lessthanjake wrote:I don’t think looking at an average of the top 10 in the league is all that responsive to what we’re talking about here, since the point is precisely that the very top offensive guys are way above the borderline top 10 guys.
That's true for both ORAPM and DRAPM.
In any event, I think you’re making a mountain out of a molehill. Just for clarity’s sake, I first want to note that the NBArapm website actually very confusingly has two different forms of 5-year RAPM. It doesn’t explain that they’re different, but the “Six-Factor RAPM” is different than the RAPM you see if you search for someone’s name. Not sure which one is better or what all the differences are, but I think one is Engelmann and one is done by the website’s creator himself. In any event, you’re looking at the “Six-Factor RAPM.” When we look at that, we see the offensive peaks being around +8 or +9. Those are mostly Steph and Nash, though, and for guys who aren’t Steph or Nash, the offensive peaks are mostly between +7 and +8. Meanwhile, while Garnett has a weird DRAPM from non-peak years that is way above anything else, the defensive peaks are generally between +6 and +7. And for anyone not named Kevin Garnett, they’re actually between +5 and +6.
So, instead of picking what matters and what not, just take a look at 10 best peaks with ORAPM vs DRAPM (only one year per player):
Offense:
2019 Curry: +9.1
2010 Nash: +8.2
2017 James: +7.8
2019 Harden: +7.8
2022 Young: +7.4
2021 Lillard: +7.2
2010 Wade: +7.2
2024 Booker: +6.8
2010 Bryant: +6.6
2019 Paul: +6.6
Average: +7.5
Defense:
2013 Garnett: +7.6
2013 Bogut: +6.4
2005 Duncan: +6.3
2021 Gobert: +6.1
2004 Bradley: +5.9
2022 Roberson: +5.8
2007 Collins: +5.7
2008 Hayes: +5.6
2004 B. Wallace: +5.5
2004 R. Wallace: +5.3
Average: +6.0
Again, not close to +3 you threw previously and again, most of the data comes from offensive minded era.
While the numbers are a bit different in the other RAPM on that website, it’s generally a similar story, so I won’t bother listing more numbers. All of this is roughly indicative of offensive peaks being about +2 higher than defensive peaks. This is significant! You can say that this measure indicates it’s +2 rather than +3, but this is only one measure, and other measures I’ve mentioned have different gaps, including ones that are bigger than +3. The point is that there is a notable gap!
Yeah, but you threw significantly bigger number in RAPM than anything we can actually see. The difference between top 10 players this year is +1.3. the difference in top 10 peaks ever is +1.5. The difference in the best seasons ever is +1.5. The difference between top 3 is +1.6.
None of these gaps suggest +3 difference, which leads me to the conclusion that hybrid stats overstate offensive impact in comparison to defensive output.
I’d also note that I did specifically say upfront that the point was the general concept, rather than the specific numbers I mentioned.There is a significant gap between peak offensive impact and peak defensive impact,
Nobody denied that, but there is a huge difference between +1.5 and +3 or +4 difference you suggested. The first is a meaningful difference, but the second is like 3 tiers of difference between two players. Again, nobody here says offense isn't overall slightly more valuable than defense now - I argue that you overstate the difference, which is probably true looking at the numbers.
such that an offense-first guy can be worse on defense than a defense-first guy is on offense and still be a more impactful player. I don’t think that can really be denied in a general sense. The question is just whether that is the situation here.
Again, I didn't say it's impossible - I just said it's very debatable in this situation.
Not sure why. All these measures are getting at impact. And if we’re talking about peak impact, they’re actually pretty clearly the best way to look at it (because peaks are inherently relatively short timeframes and RAPM itself is noisy in relatively short timeframes and hybrid models help a lot with that).
Because hybrid metrics are heavily influenced by the presuppositions of the authors creating formulas.
As you know, we do not have impact data for the 1970s and 1980s, so obviously that kind of evidence can’t be produced. But the fact that the most successful star players went from defense-focused ones to offense-focused ones paints a pretty good picture of a shift.
What teams in the 1980s were offense-focused? Certainly the Lakers, but they have two of the best offensive players ever. Definitely not Celtics, who had defense-first teams throughout the majority of decade. Definitely not the early 1980s Sixers or late 1980s Pistons. Definitely not 1980s Bucks.
And I also think that that shift is pretty evident just watching the games in those eras. In significant part due to rule changes (relaxation of dribbling rules, the three-point line, etc.), great offense eventually just started being able to usually get the better of great defense in a way that wasn’t true in earlier decades.
Three point line was irrelevant for vast majority of the 1980s. Relaxation of dribbling also happened relatively late. The only thing I can see as the legit point is the introduction of illegal defense, but I don't think it had that huge impact for the majority of the decade.
You think it's evident, but I disagree.
The sample for Hakeem isn’t massive, but it’s not incredibly tiny either. It’s like a season and a half worth of data. Definitely still small enough to be noisy. But I *really* think you shouldn’t throw Squared’s data in the trash on the basis of him providing “LOW” and “HIGH” values. RAPM is noisy enough that if others did that, you’d be very surprised how wide the error terms are. I think it’s very good that Squared provided that, and I think others should be encouraged to do so, which to some degree requires people not to say measures should be thrown away when they do so, in favor of metrics that probably have similar error terms and just don’t tell you. The error term matters, but it’s really not exclusive to Squared’s data.
So all you concluded from my post is that we should throw Squared work to the trash?
I am well aware that RAPM studies have high level of uncertainty, but I don't think you really looked at the data closely. Hakeem's sample is so small that it has like 9 points error bar. It literally says that Hakeem ends up somewhere between average player and the best player in the league. That's all we can conclude from it. Other players like Jordan or Magic, that have significantly more games tracked, have also relatively high error bars, but they look reasonable. Even at worst, they look like all-nba level players based on the data.
It's not a matter of providing error intervals (which is a very good thing, more people should do that), but a matter of signal stability and sample size. No, 1.5 worth of games from full decade isn't meaningful, especially when the data itself says we can't take much out of it. Data analysis doesn't work this way.
In any event, this is all a probabilistic exercise. I don’t think it’s particularly meaningful that Hakeem is at 5.23 and Barkley is at 4.99, because there’s enough noise in this data that that probably only means there’s barely more than a 50% chance Hakeem was more impactful. But that’s why I mentioned how far off Hakeem was from the very high end guys. While there’s error terms on both sides of this, this data would indicate that the chance that Hakeem was as impactful as Jordan or Magic (at least in the games in these samples) is very low, even if it is possible.
Hakeem's sample shows us that Squared hasn't provided any signal for him yet, that's all it says.
It reminds me people tracking 2 Wilt games and concluding he couldn't shoot fadeaways, because he went 1/5 on that shot. This is not probabilistic exercise, this is relying on unreliable data.
https://web.archive.org/web/20150329072440/http://stats-for-the-nba.appspot.com/ratings/90s.html
Thank you!
