With the loss of nbashotcharts, what is the primary source for RS and PS RAPM?

[Throwawaytheone]

.

[lessthanjake]

I’m hoping to see others’ answers to this, but one thing I do want to note is that NBAshotcharts isn’t entirely lost. I’m fairly sure that this website just is comprised of NBAshotcharts’ 3-year RAPM data (both luck-adjusted and non-luck-adjusted): https://www.nbarapm.com/

[eminence]

https://psteve.shinyapps.io/RAPM/

[Bidofo]

https://psteve.shinyapps.io/RAPM/

1-year 2024 results
Brunson 1st in ORAPM.
OG 1st in DRAPM.
4th and 5th, respectively, in total RAPM with Hartenstein at 6th.

I vote for this one as the perfect source. Hoping for 57+ish wins next year assuming we retain OG and IH and with decent injury luck.

[eminence]

https://psteve.shinyapps.io/RAPM/

1-year 2024 results
Brunson 1st in ORAPM.
OG 1st in DRAPM.
4th and 5th, respectively, in total RAPM with Hartenstein at 6th.

I vote for this one as the perfect source. Hoping for 57+ish wins next year assuming we retain OG and IH and with decent injury luck.

Y'all were crazy good with OG last season (20-3), seems possible.

[Throwawaytheone]

.

[Throwawaytheone]

.

[lessthanjake]

The ones I referenced included NBArapm that you stated above, along with:

-https://drive.google.com/drive/folders/1T3HuKvVAxBrZDt1LgSlViBczLDpM7q-Q

I really appreciate the list you’ve provided, and will probably keep this bookmarked for future use myself!

Please note that the above is actually NBAshotcharts RAPM. That’s just the raw data for it.

-https://web.archive.org/web/20150408042813/http://stats-for-the-nba.appspot.com:80/

I will note for people that my understanding (others can probably elaborate better on this) is that at least the stuff here that’s from before 1997 as well as the “entire 90s” thing isn’t really true RAPM. It’s Engelmann using quarter-by-quarter box scores to try to build a RAPM-like model. I think it’s an interesting idea and has some value, but it’s not actually really RAPM.

That’s all I have.

There’s also the 1997-2024 career RAPM that Engelmann has done. There’s two versions of it. The first is here: https://docs.google.com/spreadsheets/d/1bg8KxzagN7D0O16EmUO9_kCyXwthEUjKywlrWPQUQt8/edit#gid=0. The second is here, and this one basically differs by layering on more adjustments for things, including for age and coaches: https://docs.google.com/spreadsheets/d/1pGTFzq0eE85AP5wW8v8yFzRiJn_lfSCAzh7hd4czQI4/edit#gid=0

There’s also some RAPM stuff from Cheema. You linked to a chart that has the 25-year RAPM data for postseason and regular season, but he also did it in 5-year snippets: https://www.thespax.com/nba/quantifying-the-nbas-greatest-five-year-peaks-since-1997/

There’s also Cryptbeam’s RAPM: https://www.cryptbeam.com/rapm/

There’s also a bunch of RAPM from prior years here. I think it’s stuff done by Engelmann, but am not sure: https://web.archive.org/web/20201024055554/https://sites.google.com/site/rapmstats/rapm-by-player

Perhaps the best in terms of being similar to NBAshotcharts is here: https://thebasketballdatabase.com/2022-23RegularSeasonPlayerRAPMComprehensive.html. You can go back to different seasons and get one-year, three-year, and five-year RAPM for each year, so there’s actually a lot there.

[homecourtloss]

https://psteve.shinyapps.io/RAPM/

1-year 2024 results
Brunson 1st in ORAPM.
OG 1st in DRAPM.
4th and 5th, respectively, in total RAPM with Hartenstein at 6th.

I vote for this one as the perfect source. Hoping for 57+ish wins next year assuming we retain OG and IH and with decent injury luck.

I like this. I made quite a bit of money off of OG’s play before the lines started shifting when Vegas knew it wasn’t a fluke.

Y'all were crazy good with OG last season (20-3), seems possible.

I don’t expect them to perform this well again, but I think 52+ish wins is doable.

[homecourtloss]

-https://web.archive.org/web/20150408042813/http://stats-for-the-nba.appspot.com:80/

I will note for people that my understanding (others can probably elaborate better on this) is that at least the stuff here that’s from before 1997 as well as the “entire 90s” thing isn’t really true RAPM. It’s Engelmann using quarter-by-quarter box scores to try to build a RAPM-like model. I think it’s an interesting idea and has some value, but it’s not actually really RAPM.

Of course not because you need play-by-play data to generate an RAPM set.

I don’t know if Engelmann ever did it, but it would be interesting to see the R squared between some NPI RAPM years we do have, i.e., 1997 and onwards, and this RAPM model if he created them for years 1997+.

[eminence]

https://psteve.shinyapps.io/RAPM/

1-year 2024 results
Brunson 1st in ORAPM.
OG 1st in DRAPM.
4th and 5th, respectively, in total RAPM with Hartenstein at 6th.

I vote for this one as the perfect source. Hoping for 57+ish wins next year assuming we retain OG and IH and with decent injury luck.

Make it 60+

[Throwawaytheone]

.

[lessthanjake]

There’s also some RAPM stuff from Cheema. You linked to a chart that has the 25-year RAPM data for postseason and regular season, but he also did it in 5-year snippets: https://www.thespax.com/nba/quantifying-the-nbas-greatest-five-year-peaks-since-1997

What would you say is the difference between this and NBAshotcharts? Is one PI and the other NPI, or is it something else?

There’s also Cryptbeam’s RAPM: https://www.cryptbeam.com/rapm/

Never even knew this existed. Honestly, seems like there's so many variations it's impossible to know which one the community should collectively rely on, half the time it seems people just spout contradictory RAPM numbers from wildly different sources.

I think Cheema uses a prior, based on the player’s minutes per game and the team’s net rating. I believe NBAshotcharts doesn’t use a prior at all. Meanwhile, the Engelmann/Goldstein stuff uses a prior as well I’m pretty sure, but it’s kind of black box so we don’t really know what exactly the prior is comprised of.

The other thing is that there’s other methodological decisions that can be made even beyond a prior. For instance, what do you do in the model with players that have played very few minutes? I think different models treat those players differently—which ends up affecting numbers for everyone. Some models adjust for garbage time, since teams do worse when ahead by a lot, and so that is arguably a relevant thing to adjust for. There’s also potentially “luck” adjustments, where the model basically assumes that three-point-shooting and FT shooting variance is luck and adjusts for how much “luck” someone has had on and off the floor. RAPM models can also use different methods to scale the results. For instance, do you scale the output by standard deviations (and if so, how exactly are you doing that), or do no artificial scaling, or some other type of scaling? This is part of why you really can’t compare the actual numerical output from one model to the output from another—the fact that they likely are scaled differently means that the actual numbers aren’t really comparable.

So there’s just a lot of different variables here in terms of how exactly the RAPM model is done, such that all of these are going to be subtly different. And changes will of course end up being to the benefit and detriment of some players, such that people will be able to cite to contradictory RAPM data. I think it’s probably a good practice to look at various different RAPM sources just to get a general sense of where someone falls in the various sources in general. Usually different RAPM measures will be very similar to each other for a given player, but sometimes there will be huge differences, and in those cases you can tell that methodological differences have affected the output for that player a lot. To me, that makes me more skeptical of that player’s data, since I don’t feel like I have a confident answer as to which methodological choices are actually best. But other people decide that they’re just very confident in one version over another.

There’s also a bunch of RAPM from prior years here. I think it’s stuff done by Engelmann, but am not sure: https://web.archive.org/web/20201024055554/https://sites.google.com/site/rapmstats/rapm-by-player

Perhaps the best in terms of being similar to NBAshotcharts is here: https://thebasketballdatabase.com/2022-23RegularSeasonPlayerRAPMComprehensive.html. You can go back to different seasons and get one-year, three-year, and five-year RAPM for each year, so there’s actually a lot there.

That's sick, yeah I'm definitely sticking to that one. Has everything, the adjustability, 1/3/5 yr, RS/PS, and is well formatted. I retract what I said 5 seconds ago about lacking a consensus variant then hahah.

How similar would you say this is to something like Squared2020's tracked RAPM? This of course has a far larger sample size, but aside from that, as fundamental formulas, do they attempt to do the same thing?

Thanks for all the information, appreciate it.

So I think this is conceptually similar to Squared’s RAPM, in the sense that neither this nor Squared’s RAPM have a prior. That said, a big difference here is that Squared’s data is based on partial seasons, because Squared is compiling it himself just based on tape that he’s able to get a hold of and watch and do tracking for. That makes the samples for a lot of players smaller and less fulsome, such that we should have less confidence in it. But, of course, for the years Squared is doing it for, there’s no play-by-play data (we only have that from the 1997 season onwards), so there’s not actually any way to get better data for those years. Squared’s data is therefore inherently flawed but is the best we have for those years.