), in this era where stars are consistently coasting during the rs: those rs RAPM's are not fully representative of what they're capable of (like....in the playoffs).

'19 Kawhi's rs RAPM is a great example. The only team I watched more than the Raptors last year was the Jazz; and tbh, I'd barely have batted an eye if I saw Kawhi had a neutral RAPM. He coasted on defense all year until the playoffs (got All-D honors pretty much based on prior rep alone), and on offense he can score like hell, but was/is a bit of a ball-stopper who sometimes interrupts their flow.

Now in the playoffs, that latter feature didn't really matter all that much, because game-planning tough defenses tend to interrupt offensive flow regardless, and thus teams really need a reliable go-to isolation scorer they can fall back on in the playoffs......this is often what separates a good regular season team from a good playoff team, and it's what the Raptors have been lacking the last few seasons [and why they couldn't previously "get over the hump"].
But in the rs, when [usually] the offense is flowing and other players are able to get theirs, it's less necessary. And consequently, we didn't see Kawhi providing much lift offensively in the rs. Combine that with him coasting defensively [again: IN THE REGULAR SEASON], and I don't think one should be shocked to see an utterly pedestrian RAPM for the rs.

tbh, I'd have been surprised [and a little skeptical] if he'd had a monster rs RAPM.

Thanks for the clarification on the NPI link.

If anyone finds a good PI RAPM source perhaps they might consider sharing it here. I know of a source for PIPM -- does anyone know how that compares to RAPM in methodology?

https://www.bball-index.com/18-pipm/

[udfa]

Good posts here, and I'll just add:

If I'm using only one stat, I'd prefer a stat in the XRAPM family (such as RPM or now, RAPTOR). It does not make sense to ignore the box score.

But if I'm going in depth, I have a better ability to understand what's driving the +/- data if I can see it without it previously being merged into a black box, a la XRAPM stats.

The best of all worlds is to have access to it all of course.

If you don't mind my asking, what source do you use for researching RAPTOR stats? I found a bunch of RAPTOR pages on 538 but couldn't find my way there to where historical seasonal ratings tables were.

[GeorgeMarcus]

If I'm using only one stat, I'd prefer a stat in the XRAPM family (such as RPM or now, RAPTOR). It does not make sense to ignore the box score.

Personally I'm not a fan of box priors, or any priors that don't stem from the broader implications of "impact". To me it dilutes the fundamental purpose of RAPM. Not to say box stats should be ignored (they shouldn't), but rather used to compliment/contextualize impact data.

That said, you're a thoughtful poster and I like to keep an open mind. What would you say is the reason for your preference?

[Doctor MJ]

Good posts here, and I'll just add:

If I'm using only one stat, I'd prefer a stat in the XRAPM family (such as RPM or now, RAPTOR). It does not make sense to ignore the box score.

But if I'm going in depth, I have a better ability to understand what's driving the +/- data if I can see it without it previously being merged into a black box, a la XRAPM stats.

The best of all worlds is to have access to it all of course.

If you don't mind my asking, what source do you use for researching RAPTOR stats? I found a bunch of RAPTOR pages on 538 but couldn't find my way there to where historical seasonal ratings tables were.

https://github.com/fivethirtyeight/data/tree/master/nba-raptor

Don't need to do anything that elaborate. Just click on their csv files, copy the text, paste into a spreadsheet, use Text-to-Columns separating by commas, and you're set.