25-Year RAPM (RS and Playoffs)

[OhayoKD]

ftr, I'm not super-in-love with ultra-long RAPM samples.

I think working in 3 or 5 year chunks, and in general putting more focus on the ceiling we see from a guy tells us how good he was, though of course that doesn't factor in all context either.

Edited the post with that in mind fwiw. There's also 5-year data from ahmed in a post above.

One thing I think is pretty funny is the idea that Nash disappointed +/- in the playoffs. I'm having trouble finding my initial study on this, but when I looked at a bunch of stars and compared their playoff Team Wins to their playoff OnWins (positive +/- in the game), Nash stood out above anyone else I saw.

A quick bkref query shows Nash:

69 OnWins
57 team wins
That's a +12 edge.

While I understand this is a very coarse metric and that someone's impact can still disappoint relative to the regular season even while coming off well with the metric, just for perspective here, if I do the same thing in the regular season:

748 OnWins
764 team wins
Down 16 instead of up 12.

If that seems bizarre to you given that Nash is known for his regular season impact, just know that this is the norm. Stars on great teams often have teams capable of outscoring the opponents on the bench. So this data, while unexpected to most, is actually pretty normal from my analysis.

The spike in the playoffs though, while not utterly unique, is a definite stand out.

What does it say about a guy when is RAPM says he disappoints some in the playoffs, but a study like this shows him as a positive outlier? At the very least I'd say, it means we shouldn't run too far too fast with the data.

Could you refresh me on what "on-wins" is again?

[Doctor MJ]

ftr, I'm not super-in-love with ultra-long RAPM samples.

I think working in 3 or 5 year chunks, and in general putting more focus on the ceiling we see from a guy tells us how good he was, though of course that doesn't factor in all context either.

Edited the post with that in mind fwiw. There's also 5-year data from ahmed in a post above.

One thing I think is pretty funny is the idea that Nash disappointed +/- in the playoffs. I'm having trouble finding my initial study on this, but when I looked at a bunch of stars and compared their playoff Team Wins to their playoff OnWins (positive +/- in the game), Nash stood out above anyone else I saw.

A quick bkref query shows Nash:

69 OnWins
57 team wins
That's a +12 edge.

While I understand this is a very coarse metric and that someone's impact can still disappoint relative to the regular season even while coming off well with the metric, just for perspective here, if I do the same thing in the regular season:

748 OnWins
764 team wins
Down 16 instead of up 12.

If that seems bizarre to you given that Nash is known for his regular season impact, just know that this is the norm. Stars on great teams often have teams capable of outscoring the opponents on the bench. So this data, while unexpected to most, is actually pretty normal from my analysis.

The spike in the playoffs though, while not utterly unique, is a definite stand out.

What does it say about a guy when is RAPM says he disappoints some in the playoffs, but a study like this shows him as a positive outlier? At the very least I'd say, it means we shouldn't run too far too fast with the data.

Could you refresh me on what "on-wins" is again?

Just something I was tracking a while back that I felt needed a succinct name.

So, an OnWin is a game where your team outscored the opponent when you were on the floor, aka, a positive +/- game.

The idea behind this is that it's not actually a player's goal to help his team win every game by as many points as possible, but just to outscore the other team. So why not make a simple threshold metrics that can be juxtaposed with actual team wins?

Actually just found the spreadsheet I was thinking of. If there's enough interest I can upload as a Google sheet.

To just give some barebones:

* I went through 100 of the guys from the +/- era and compered there team Wins to OnWins for both the regular seasons and the playoffs through their career.

* If we look at OnWin+, my short name for how many more OnWins than Wins, one of the unexpected things is that most of these great players break negative.

* When I tally this up over career, RS & PS combined, there are two players way out in front: Kevin Garnett & Dirk Nowitzki, which isn't necessarily a surprise to anyone who knows their +/- reputations, but other guys are unexpected.

Here's where the Top 10 stood when I tallied these - which I think was last off-season:

1. Kevin Garnett 37
2. Dirk Nowitzki 36
3. Thaddeus Young 20
(tie) Antawn Jamison 20
5. Kyle Lowry 18
6. Vince Carter 17
7. Steph Curry 16
8. Mike Conley 11
9. Rudy Gay 10
(tie) Lou Williams 10

And when I do it only for the playoffs:

1. Steve Nash 12
2. Jason Kidd 10
3. Chauncey Billups 9
4. Manu Ginobili 8
5. Kevin Garnett 7
(tie) James Harden 7
(tie) JJ Redick 7
(tie) Lou Williams 7
9. Paul George 6
(tie) Kyle Korver 6

I'll note that I also have a ratio stat comparing OnWin/Win, which I labeled OnWin/. I did this because it was a simple way to normalize for how much you played. When you do this, it's noteworthy in the way the top guys stand out in the playoffs.

In the regular season, Garnett had an OnWin/ of 1.043 while Nowitzki's stood at 1.037.

In the playoffs? Nash has a monster OnWin/ of 1.211. After that there's a huge drop off to Kidd at #2 with 1.127, but note how much bigger that is than 1.043. Everybody in the top 10 here has a PS OnWin/ bigger than than Garnett's RS OnWin/...which of course includes Garnett himself.

Now, big caveat: This was a hand-made list. I make no claims that I've identified the ultimate leaderboard here. It would make sense for me to scrape the necessary data, and maybe at some point I or someone else will, but this is what I have right now.

[Doctor MJ]

A couple other things here:

Curry had 42 OnWins this year while playing 44 wins, so I believe he'd drop from a career OnWin+ 16 to 14. Harsh mistress this stat.

In terms of the all-season season OnWin leaderboard:

1. Jokic 69
2. KCP 67
3. Tatum 65
4. White 65
5. Gordon 64

Also probably no surprise that Jokic tops the playoff OnWin leaderboard, but it's particularly interesting because of high he ranks on a list from '96-97 onward. Here's a list of the highest OnWin playoff seasons, with other games played after the dash:

1. Duncan '03 20-4
2. Ginobili '05 19-4
3. Jokic '23 18-2
(tie) Wade '12 18-5
(tie) Ginobili '03 18-6

I'll note that Draymond & Steph had a 16-1 OnWin record in '17, so Jokic '23 isn't quite the highest OnWin%, but he's up there.

[AEnigma]

Hm, that is an interesting metric, albeit not one I am inclined to give much weight based off those leaderboards.

Outside of the usual pitfalls with these raw plus/minus measures, one significant limitation is the lack of application for minutes. A player who never left the game but was -1 was certainly more “valuable” than a player who was +1 but in only 28 minutes.

Still, for Dirk, Garnett, and Nash, a nice feather in their caps.

[One_and_Done]

Unless you have Draymond Green in your top 20 all-time I don't think this is much of an argument for anything. It's nice to add data points to the other evidence about who was better, but you can't just be a model junkie. You have to apply some common sense and context.

A guy being 0.4 better or whatever over their careers is a weak argument no matter how you slice it. There's too much noise for such small differentiators to be compelling. Also people don't use these stats consistently, only when they want to advance their narrative like Stockton was secret franchise player, or KG is secretly the GOAT, neither of which is true.

I'd also agree that these stats are more useful, in a comparative sense, when limiting them to the players peak (or a small 3 year stretch or something). Career is problematic for the reasons pointed out.

[Doctor MJ]

Unless you have Draymond Green in your top 20 all-time I don't think this is much of an argument for anything. It's nice to add data points to the other evidence about who was better, but you can't just be a model junkie. You have to apply some common sense and context.

So, I'm particularly struck by the first sentence, as to me this implies that you either use a metric as your literal opinion, or it doesn't amount to much. I strongly disagree with this.

I look at the data, use it to make holistic assessments, and from there it informs rankings. Data like what's shown in this thread I find to be useful, even if I don't consider Green to be a Top 20 all-time player.

[eminence]

I haven't really thought about where I'll be voting for Dray. I imagine kinda high, though not as high on his recent years.

[OhayoKD]

So looking at this data,

In the regular season, it's:

1. Lebron
2. Garnett
... gap
(everyone else)

In the post season, it's:

1. Lebron
2. Draymond
3. Manu
4. Garnett
...

No? The gap between Lebron and Garnett is bigger than the gap between Garnett and Duncan and Duncan has 40,000 more possessions on him while Lebron has 70,000 possessions on him. Draymond and Many are at a massive possession disadvantage without a significant rate one.

I don't understand this interpretation of the data at all. Even if you were just going per-possession, KG is not closer to Lebron than the field

[uberhikari]

As somebody who's thought about the GOAT debate for a lot over the last 6-7 years, I've only recently come to the conclusion (within the last year or so) that LBJ, MJ, Russell, and KAJ are in the GOAT tier. And I think you can have them in any order from 1-4, and I won't knock you. To me, somebody with this list is imminently reasonable:

1. Russell
2. KAJ
3. LeBron
4. MJ

But if somebody primarily values both prime and longevity I can't see how they could ever have MJ or Russell over LeBron. The NBA has been around for 77 years, and LeBron has literally been in the NBA for 26% of its existence. The fact that he's #1 in both the RS & PS AND has 25k more possessions than the next closest player when it comes to longevity (Kobe Bryant) is downright crazy.

Given the current and future state of the NBA trending towards load management, I don't even think this can be replicated.

[Tim Lehrbach]

And with another round of APM data comes a great big heap of inappropriate commentary.

So many references to goodness or one player being better than another -- impact data doesn't say that. Plus-minus reports what happens to the score when lineups are on the floor, and, depending on the data set, formula, prior, etc., adjusts for teammates, opponents, and anything else you want to throw in there. It does not spit out a ranking of the best players but a figure for the impact of each player as an input within and against the set of lineups. It doesn't explain how or why that score is what it is, only what effect that input has on the score. We all know this intuitively because none among us uses impact metrics as our lone factor in grading players. Yet, we slip so easily into this language of comparing players' APM scores against each other as measures of player quality.

I'm sorry to be grumpy about this but it's an important issue for me. I am a huge advocate for the proper use of impact stats. I think their best use is falsifying affirmative statements about players. For example, if I say so-and-so was really killing his team out there on defense, and you bring me a sample showing a positive DRAPM impact, well, my statement isn't really worth much. But what your sample doesn't show is that the guy "played good defense." It just tells us the the input of this player on lineups did not result in bad defense as I had claimed.

This is my favorite family of stats by far, but its use can be maddening.

Also me: ZOMG DAME!!!!!

[homecourtloss]

And with another round of APM data comes a great big heap of inappropriate commentary.

So many references to goodness or one player being better than another -- impact data doesn't say that. Plus-minus reports what happens to the score when lineups are on the floor, and, depending on the data set, formula, prior, etc., adjusts for teammates, opponents, and anything else you want to throw in there. It does not spit out a ranking of the best players but a figure for the impact of each player as an input within and against the set of lineups. It doesn't explain how or why that score is what it is, only what effect that input has on the score. We all know this intuitively because none among us uses impact metrics as our lone factor in grading players. Yet, we slip so easily into this language of comparing players' APM scores against each other as measures of player quality.

I'm sorry to be grumpy about this but it's an important issue for me. I am a huge advocate for the proper use of impact stats. I think their best use is falsifying affirmative statements about players. For example, if I say so-and-so was really killing his team out there on defense, and you bring me a sample showing a positive DRAPM impact, well, my statement isn't really worth much. But what your sample doesn't show is that the guy "played good defense." It just tells us the the input of this player on lineups did not result in bad defense as I had claimed.

This is my favorite family of stats by far, but its use can be maddening.

Also me: ZOMG DAME!!!!!

Was the bolded meant for the GB because nobody is doing that in this thread.

[ceiling raiser]

Obviously APM is just one data point, but it's really hard to reconcile Shaq's reputation with his impact footprint.

[eminence]

I'm comfortable saying APM data is decent data for who is 'better'. That is not saying it's perfect and without qualifiers (minutes played is the most obvious one, but plenty of other ones from sample size to aging curves and everything in between). But if APM data isn't acceptable as evidence/support in determining which players are better I legitimately don't know what possibly could be acceptable.

With any stat I think folks often forget to properly account for role (how often have we seen a 31 ppg scorer is 'better' than a 29 ppg scorer?) and make their conclusions too strongly, but I don't think that's remotely unique to APM stats.

[eminence]

Obviously APM is just one data point, but it's really hard to reconcile Shaq's reputation with his impact footprint.

Are you talking career numbers here?

I find his peak/prime numbers pretty convincing. In a tier with Duncan/KG/Curry/Dray behind LeBron doesn't seem a bad place to be.

[OhayoKD]

I'm comfortable saying APM data is decent data for who is 'better'. That is not saying it's perfect and without qualifiers (minutes played is the most obvious one, but plenty of other ones from sample size to aging curves and everything in between). But if APM data isn't acceptable as evidence/support in determining which players are better I legitimately don't know what possibly could be acceptable.

With any stat I think folks often forget to properly account for role (how often have we seen a 31 ppg scorer is 'better' than a 29 ppg scorer?) and make their conclusions too strongly, but I don't think that's remotely unique to APM stats.

apm is legitimate. Nut it is a rate-metric. Eben going by the numbers, a player having a higher value per-possession over signficantly less possessions does not necessarily suggest superiority. Volume also matters. See: 2008 KG vs 2008 Kobe

[ceiling raiser]

Obviously APM is just one data point, but it's really hard to reconcile Shaq's reputation with his impact footprint.

Are you talking career numbers here?

I find his peak/prime numbers pretty convincing. In a tier with Duncan/KG/Curry/Dray behind LeBron doesn't seem a bad place to be.

Peak yes, larger prime, no.

Justifies top 5 all-time peak discussion but not sure if he's a lock top 10 GOAT career guy.

[CptCrunch]

What kind of ridge regression are you running here? How did you choose your shrinkage parameter lambda? Are you running a Bayesian version or frequentist version with no priors. You did mention a 'box-prior' in one post above.

Where are you getting / scraping the play by play data? And how are data elemnts treated. Everyone is calculting RAPM these days and calling it by the same name despite 100 different things that could set each calculations apart.

[rk2023]

Read on Twitter

?s=20

Here's JE's Model from 97-22 if it intrigues anyone. Figured it would be appropriate to share in this thread, believe this is RS+PS inclusive.

[OhayoKD]

Read on Twitter

?s=20

Here's JE's Model from 97-22 if it intrigues anyone. Figured it would be appropriate to share in this thread, believe this is RS+PS inclusive.

Is there a possession-count?

[eminence]

-> For RAPM, A penalty is applied to "outliers" so that they "converge towards zero". In other words, the gaps between players here are suppressed, and not actually indicative of what they would be in the real-world. "Closeness" can only really be gauged here in a relative sense(ex: gap between #1 and #2 vs #2 and #3), an extrapolation like "player a is worth 15% more than player b" doesn't really work.

So I haven't done any large sample APMs/RAPMs recently, but from my understanding I don't think the regularization penalties are a serious concern on a sample of this size (at least looking at players with somewhat similar RAPM numbers, it's probably a little concerning if you're looking at the guys at the top and bottom of the list).

Your optimal regularization value will grow with approximately a parameters/sample size ratio. Very large sample size = relatively small optimal regularization value.

This is a lot of guess, as I don't really know how Cheema did theirs, but on past experience I would guess on a sample this size it is only having a couple tenths of a RAPM point pull at the edges, scaling down to 0 at 0 (duh).

Eg: LeBron's 5.54 might 'really' be a 5.8 (+0.26), while KD's 3.54 might 'really' be a +3.7 (+0.16) and so on.