I see we're doubling down...
f4p wrote:OhayoKD wrote:Well, we're not off to a great start:
f4p wrote:
You referenced this once, and then proceeded to ignore the explanation:
please explain why my explanation of the article is wrong. in fact, i proffered my same explanation before and someone on here said:
"Seems about right, yeah."
Another someone on here who I shan't name also saw that explanation, called you an idiot, offered a much ruder variant of my response, and then deleted it in timely manner after realizing their emotions had gotten the best of them. I'm glad you felt validated by a line of support, but that approval was not unanimous.
But whatever, maybe mathing this out will illustrate my point more clearly
Inserting question marks so you wouldn't read it so aggressively seems to have backfired, so let me be more definitive:
1. The scale is artificial, you cannot extrapolate "closeness" like that
based on what? where the does the article say it is artificial? it says they are predicting net rating. and how would even an artificial RMSE not be close if the numbers are 2.6 and 2.85? it's still only a difference of 10% of the overall error.
If I just guessed the net-rating of every team was 0, I would get an error-rate around
3. Actually IIRC, Unibro did that calc in January and got an error rate of
2.8 which you might notice is
"lower" than
2.85. But hey, let's just be nice and go with
3. Maybe the blowouts got more frequent since?
Taking this scale at face-value, PER is much closer to a random guess than EPM is. Quantifying things, we are comparing a difference of
.25 to
.05. You might notice that first number is
5x bigger than the second.
And, as you've managed to side-step(again), that gap comes from looking at 30 teams. Assuming the box-stuff offers any improvement over random guessing, the fluctuations naturally even out when you're measuring bigger sets of data. So the gap could get much bigger when we look at a much smaller part of the whole(an individual player)
In fact...
well yes, we know the box score doesn't see everything. but it does track direct things that you do. the things you describe all have to be sussed out secondarily from gobs of lineup data, where we hope the metrics figure out who gets credit for what.
Well, with all due respect to "your own experience", the lineup-data does a better job of sussing out who deserves what credit based on the article you referenced. You keep forgetting this, but those metrics didn't just grade out as more predictive, they graded out as less susceptible to roster turnover. TLDR: They're less...
IOW, the accuracy gap we looked at above is one we'd expect to widen as we started looking at individual players.
We also don't have to "blindly hope", we can look at winning directly(raw-signals), track historical trends, and directly vet how these metrics are evaluating specific types of players in specific situations. In this case the raw stuff, rapm, and the stuff that is both stabler and more predictive than all the box-stuff you like seem to disagree with you. Hence you've resorted to blurring the lines between different approaches so you can argue based on "my own experience" that they should all be disregarded because they all have the same
definitely real disadvantages compared to box-aggregates.
The "box-score" does actually need "larger sample sizes" than something like EPM and LEBRON. The "box-score" sees less than something like RAPM and sees alot less than something like wowy, let alone an indirect full season sample.
Counting things with some loose correlation with "goodness" can make a metric less accurate if it overvalues those contributions(uncontested rebounds are worth alot less than contested rebounds), or it fails to count negative counterparts(a blown steal attempt can lead to a good shooting opportunity), or it fails to account for schematic or team-based context(Curry for example is a player who we've repeatedly seen trade scoring for off-ball creation when playing with more capable scorers(the 2017 finals being a premier example))
You're just saying things that you think sound good without actually vetting if they hold up under any sort of scrutiny. What exactly is your basis for "the best players" if you're going to disregard that which is more directly related to winning?
All this "gish galloping" leads to an analogy that sucks...
your aversion to anything box score related, based on the errors from that article, is like saying hurricane tracker A has proven slightly better at predicting than hurricane tracker B in certain situations, and since A says it is going to miss your hometown and B says it is going to be a direct category 5 direct hit, you aren't making any preparations.
I'm going to explain to you all the reasons this analogy doesn't work and then replace it with one that actually makes sense
1. It's not "just based on errors from that article", it's based on errors from many articles, and historical trends(what correlates with team success, and what players/skills correlate directly with team success), and film-tracking(ex: generationally gifted ball passers -> "frequency of good passes per 100 is much higher even with similar box OC" -> all-controlling guys like Magic, Nash, and Lebron direct impact/team results outpace their box-score while "delegators" see the opposite effect)
2. Following A incorrectly and following B incorrectly lead to the exact same outcome(being wrong), yet somehow your hypothetical sees a much higher downside for following A.
3. As covered above, we have no reason to prediction A think is "slightly" more accurate
A proper analogy would go something like this:
A hurricane is coming but we don't know which city. A weather predictor that is consistently more accurate in all sorts of situations says it's going to come at Tokyo. A less accurate predictor says it's going to come in Kyoto. Moreover, predictor A's prediction rate declines less when there are more situational variables(like a sudden influx of weather mages) while predictor B does about as well as Steve who decides which city a hurricane will hit based on how much he likes the pancakes his wife, Clara, makes him in the morning.
There also happens to be a guy named Ohtani, named after the late great baseball player after he was killed by a blood-thirsty unicorn, who has memorized every single time a hurricane has hit a city and can see all of Japan with a bird's eye view. He happens to see more warning signs for a hurricane near Tokyo than Kyoto and these same signs have repeatedly correlated with Tokyo being hit over the last 10 years.
Finally we have you, claiming that actually the hurricane is going for Kyoto because predictor B has brilliantly ascertained that things there are more felled trees when a hurricane is around. Did our predictor check with Ohtani? No. He blames Ohtani for cursing his cow and insists that because there are 4 more felled trees in Kyoto, the hurricane is more likely to go there.
You, hearing this, decide the hurricane "is much more likely" to head towards Kyoto and Tokyo actually doesn't have a high chance of getting a hurricane at all because there aren't so many felled trees.
And when I, Ohtani's secret lover(he's got great legs fwiw), point out that's a bad argument, you pivot to "a hurricane hitting tokyo is less likely than it was last year" because the original thing you were arguing isn't something you can properly defend.
Again, I challenged "2022 Lebron clearly had a much better regular season than 2022 Curry". This is a dodge. Ohtani would never.
You have a talent for stretching out bad logic and weak argumentation into moderately amusing word salads so I'm going to attempt a concise summary of what you've argued in this post:
1. Holistic data(lineup-adjusted, raw or otherwise) is wrong because... Curry's scoring efficiency fell
not "wrong". simply too situation-dependent and with error bars too large to trust that it has to be right, especially in the face of more primary evidence of steph playing worse.
The primary evidence supports the evidence with
lower error bars. Unless we don't agree that a player's job is to make teams win, box-aggregates are not "primary", they are secondary and have a weaker connection to our "primary" evidence both methodically and in terms of results.
You're
0 for 12. Holistic data(lineup-adjusted, raw or otherwise) is wrong because... I have a different opinion on Steph Curry
no, because i saw him look way worse and he probably didn't develop massive new levels of intangibles specifically in the 2021 offseason, after having a whole career to develop them, to offset his worse play.
"Intangibles" being whatever isn't captured in the box-score. And somehow this gets us to 2022 Lebron>>>2022 Steph
You're
0 for 23. 2014 is basically peak Lebron because...it's one year removed from arguable peak Lebron which is arguable because PER and PER is good here because.... the 4 consensus best regular seasons of Lebron's career happen to be the four highest PER scores of Lebron's career
yeah, basically. along with WS48 and BPM and watching him play.
So two metrics with the same issues as PER propping up a season that one, I never compared with Steph's, and two is not actually seen as a high-end regular season by the same consensus you're using to justify "near 2013" specifically because of something that PER would do a terrible job accounting for.
0 for 34. You think Lebron is great because WOWY(the straw strikes again!) so that means a PER advantage Lebron holds over someone must be accurate
it means lebron isn't empty calories and has a career of both box score and impact/WOWY dominance, indicating they are closely tied together and wouldn't separate just for one particular season, before becoming closely tied together again.
Yeah, but being "dominant" does not mean the metrics view him as comparably dominant, nor does it mean that when he's nearing 40, and the main source of him being underrated has gone away(his ability to function as a defensive anchor), his PER is bound to be accurate vs a guy who is
tangibly an all-time creator(effect on teammate efficiency is all-time) in spite of simple box-score not seeing him as one(assists require you complete the final pass).
0 for 45. 2022 Lebron is awesome because... he could have won a scoring title
lebron wasn't old and falling apart and a shell of himself because... he could have won a scoring title.
And 06 Lebron is obviously peak Lebron because he
actually won a scoring title.
0 for 56. The stats are close in accuracy because... 2.48!
because the RMSE is 2.48 net rating points and for like 10 other metrics, it's 2.85 or lower, meaning they are all basically the same.
And that makes you
0 for 6Ohtani would never.
EDIT: The 2.8 comes from "MAE"(margin average error) as opposed to the method used in the article, "RSME"(Root-squared Margin error). Since these are different methods you can't derive a gap this way. In order to make the case for "closeness" or lackthereof we'd probably need to take the "RSME" of our control methodology and then look at the difference between the methods. Taking the percentage difference between 2.45 and 2.85 doesn't really achieve anything without an additional reference.
