2015 APM/RAPM/etc.

[cpower]

Those are wild swings in just 12 days. Look at the top 10 DS/lorak posted on March 15th:

Code: Select all

ORAPM   DRAPM   RAPM   Name
3,18   -4,23   7,40   Manu Ginobili
4,53   -2,38   6,91   James Harden
2,83   -3,29   6,11   Draymond Green
3,78   -1,94   5,72   Stephen Curry
4,09   -1,04   5,14   Kyle Korver
1,40   -3,65   5,05   Khris Middleton
1,77   -3,03   4,80   Jeff Teague
2,32   -2,42   4,74   Zach Randolph
3,25   -1,44   4,69   Kyle Lowry
4,05   -0,61   4,66   LeBron James


This is the current top 10:

Harden +10.06
Green +8.38
Curry +8.35
Ginobili +8.22
Morrow +7.97
Korver +7.04
Davis +6.81
Leonard +6.73
Lowry +6.66
Lillard +6.42

These guys slipped out of the top 10:

Teague +4.68
Middlteon +5.52
Randolph +6.39
James +5.91

Do these huge swings not reinforce the instability of NPI RAPM (which I've said year after year)? Considering we only added a week and a half worth of data 80% into the season, this should be a warning sign at the least.

I'm still hoping JE updates his 2015 PI set soon, which I've always been more confident with.

Morrow over Davis. I am not sure if this is noise or nonsense. lol

[Quotatious]

Morrow over Davis. I am not sure if this is noise or nonsense. lol

You have to remember that RAPM has to be used in conjunction with usage/role. Morrow is just a role player, his usage is not nearly as high as AD's, and opposing teams don't gameplan against him like they do against Davis.

RAPM results should only be compared among players of similar caliber/role/usage. Otherwise you'll often get nonsense like Nick Collison ahead of Dirk etc.

[cpower]

Morrow over Davis. I am not sure if this is noise or nonsense. lol

You have to remember that RAPM has to be used in conjunction with usage/role. Morrow is just a role player, his usage is not nearly as high as AD's, and opposing teams don't gameplan against him like they do against Davis.

RAPM results should only be compared among players of similar caliber/role/usage. Otherwise you'll often get nonsense like Nick Collison ahead of Dirk etc.

that's part of the things that advance stat should do already. Refine the model to eliminate more noise, giving better weights on different playing style, etc. Harden and Curry play the game very differently, high efficiency vs high usage, how does RAPM measure one against the other?

[Quotatious]

that's part of the things that advance stat should do already. Refine the model to eliminate more noise, giving better weights on different playing style, etc.

This is a valid point, but it's extremely hard to create a really accurate "all in one" metric. That's why it's important to gather as much knowledge as possible (boxscore, plus/minus, eye-test, everything), and try to make sense of it. Every stat can tell you something useful, if you put it in proper context.

Harden and Curry play the game very differently, high efficiency vs high usage, how does RAPM measure one against the other?

Curry's and Harden's numbers can be compared at face value - they're both high usage/high efficiency, offense-first players, #1 options of their respective teams. You make it seem like Harden vs Curry is like Iverson vs Reggie Miller, where one player has huge usage, but mediocre efficiency, and the other has great efficiency but relatively low usage. Harden vs Curry isn't like that, at all. Harden has a bit higher usage (31.3 to 29.0), Curry is a bit more efficient as a scorer (63.0% TS to 60.4), but the gap is fairly small.

[cpower]

that's part of the things that advance stat should do already. Refine the model to eliminate more noise, giving better weights on different playing style, etc.

This is a valid point, but it's extremely hard to create a really accurate "all in one" metric. That's why it's important to gather as much knowledge as possible (boxscore, plus/minus, eye-test, everything), and try to make sense of it. Every stat can tell you something useful, if you put it in proper context.

Harden and Curry play the game very differently, high efficiency vs high usage, how does RAPM measure one against the other?

Curry's and Harden's numbers can be compared at face value - they're both high usage/high efficiency, offense-first players, #1 options of their respective teams. You make it seem like Harden vs Curry is like Iverson vs Reggie Miller, where one player has huge usage, but mediocre efficiency, and the other has great efficiency but relatively low usage. Harden vs Curry isn't like that, at all. Harden has a bit higher usage (31.3 to 29.0), Curry is a bit more efficient as a scorer (63.0% TS to 60.4), but the gap is fairly small.

I am just surprised that Harden has a massive lead in RAPM by about 20%, that's noise to me already. If Harden was the more efficient player or his team had a better SRS, or his on-off numbers were better or his WS48/BPM/RPM were better, then it would make a little sense. I think the difference between the two is definitely less than 5%, anything beyond that is simply not possible. In short, the stat fails to measure the top players.

[colts18]

Do we have instances in previous years where multiple NPI samples are available? I'm almost certain this exact scenario played out in 2012 as well, we had NPI from like April and then got an update around the Finals which had similarly wild swings, and the conclusion at that time was the same - that NPI was just too noisy within single seasons.

My gut tells me these discussions were either in the Peaks project (it's been 3 years, we should consider this for the summer) or the POY threads for that year. I'll start skimming through them to see if I can find anything.

colts has the older 2012 NPI set archived: Link.

There was an update in I want to say June and that's the one I'm looking for.

I clearly remember Elgee complaining about the wild swing in NPI RAPM during the 2012 season. It was probably the Peaks project because we ran that in 2012. Look through ElGee's posts and you should find it.

[lorak]

I am just surprised that Harden has a massive lead in RAPM by about 20%, that's noise to me already. If Harden was the more efficient player or his team had a better SRS, or his on-off numbers were better or his WS48/BPM/RPM were better, then it would make a little sense.

Everything here has a lot of sense here, we just have to look at full context. Harden has +10.1 RAPM, Curry +8.4, but James plays on +3.5 SRS team, while Steph +10.6, so what Curry has done is more impressive. It's like - for example - improving 50 W% team by 10 wins vs 65 W% by 8 wins. "8" is lower number here, but in fact means impact is greater, because it's more difficult to improve so much better team.

[cpower]

I am just surprised that Harden has a massive lead in RAPM by about 20%, that's noise to me already. If Harden was the more efficient player or his team had a better SRS, or his on-off numbers were better or his WS48/BPM/RPM were better, then it would make a little sense.

Everything here has a lot of sense here, we just have to look at full context. Harden has +10.1 RAPM, Curry +8.4, but James plays on +3.5 SRS team, while Steph +10.6, so what Curry has done is more impressive. It's like - for example - improving 50 W% team by 10 wins vs 65 W% by 8 wins. "8" is lower number here, but in fact means impact is greater, because it's more difficult to improve so much better team.

If that's the case that should be factored too. SRS and team record are not hard to look up and get weighted into the model.
Also GSW is 59-13, if Curry is only +8.4, they are a 51 win team without him? something doesn't add up.

[bondom34]

I am just surprised that Harden has a massive lead in RAPM by about 20%, that's noise to me already. If Harden was the more efficient player or his team had a better SRS, or his on-off numbers were better or his WS48/BPM/RPM were better, then it would make a little sense.

Everything here has a lot of sense here, we just have to look at full context. Harden has +10.1 RAPM, Curry +8.4, but James plays on +3.5 SRS team, while Steph +10.6, so what Curry has done is more impressive. It's like - for example - improving 50 W% team by 10 wins vs 65 W% by 8 wins. "8" is lower number here, but in fact means impact is greater, because it's more difficult to improve so much better team.

If that's the case that should be factored too. SRS and team record are not hard to look up and get weighted into the model.
Also GSW is 59-13, if Curry is only +8.4, they are a 51 win team without him? something doesn't add up.

8.4 is points per 100, it has nothing to do with wins. They're measuring entirely separate things.

[bondom34]

Speaking of APBR, I've been trying to join for a while now and they won't send me the activation email. I sent an email to the guy who posted the sticky about it but I haven't heard anything from him either. Have any of you guys ever had problems with this? And those of you who are on the board, could you point me to someone I could talk to?

I haven't but who are you emailing.

And on topic, this is one of those things where the more I learn, the more I'm questioning reliance on certain metrics. This blew me away w/ how big a shift there is, and its really skewing things.

[SideshowBob]

Single-year NPI RAPM has always been noted for being unstable though. It's not completely useless of course, it's obviously good for noting trends and it does ballpark the high vs. low impact players pretty well. But keep in mind that there is potential for differences in the way the regression is run, so direct comparisons between samples/years should be restricted to the same sources, and this applies for PI RAPM as well.

Single-year PI RAPM was the gold standard for predictive stats prior to the popularization of the blended metrics (IPV/xRAPM) and it's what was conventionally referred to as RAPM (aka "pure" RAPM) when it became popular on this board. Because of the priors, we're essentially injecting a little bit of context into the regression, so on the whole we get better figures.

Amongst all of these RAPM datasets, there are also multi-year NPI RAPM sets in the mix (I've seen this referred to as "vanilla" RAPM).

The lack of organization/proper labeling/standardization here is a bit of a mess (okay no, it's a huge problem). Consider we also have single-year RPM (which is also xRAPM with no-prior), and prior informed xRAPM (which is what JE's site is currently dominated by from 01-13 and 14), "fake" RAPM from 91-00, which is derived via PbP simulation from 90s box-score and may be either xRAPM or RAPM (single-year, may or may not be prior-informed), but is misleadingly labeled simply as RAPM.

And all of these and above are purely what we've been provided by Engelmann, who also has tinkered with and made public a few SPM sets as well, not to mention the long-term multi-year xRAPM and RAPM sets that have been used for derivation of ASPM/BPM coefficients.

[RSCD3_]

Single-year NPI RAPM has always been noted for being unstable though. It's not completely useless of course, it's obviously good for noting trends and it does ballpark the high vs. low impact players pretty well. But keep in mind that there is potential for differences in the way the regression is run, so direct comparisons between samples/years should be restricted to the same sources, and this applies for PI RAPM as well.

Single-year PI RAPM was the gold standard for predictive stats prior to the popularization of the blended metrics (IPV/xRAPM) and it's what was conventionally referred to as RAPM (aka "pure" RAPM) when it became popular on this board. Because of the priors, we're essentially injecting a little bit of context into the regression, so on the whole we get better figures.

Amongst all of these RAPM datasets, there are also multi-year NPI RAPM sets in the mix (I've seen this referred to as "vanilla" RAPM).

The lack of organization/proper labeling/standardization here is a bit of a mess (okay no, it's a huge problem). Consider we also have single-year RPM (which is also xRAPM with no-prior), and prior informed xRAPM (which is what JE's site is currently dominated by from 01-13 and 14), fake RAPM from 91-00, which is derived via PbP simulation from 90s box-score and may be either xRAPM or RAPM (single-year, may or may not be prior-informed), but is misleadingly labeled simply as RAPM.

And all of these and above are purely what we've been provided by Engelmann, who also has tinkered with and made public a few SPM sets as well, not to mention the long-term multi-year xRAPM and RAPM sets that have been used for derivation of ASPM/BPM coefficients.

Too many PM's

[Keller61]

Apparently Kobe is still a positive force on offense (+0.90) despite his 48% TS.

253. Nikola Mirotic -0.20
^ That's ridiculous.

Also, DRose being the Bulls' best defender doesn't quite match the eye test.

[Dr Spaceman]

I haven't but who are you emailing.

EDIT: Well I just checked this morning and I can log in now, so apparently he did it. Cool! Thanks for your help

[Krodis]

I am just surprised that Harden has a massive lead in RAPM by about 20%, that's noise to me already. If Harden was the more efficient player or his team had a better SRS, or his on-off numbers were better or his WS48/BPM/RPM were better, then it would make a little sense. I think the difference between the two is definitely less than 5%, anything beyond that is simply not possible. In short, the stat fails to measure the top players.

Well I think there are a couple of things. For one, NPI RAPM is in fact noisy, so yeah, it's possible that it's just not being accurate. But beyond that you have to keep in mind that RAPM and plus-minus stats in general are trying to capture non-box score impact, so the fact that Curry grades out slightly better in box score advanced stats doesn't really discredit it.

Furthermore, while, yeah, Curry's team is better and his raw on-off numbers are better, part of the adjustments in RAPM is to try to disentangle how much of that is the player and how much of that is other factors. For example, Curry's fellow starters are also very, very, very good players. He plays in the same starting lineup as Klay Thompson, Draymond Green, and Andrew Bogut, all of whom are great players who grade out well in RAPM (especially Green). And raw on-off is especially prone to teammate factors.

[Dr Spaceman]

2009 LeBron was ranked 10th

BTW- if we move forward with the peaks project I really think we should just vote players. The issue with LeBron was I believe people were split between 09 and 12 (seems strange in hindsight), but this can happen a lot with players depending on what one values.

I think the logical thing in situations like this is to assume that if people are split on a player's peak season then we should assume all seasons involved are essentially the same level and just have votes count for the player in question and then have maybe a runoff to decide the actual year. Some guys have very clear-cut peaks (Shaq, Wilt, McHale to name a few), while others can be split into two or even three seasons (Dirk, Kobe, Barkley, etc.) It would be very unfair if, say, 13 James gets 10 votes, 09 James gets 9 votes, and Duncan 03 wins with 11.

[SideshowBob]

Peaks project (it's been 3 years, we should consider this for the summer)

Absolutely we should. I wasn't around the last time they did it, and that project and the RPOY project are things I'm very sorry I missed out on. RPOY obviously can't really be redone due to logistics and attrition, but the peaks project is absolutely something that could be updated.

If I'm not mistaken, that project died last time around #30, correct? We'd have to figure out a way to prevent that, and whether we can do both that and Hibachi's project at the same time (is that still going forward?).

I think we should tackle the RPOY project again in a few years, maybe 2020 after a 10 year gap.

Yeah it started falling hard late 20s and got to 33 I believe. Maybe restrict it to 25 this time around. More importantly, I think we need a better preliminary process with these projects, folks should come in with a list they're fairly sound with, as opposed to just going with the flow so they can sustain argument in later stages rather than having the project fizzle out. If that means cutting the size then so be it.

[E-Balla]

2009 LeBron was ranked 10th

BTW- if we move forward with the peaks project I really think we should just vote players. The issue with LeBron was I believe people were split between 09 and 12 (seems strange in hindsight), but this can happen a lot with players depending on what one values.

I think the logical thing in situations like this is to assume that if people are split on a player's peak season then we should assume all seasons involved are essentially the same level and just have votes count for the player in question and then have maybe a runoff to decide the actual year. Some guys have very clear-cut peaks (Shaq, Wilt, McHale to name a few), while others can be split into two or even three seasons (Dirk, Kobe, Barkley, etc.) It would be very unfair if, say, 13 James gets 10 votes, 09 James gets 9 votes, and Duncan 03 wins with 11.

Problem with that is that's a major assumption. If I feel 09 Lebron is better than 03 Duncan but not 12 Lebron and I don't even rank 13 Lebron all that close to 03 Duncan (which is all true) it should matter.

I feel the other way would make the list end up basically looking like a top 3 year peak or primes list. The best thing about the peaks project was that it had very specific goals.

[Dr Spaceman]

I think we should tackle the RPOY project again in a few years, maybe 2020 after a 10 year gap.

Wow I hadn't realized it's been 5 years already. I think that project was RealGM's most successful and looked like the most fun as well. Maybe it's something we could do earlier, I mean top 100 is done every 3 years.

Yeah it started falling hard late 20s and got to 33 I believe. Maybe restrict it to 25 this time around. More importantly, I think we need a better preliminary process with these projects, folks should come in with a list they're fairly sound with, as opposed to just going with the flow so they can sustain argument in later stages rather than having the project fizzle out. If that means cutting the size then so be it.

I'd be happy with making a pre-list mandatory for participation. I think another big problem was that people would vote a player for one spot and then just repeat the same thing in every subsequent spot until they got voted in. We'd have to think of a way to deal with that as well.

[bondom34]

I haven't but who are you emailing.

EDIT: Well I just checked this morning and I can log in now, so apparently he did it. Cool! Thanks for your help

I didn't do anything, but I'll take the credit!