RE: Kevin Pelton & Realgm - Yes, Kobe IS A Top 10 Player of All Time

[D Nice]

I don't have an issue with a pro-Kobe post. I agree that Kobe often gets an unfair rap on this board and I'm all for someone making a reasoned defense of Kobe. It's clear from the posts of Doc and drza that some of his work in this regard has merit which is great. But sadly imo at least it loses much of its punch with the way he manipulates the data. I won't get into specifics as others have touched on that in some detail already. He did address the two guys who complemented his work, but has ignored other critics who were very polite and respectful in pointing out some issues with his work. That's not how someone behaves who is interested in honest debate. He had an agenda(fine) and clearly only wanted to hear praise for both his methodology and his conclusions.

You have got to be ******** kidding me. I have literally responded to everyone here that has raised points except for the people who asked me to write more stuff on Chuck/Hakeem , brought up Russ/Wilt as the more potentially contentious elements of the overall argument, and parapooers, who as far as I can tell didn't make a point (the "approach" he seems to think is better than my valuation is effectively what I did for the first impact study...have no idea why I am being linked to a BPM plot over a worse window but K -- also that data literally cannot be real because there is no complete play-by-play data before 2001 so I'm not sure how the window extends to 2000...). That is it. Magically somehow 3 of the 11 people on my ignore list have responded to this thread (I'm sure that's just a huge coincidence), so if THOSE are the "posts" you were referring to...lol. But because I remembered you being a far better poster than you've shown here I actually read their posts and, of course, no. The fact that you think they raised "good arguments" underscores limitations in your understanding far more than it does anything else. Literally everything they said either misses the point, is addressed in the post, speaks to something ALREADY QUALIFIED WITH NO ACTUAL WAY OF EMPIRICALLY KNOWING, or is just...there was literally 1 thing Frosty posted that might seem like it's a good point so will address here.

I mean you ask "why is this here" when you have 2 STICKIED THREADS with Kobe at 13 All-Time and 23rd Peak, A PC Board thread with "respectable posters" echoing Mr. Pelton's obviously super awesome "argument," but are continuing to ask why this thread was posted here? Seriously? Please stop leveraging blue letters to troll my thread. Thanks.

Interesting...Ummm wait a sec.. the original data for Kobe was 2008-2011 and you were crusing along with that data...then bam suddenly we add in 2 post seasons where Kobe exited in the first round playing against Phoenix (not known for their defense) in both years and remove his 2011 PS which isn't a very good season for him. hmmmmm

The point of the entire thing is to get as clean a comparison in terms of their extended primes as possible. Including the 2006/2007 Lakres (bottom-feeder supporting casts) in a comparison based primarily on Wins Accrued and Team SRS only works if the support is comparable. So I literally HAD to include a season of post-prime Kobe just to get even 4 "contender seasons as leader" for a comp (because nobody will credit Kobe for 2000-2004...again anybody who actually read the post will have all of this information). But aligning their box scores using 06-10 is the best window to put up against 84-88 Bird. The argument that I "shifted" the window to include 2011 Kobe only hurts Bird's case...using just the 08-10 Lakers makes Kobe's "contender data" look better, not worse.

Bird's Rookie Season +30 wins season doesn't mean anything without context...is the assertion that '80 Bird = GOAT? What is the assertion here? Kobe won 45 games in a much tougher conference with nobody nearly as good as Cedric Maxwell and 2 non-NBA players in the starting 5. None of this means anything by itself.

(Yes, I read disclaimer) Unless I missed something, this is one of those instances where you've done something fairly arbitrary that's actually a large factor in your outcome (top-30 rankings). I'm not sure where you came up with the scoring system (or 2.2x multiplier relative to expected titles) but they are sometimes noticeably different than what I've calculated for top seasons. For me, if I input these with normal portability I get something closer to:

9.3 = 1.09 (0.34 diff)
7.5 = 0.72 (0.43)
6.3 = 0.52 (0.48)
5.3 = 0.39 (0.41)
4.3 = 0.28 (0.27)
3.5 = 0.21 (0.09)

So guys in the "MVP" tier (or +/- one tier) get the largest "boost" based on the estimations your working with. You give Kobe 7-MVP seasons and 2 top-5's. That will slowly add up. More importantly, you've got some very crude buckets that you're forcing players to fall into, when there's a tremendous amount of nuance that can also add up over time. This leads to strange jumps -- for instance, you're positing that in 1987, Jordan improves 3 points per game (!) and gets 0.18 extra titles in 88. (For comparison, I think he improved a point and added 0.07 extra titles -- an enormous 2.6x difference for you. Bird in 83 to 84 does a similar thing.) These are crude jumps on the steepest part of the curve.

Kobe has 9/14 seasons in those tiers
Bird has 9/10 seasons in those tiers
Magic has 11/11 seasons in those tiers
Lebron has 11/12 seasons in those tiers
Jordan has 5/11 seasons in those tiers
Duncan has 12/17 seasons in those tiers
Shaq has 9/15 seasons in those tiers

Who is this supposed to be slanted towards again? And the point of the valuations was not supposed to be tied to anything statistical short of extended prime-based analysis because then you start the cherry-picking game where worse versions of players are on higher tiers than worse versions of said player because of circumstance...not anything to do with their skillset, motor, or understanding of the game. Obviously penalize situations like 04 Kobe or 11 Lebron or D-Rob/Karls PS meltdowns but short of that I was trying not to get bogged down in that stuff just for that part. Obviously I was expecting too much. If I were doing this the way you're purporting either Bird/Magic lose all of their ATG seasons except for MAYBE 86/87 (at NO OTHER POINT do they demonstrate ANYTHING that can qualify as separation from Prime Kobe) and/or 08/09 Kobe and 09/10 Wade (and 2010 LBJ) would all be bumped up a tier.

So many people missing the point though I'm just going to remove the valuations and scores. Should have done this more in lawyer mode.

I only have 8 players ahead of Kobe for all-time offensive peak. So one of the biggest sticking points for me with him are on defense, not offense. Going through your valuations, you likely rate Bryant higher than I do in every season but it's hard for me to see a pattern where this is really offense-based. I imagine most of this is from defense and the crude bucketing of value (since you never cite his peak as ATG but do so with almost all peak players I have ahead of him, save for Robinson, Russell and Dr. J.). You might like his 01 and 06 seasons a touch more than me offensively, but otherwise it's difficult to infer real tangible differences in the yearly valuations.* You do credit him for 2013 when he was injured (?) -- I don't entirely get the logic there and thus I don't count that as value-added either. Either that, or you think he had a GOAT-level offensive peak and then you should just start a separate thread.

Read the 4th impact study dude. Should take...5-10 minutes tops. While I've got you though, can you link me to your 2012-2014 Pop Post?

The issue with the quoted text above is the time period. If you play a time-machine game, then it kind of makes sense to limit a defender like Russell based on available data. However, before the 3-point shot, when spacing and rules were entirely different, it's not only possible for Russell (and, IMO, Thurmond and Wilt) to exceed the impact of today's bigs but the faint statistical signal suggests this was the case. Remember, those guys could only play by the rules of their time, and Russell dominated by those rules.

Not playing any time machine game. All I care about is best at playing basketball. MJ, Kobe, and Lebron would HAPPILY trade the existence of a 3pt line for the ridiculous FTR boost they gain from going backwards, not to mention being guarded by the perimeter defenders of old. They only get more dominant. If "In era" is your primary thing, fine, show me the list with Mikan in top 10 and Pettit in Top 15 and then we can agree to disagree. I'm probably more confident about the Russell stuff than anything written about Kobe here...the amount of data regarding GOAT D vs. GOAT O is staggering. Unless of course Russ is literally 30% more impactful on a per-possession basis than peak Duncan/KG/D-Rob/Hakeem/Ben on that end.

When you were going through the whole "Lebron has terrible portability" argument and using Bosh/Love as your only test cases (I didn't see a link your Lift spreadsheet) I noticed you conveniently left out Kyrie's massive efficiency gain and Wade massive efficiency loss from 2014 -> 2015.

1. Wade's massive decline between 2013-2015 really has nothing to do with anything Lebron was doing/not doing. I stressed the utilization of in-prime data only for primarily that reason.

2. I repeatedly refer to 13/14 Lebron as highly portable. So referring to his impact in 2014 on wade is redundant. Any mentions are strictly 2004-2012/2015-2016 LBJ. We've really never seen anything like this in history where a guy's skillset is that far apart from the rest of his career in 2 seasons.

3. Kyrie Irving is literally on the sheet. A better argument would be that Lebron's boost on Kyrie is a bit blunted by Irving's injuries/rust this year, but I'm pretty sure I made a note of this on the sheet. If I didn't, I should have, wasn't really trying to slight anything towards or away from Bron there. Even if Kyrie wasn't on the sheet (which again, he is) I state more than once that Kyrie is the player archetype Lebron "synchs up" with best.

Methinks I'm mostly done here for now. Nobody seems to have much relevant basketball/math/logic stuff to say to get any real movement on stuff. Still would love to hear more about what makes me wrong on Wilt.

[ElGee]

(Yes, I read disclaimer) Unless I missed something, this is one of those instances where you've done something fairly arbitrary that's actually a large factor in your outcome (top-30 rankings). I'm not sure where you came up with the scoring system (or 2.2x multiplier relative to expected titles) but they are sometimes noticeably different than what I've calculated for top seasons. For me, if I input these with normal portability I get something closer to:

9.3 = 1.09 (0.34 diff)
7.5 = 0.72 (0.43)
6.3 = 0.52 (0.48)
5.3 = 0.39 (0.41)
4.3 = 0.28 (0.27)
3.5 = 0.21 (0.09)

So guys in the "MVP" tier (or +/- one tier) get the largest "boost" based on the estimations your working with. You give Kobe 7-MVP seasons and 2 top-5's. That will slowly add up. More importantly, you've got some very crude buckets that you're forcing players to fall into, when there's a tremendous amount of nuance that can also add up over time. This leads to strange jumps -- for instance, you're positing that in 1987, Jordan improves 3 points per game (!) and gets 0.18 extra titles in 88. (For comparison, I think he improved a point and added 0.07 extra titles -- an enormous 2.6x difference for you. Bird in 83 to 84 does a similar thing.) These are crude jumps on the steepest part of the curve.

Kobe has 9/14 seasons in those tiers
Bird has 9/10 seasons in those tiers
Magic has 11/11 seasons in those tiers
Lebron has 11/12 seasons in those tiers
Jordan has 5/11 seasons in those tiers
Duncan has 12/17 seasons in those tiers
Shaq has 9/15 seasons in those tiers

Who is this supposed to be slanted towards again? And the point of the valuations was not supposed to be tied to anything statistical short of extended prime-based analysis because then you start the cherry-picking game where worse versions of players are on higher tiers than worse versions of said player because of circumstance...not anything to do with their skillset, motor, or understanding of the game.

Slanted? You've missed my point completely. I'm saying you're talking about rounding errors.

You're entire premise is that Kobe is a top-10 player of all-time. You look at a large statistical body of evidence to support that. There's nothing you really say that I balk at. Yet if you've got tiers with 40% increases between the tier, it doesn't make room for a guy being +6.5 and the other +5.75. The "just-below-the-tier" score and "just-above-the-tier" score compound to create large difference over 10 years that aren't accounted for here -- it stands out to me because that's largely the difference between these guys and yet you're accounting method to approximate an all-time list has 9 Kobe years equal to 9 Bird, 9 Shaq, 11 Magic, 11 LeBron and 12 Duncan seasons. That's smoothing out the differences between their prime seasons.

I only have 8 players ahead of Kobe for all-time offensive peak. So one of the biggest sticking points for me with him are on defense, not offense. Going through your valuations, you likely rate Bryant higher than I do in every season but it's hard for me to see a pattern where this is really offense-based. I imagine most of this is from defense and the crude bucketing of value (since you never cite his peak as ATG but do so with almost all peak players I have ahead of him, save for Robinson, Russell and Dr. J.). You might like his 01 and 06 seasons a touch more than me offensively, but otherwise it's difficult to infer real tangible differences in the yearly valuations.* You do credit him for 2013 when he was injured (?) -- I don't entirely get the logic there and thus I don't count that as value-added either. Either that, or you think he had a GOAT-level offensive peak and then you should just start a separate thread.

Read the 4th impact study dude. Should take...5-10 minutes tops. While I've got you though, can you link me to your 2012-2014 Pop Post?

The issue with the quoted text above is the time period. If you play a time-machine game, then it kind of makes sense to limit a defender like Russell based on available data. However, before the 3-point shot, when spacing and rules were entirely different, it's not only possible for Russell (and, IMO, Thurmond and Wilt) to exceed the impact of today's bigs but the faint statistical signal suggests this was the case. Remember, those guys could only play by the rules of their time, and Russell dominated by those rules.

Not playing any time machine game. All I care about is best at playing basketball. MJ, Kobe, and Lebron would HAPPILY trade the existence of a 3pt line for the ridiculous FTR boost they gain from going backwards, not to mention being guarded by the perimeter defenders of old. They only get more dominant. If "In era" is your primary thing, fine, show me the list with Mikan in top 10 and Pettit in Top 15 and then we can agree to disagree. I'm probably more confident about the Russell stuff than anything written about Kobe here...the amount of data regarding GOAT D vs. GOAT O is staggering. Unless of course Russ is literally 30% more impactful on a per-possession basis than peak Duncan/KG/D-Rob/Hakeem/Ben on that end.

-I read the 4th study -- can you explain what I'm supposed to look at there? I don't see the point...

-Not sure which Pop post you're referencing.

-Yes, absolutely suggesting that there's a 30% difference on defense in Russell's day.

-"In era" is my thing but I care about a more mature product versus an emerging one.

[ardee]

Bumping this in anticipation of providing good information to some people before the top 100 project.

[THKNKG]

Bumping this in anticipation of providing good information to some people before the top 100 project.

Definitely has lots of bias/agenda, but even in spite of that, still good stuff overall. I've been higher on Kobe the last month or two than I ever have, and this helps reinforce some of it. Still don't think he's top 10, but much higher than I originally held him. Thanks for bumping it.

[ardee]

Bumping this in anticipation of providing good information to some people before the top 100 project.

Definitely has lots of bias/agenda, but even in spite of that, still good stuff overall. I've been higher on Kobe the last month or two than I ever have, and this helps reinforce some of it. Still don't think he's top 10, but much higher than I originally held him. Thanks for bumping it.

For what it's worth, D Nice is a professional data scientist. His knowledge of how these numbers work is better than anyone's.

Sent from my SM-J700F using RealGM mobile app

[THKNKG]

Bumping this in anticipation of providing good information to some people before the top 100 project.

Definitely has lots of bias/agenda, but even in spite of that, still good stuff overall. I've been higher on Kobe the last month or two than I ever have, and this helps reinforce some of it. Still don't think he's top 10, but much higher than I originally held him. Thanks for bumping it.

For what it's worth, D Nice is a professional data scientist. His knowledge of how these numbers work is better than anyone's.

Sent from my SM-J700F using RealGM mobile app

Oh, for sure, those are the components I appreciated most. By biases, I'm more so referring to some of the conclusions drawn, and that sort of thing. Certainly not the data itself.


Sent from my iPhone using RealGM mobile app

[eminence]

When did the term data scientist overtake statistician anyways? Same dang thing... grump grump grump.

[Blackmill]

When did the term data scientist overtake statistician anyways? Same dang thing... grump grump grump.

I would use "statistician" to refer to some one involved in pure mathematics and "data scientist" to mean some one working on the applied side.

[kennygee90]

This is incredible and as a lakers and Kobe fan I salute you for the hardwork you put in.

[rebirthoftheM]

The data breakdown here was well beyond my knowledge capacity, but I know a lot of posters who will be voting in the top 100 tend to have strong knowledge of, and put heavy stock into data computations. I think what this all confirms is that context is always very important.

[rebirthoftheM]

When did the term data scientist overtake statistician anyways? Same dang thing... grump grump grump.

If you could eminence (since D-Nice went MIA) please extrapolate on the below re the computation of RAPM and how sample sizes impact it:

Take for example Kobe’s NPI single-season RAPM in 2006-2007. Kobe’s ORAPM was 2.35, and DRAPM was 0.2. His ORAPM was almost half of the previous season, but his DRAPM was now in the positives as compared to being negative in 06. Kobe played a total of 3140 minutes, and sat 866 minutes. Of those 866 minutes, 240 (~28%) were the result of Kobe missing 5 full games during the RS. The Lakers in those 5 games had the following ORTGs/DRTGs (will note where either Lamar Odom or Luke Walton, or both of them played in these games because they were the two best Laker players):


• Game 01: Suns: 113.9 (1ST in league)/106.4 DRTG (13th in league)- 114-106 Win- LO & Walton in
• Game 02: Warriors: 107.0 (10th in league)/107.4 DRTG (19th in the league)- 110-98 win- LO & Walton in
• Game 19: Hawks: 103.0 (29TH in league)/108.3 DRTG (23RD in the league)- 106-95 Win- Lamar in, Walton Out
• Game 45: Knicks: 105.7 (17th in league)/108.8 DRTG (25th in the league)- 94-99 Loss- LO in and Walton Out
• Game 62: Bucks 107.0 (Tied for 10th )/ 111.7 DRTG (29th in the league)- 99-110 Loss- LO & Walton out
• Average ORTG of opponents= 107.32 (9th in the league)
• Average ORTG of opponents= 108.52 (23rd in the league)

On average, the Lakers were playing very solid offenses and atrocious defenses in the games that Kobe missed. Logically then, the Lakers would have over performed on offense (even without Kobe, as LO/Luke were rolling earlier on in the season) because of the atrocious defenses they faced, and under-performed defensively because of the very solid offenses they were facing in these games.

Given that about 28% of Kobe’s “off’ minutes occurred during this sample, how much of the above fed into Kobe’s RAPM? Would there likely have been an internal adjustment to take into account the specific quality of the teams faced during the games Kobe missed? Or would it have fed it uncritically? I don't get the process being the computation so please do explain, as I understand RAPM is only useful if you have "off" data (hence if RAPM existed in Wilt's day, it would be utterly useless in determining his value)

I note again that Kobe’s ORAPM was roughly cut in half from 06, whist his DRAPM almost doubled. I find this very suspicious, because 07 Kobe was slower, less laterally quick and less in shape in 07 (a lot of it the result of his off-season surgery), and got lit up even in high profile match-ups- ones he usually would get up to (see Michael Redd/Gilbert Arenas as examples). And also, Kobe was only marginally less dominant on offense in 07 (his slashing game weakened, but his jump shot was at its highest, his play-making higher and he was generally more efficient).

It is unfathomable to have watched the games and conclude that 06 Kobe was a negative defender and then in the same breath suggest 07 was a positive defender. Yet DRAPM suggests this. My tingling sensations asks, whether the fact that 28% of Kobe’s total ‘off’ minutes happened to occur against bottom 10 defenses, meaning the Lakers as a collective would have over-performed on O (hence Kobe’s deflated ORAPM) and against top 10 offenses, meaning the Lakers would have naturally under-performed on D (hence the over-inflated DRAPM)?

Again, I should note that I don’t really rate DRAPM as a judge of a perimeter player’s defense, unless of course they seen as high-impact defenders, in which case I’d put stock in it. But because perimeter players tend to have far less control on the defense of their team, I find that relying on is without seriously examining the game-tape is misleading. ORAPM is also an issue, but given that perimeter players, particularly elite one’s, have a very strong control over their team’s offense, I tend to utilize it more routinely.

[JordansBulls]

Did anyone see Kobe on Jimmy Kimmel a few weeks back? He said something in the effect that he thinks Harden and company can take the league further than he did. I don't see that though.

[bmurph128]

I didn't read it, but I do think Kobe is a top 10 player

[drza]

When did the term data scientist overtake statistician anyways? Same dang thing... grump grump grump.

If you could eminence (since D-Nice went MIA) please extrapolate on the below re the computation of RAPM and how sample sizes impact it:

Take for example Kobe’s NPI single-season RAPM in 2006-2007. Kobe’s ORAPM was 2.35, and DRAPM was 0.2. His ORAPM was almost half of the previous season, but his DRAPM was now in the positives as compared to being negative in 06. Kobe played a total of 3140 minutes, and sat 866 minutes. Of those 866 minutes, 240 (~28%) were the result of Kobe missing 5 full games during the RS. The Lakers in those 5 games had the following ORTGs/DRTGs (will note where either Lamar Odom or Luke Walton, or both of them played in these games because they were the two best Laker players):


• Game 01: Suns: 113.9 (1ST in league)/106.4 DRTG (13th in league)- 114-106 Win- LO & Walton in
• Game 02: Warriors: 107.0 (10th in league)/107.4 DRTG (19th in the league)- 110-98 win- LO & Walton in
• Game 19: Hawks: 103.0 (29TH in league)/108.3 DRTG (23RD in the league)- 106-95 Win- Lamar in, Walton Out
• Game 45: Knicks: 105.7 (17th in league)/108.8 DRTG (25th in the league)- 94-99 Loss- LO in and Walton Out
• Game 62: Bucks 107.0 (Tied for 10th )/ 111.7 DRTG (29th in the league)- 99-110 Loss- LO & Walton out
• Average ORTG of opponents= 107.32 (9th in the league)
• Average ORTG of opponents= 108.52 (23rd in the league)

On average, the Lakers were playing very solid offenses and atrocious defenses in the games that Kobe missed. Logically then, the Lakers would have over performed on offense (even without Kobe, as LO/Luke were rolling earlier on in the season) because of the atrocious defenses they faced, and under-performed defensively because of the very solid offenses they were facing in these games.

Given that about 28% of Kobe’s “off’ minutes occurred during this sample, how much of the above fed into Kobe’s RAPM? Would there likely have been an internal adjustment to take into account the specific quality of the teams faced during the games Kobe missed? Or would it have fed it uncritically? I don't get the process being the computation so please do explain, as I understand RAPM is only useful if you have "off" data (hence if RAPM existed in Wilt's day, it would be utterly useless in determining his value)

I note again that Kobe’s ORAPM was roughly cut in half from 06, whist his DRAPM almost doubled. I find this very suspicious, because 07 Kobe was slower, less laterally quick and less in shape in 07 (a lot of it the result of his off-season surgery), and got lit up even in high profile match-ups- ones he usually would get up to (see Michael Redd/Gilbert Arenas as examples). And also, Kobe was only marginally less dominant on offense in 07 (his slashing game weakened, but his jump shot was at its highest, his play-making higher and he was generally more efficient).

It is unfathomable to have watched the games and conclude that 06 Kobe was a negative defender and then in the same breath suggest 07 was a positive defender. Yet DRAPM suggests this. My tingling sensations asks, whether the fact that 28% of Kobe’s total ‘off’ minutes happened to occur against bottom 10 defenses, meaning the Lakers as a collective would have over-performed on O (hence Kobe’s deflated ORAPM) and against top 10 offenses, meaning the Lakers would have naturally under-performed on D (hence the over-inflated DRAPM)?

Again, I should note that I don’t really rate DRAPM as a judge of a perimeter player’s defense, unless of course they seen as high-impact defenders, in which case I’d put stock in it. But because perimeter players tend to have far less control on the defense of their team, I find that relying on is without seriously examining the game-tape is misleading. ORAPM is also an issue, but given that perimeter players, particularly elite one’s, have a very strong control over their team’s offense, I tend to utilize it more routinely.

My (very broad, very brief) 2 cents is that there's a more likely reason for the phenomenon you note in the lower two underlined sections than the 28% off argument you make...namely, your first underlined. You were using NPI RAPM. It is my opinion that NPI RAPM is not as useful of a stat, specifically because there's no priors. RAPM is influenced by the priors that you use, and a single season of data often isn't enough for the algorithm to correct itself from starting with priors of essentially 0. This was the whole reason PI RAPM came into effect, as it gives the RAPM algorithm a better guess as to the starting points.

In DocMJ's scaled PI RAPM spreadsheet, 2006 and 2007 Kobe measured out very similarly in both ORAPM and DRAPM, less than a point difference for either. I think that's likely to be much more accurate than the NPI approach, as to me NPI single season just isn't enough for the algorithm to give the most useful data.

[prejt]

Bumping this in anticipation of providing good information to some people before the top 100 project.

Definitely has lots of bias/agenda, but even in spite of that, still good stuff overall. I've been higher on Kobe the last month or two than I ever have, and this helps reinforce some of it. Still don't think he's top 10, but much higher than I originally held him. Thanks for bumping it.

well where did you have him? somewhere in the top50?

[rebirthoftheM]

When did the term data scientist overtake statistician anyways? Same dang thing... grump grump grump.

If you could eminence (since D-Nice went MIA) please extrapolate on the below re the computation of RAPM and how sample sizes impact it:

Take for example Kobe’s NPI single-season RAPM in 2006-2007. Kobe’s ORAPM was 2.35, and DRAPM was 0.2. His ORAPM was almost half of the previous season, but his DRAPM was now in the positives as compared to being negative in 06. Kobe played a total of 3140 minutes, and sat 866 minutes. Of those 866 minutes, 240 (~28%) were the result of Kobe missing 5 full games during the RS. The Lakers in those 5 games had the following ORTGs/DRTGs (will note where either Lamar Odom or Luke Walton, or both of them played in these games because they were the two best Laker players):


• Game 01: Suns: 113.9 (1ST in league)/106.4 DRTG (13th in league)- 114-106 Win- LO & Walton in
• Game 02: Warriors: 107.0 (10th in league)/107.4 DRTG (19th in the league)- 110-98 win- LO & Walton in
• Game 19: Hawks: 103.0 (29TH in league)/108.3 DRTG (23RD in the league)- 106-95 Win- Lamar in, Walton Out
• Game 45: Knicks: 105.7 (17th in league)/108.8 DRTG (25th in the league)- 94-99 Loss- LO in and Walton Out
• Game 62: Bucks 107.0 (Tied for 10th )/ 111.7 DRTG (29th in the league)- 99-110 Loss- LO & Walton out
• Average ORTG of opponents= 107.32 (9th in the league)
• Average ORTG of opponents= 108.52 (23rd in the league)

On average, the Lakers were playing very solid offenses and atrocious defenses in the games that Kobe missed. Logically then, the Lakers would have over performed on offense (even without Kobe, as LO/Luke were rolling earlier on in the season) because of the atrocious defenses they faced, and under-performed defensively because of the very solid offenses they were facing in these games.

Given that about 28% of Kobe’s “off’ minutes occurred during this sample, how much of the above fed into Kobe’s RAPM? Would there likely have been an internal adjustment to take into account the specific quality of the teams faced during the games Kobe missed? Or would it have fed it uncritically? I don't get the process being the computation so please do explain, as I understand RAPM is only useful if you have "off" data (hence if RAPM existed in Wilt's day, it would be utterly useless in determining his value)

I note again that Kobe’s ORAPM was roughly cut in half from 06, whist his DRAPM almost doubled. I find this very suspicious, because 07 Kobe was slower, less laterally quick and less in shape in 07 (a lot of it the result of his off-season surgery), and got lit up even in high profile match-ups- ones he usually would get up to (see Michael Redd/Gilbert Arenas as examples). And also, Kobe was only marginally less dominant on offense in 07 (his slashing game weakened, but his jump shot was at its highest, his play-making higher and he was generally more efficient).

It is unfathomable to have watched the games and conclude that 06 Kobe was a negative defender and then in the same breath suggest 07 was a positive defender. Yet DRAPM suggests this. My tingling sensations asks, whether the fact that 28% of Kobe’s total ‘off’ minutes happened to occur against bottom 10 defenses, meaning the Lakers as a collective would have over-performed on O (hence Kobe’s deflated ORAPM) and against top 10 offenses, meaning the Lakers would have naturally under-performed on D (hence the over-inflated DRAPM)?

Again, I should note that I don’t really rate DRAPM as a judge of a perimeter player’s defense, unless of course they seen as high-impact defenders, in which case I’d put stock in it. But because perimeter players tend to have far less control on the defense of their team, I find that relying on is without seriously examining the game-tape is misleading. ORAPM is also an issue, but given that perimeter players, particularly elite one’s, have a very strong control over their team’s offense, I tend to utilize it more routinely.

My (very broad, very brief) 2 cents is that there's a more likely reason for the phenomenon you note in the lower two underlined sections than the 28% off argument you make...namely, your first underlined. You were using NPI RAPM. It is my opinion that NPI RAPM is not as useful of a stat, specifically because there's no priors. RAPM is influenced by the priors that you use, and a single season of data often isn't enough for the algorithm to correct itself from starting with priors of essentially 0. This was the whole reason PI RAPM came into effect, as it gives the RAPM algorithm a better guess as to the starting points.

In DocMJ's scaled PI RAPM spreadsheet, 2006 and 2007 Kobe measured out very similarly in both ORAPM and DRAPM, less than a point difference for either. I think that's likely to be much more accurate than the NPI approach, as to me NPI single season just isn't enough for the algorithm to give the most useful data.

Cheers on that Drza... again, respond if you have the time, but what is the methodology behind the scaled PI RAPM by DocMJ, and how does it differ to regular quoted PI RAPMs we see?

Also, as a very general question since I have little knowledge... when large parts of the 'off' category occur under should say I say 'unique' context (say over a particular stretch where X team is facing a certain quality of teams) how is this controlled for? Does the prior data control any outliers from making a large imprint? Or is there some other 'balancing' control?

[magicmerl]

If you are saying that Kobe is a top10 player all time, then the question is 'over who'?

Now, if you're pretending that NBA history starts around 1980, then Kobe definitely is a top 10 player all time. But if you're including the early days as well, including Kobe in the top10 means cutting out one of the players who is towards the bottom of the list.

For me that means Bird and Hakeem. Which of those guys should be squeezed out of the top 10 so Kobe can take their place?

[drza]

Spoiler:

If you could eminence (since D-Nice went MIA) please extrapolate on the below re the computation of RAPM and how sample sizes impact it:

Take for example Kobe’s NPI single-season RAPM in 2006-2007. Kobe’s ORAPM was 2.35, and DRAPM was 0.2. His ORAPM was almost half of the previous season, but his DRAPM was now in the positives as compared to being negative in 06. Kobe played a total of 3140 minutes, and sat 866 minutes. Of those 866 minutes, 240 (~28%) were the result of Kobe missing 5 full games during the RS. The Lakers in those 5 games had the following ORTGs/DRTGs (will note where either Lamar Odom or Luke Walton, or both of them played in these games because they were the two best Laker players):


• Game 01: Suns: 113.9 (1ST in league)/106.4 DRTG (13th in league)- 114-106 Win- LO & Walton in
• Game 02: Warriors: 107.0 (10th in league)/107.4 DRTG (19th in the league)- 110-98 win- LO & Walton in
• Game 19: Hawks: 103.0 (29TH in league)/108.3 DRTG (23RD in the league)- 106-95 Win- Lamar in, Walton Out
• Game 45: Knicks: 105.7 (17th in league)/108.8 DRTG (25th in the league)- 94-99 Loss- LO in and Walton Out
• Game 62: Bucks 107.0 (Tied for 10th )/ 111.7 DRTG (29th in the league)- 99-110 Loss- LO & Walton out
• Average ORTG of opponents= 107.32 (9th in the league)
• Average ORTG of opponents= 108.52 (23rd in the league)

On average, the Lakers were playing very solid offenses and atrocious defenses in the games that Kobe missed. Logically then, the Lakers would have over performed on offense (even without Kobe, as LO/Luke were rolling earlier on in the season) because of the atrocious defenses they faced, and under-performed defensively because of the very solid offenses they were facing in these games.

Given that about 28% of Kobe’s “off’ minutes occurred during this sample, how much of the above fed into Kobe’s RAPM? Would there likely have been an internal adjustment to take into account the specific quality of the teams faced during the games Kobe missed? Or would it have fed it uncritically? I don't get the process being the computation so please do explain, as I understand RAPM is only useful if you have "off" data (hence if RAPM existed in Wilt's day, it would be utterly useless in determining his value)

I note again that Kobe’s ORAPM was roughly cut in half from 06, whist his DRAPM almost doubled. I find this very suspicious, because 07 Kobe was slower, less laterally quick and less in shape in 07 (a lot of it the result of his off-season surgery), and got lit up even in high profile match-ups- ones he usually would get up to (see Michael Redd/Gilbert Arenas as examples). And also, Kobe was only marginally less dominant on offense in 07 (his slashing game weakened, but his jump shot was at its highest, his play-making higher and he was generally more efficient).

It is unfathomable to have watched the games and conclude that 06 Kobe was a negative defender and then in the same breath suggest 07 was a positive defender. Yet DRAPM suggests this. My tingling sensations asks, whether the fact that 28% of Kobe’s total ‘off’ minutes happened to occur against bottom 10 defenses, meaning the Lakers as a collective would have over-performed on O (hence Kobe’s deflated ORAPM) and against top 10 offenses, meaning the Lakers would have naturally under-performed on D (hence the over-inflated DRAPM)?

Again, I should note that I don’t really rate DRAPM as a judge of a perimeter player’s defense, unless of course they seen as high-impact defenders, in which case I’d put stock in it. But because perimeter players tend to have far less control on the defense of their team, I find that relying on is without seriously examining the game-tape is misleading. ORAPM is also an issue, but given that perimeter players, particularly elite one’s, have a very strong control over their team’s offense, I tend to utilize it more routinely.

My (very broad, very brief) 2 cents is that there's a more likely reason for the phenomenon you note in the lower two underlined sections than the 28% off argument you make...namely, your first underlined. You were using NPI RAPM. It is my opinion that NPI RAPM is not as useful of a stat, specifically because there's no priors. RAPM is influenced by the priors that you use, and a single season of data often isn't enough for the algorithm to correct itself from starting with priors of essentially 0. This was the whole reason PI RAPM came into effect, as it gives the RAPM algorithm a better guess as to the starting points.

In DocMJ's scaled PI RAPM spreadsheet, 2006 and 2007 Kobe measured out very similarly in both ORAPM and DRAPM, less than a point difference for either. I think that's likely to be much more accurate than the NPI approach, as to me NPI single season just isn't enough for the algorithm to give the most useful data.

Cheers on that Drza... again, respond if you have the time, but what is the methodology behind the scaled PI RAPM by DocMJ, and how does it differ to regular quoted PI RAPMs we see?

Also, as a very general question since I have little knowledge... when large parts of the 'off' category occur under should say I say 'unique' context (say over a particular stretch where X team is facing a certain quality of teams) how is this controlled for? Does the prior data control any outliers from making a large imprint? Or is there some other 'balancing' control?

I'll answer the second question first. It seems, from your question, that you see the RAPM calculation as an extension of on/off +/-, but it's not really that. The APM calculation takes as inputs every 5-man line-up on a given team, the 5-man line-up that they were facing at all times, and the point differential during each of those time chunks. The regression for a given player, then, only includes "on" times for that player, as a direct measure. Where the "off" part would factor in would be secondary...how 5 man units without him did against other 5-man units would shape those player's regressed values, which would in-turn affect that player's regression in line-ups when he played with those guys.

You may have already understood that, but I wanted to start there because that process does take the opponent's strength into account. Remember...it's all 5-man units against 5-man units. So, getting back to your question, if one player is "off" against bad teams, then those bad teams would show up as weaker competition, and thus the 5-man units (without the main guy) wouldn't get any bonus for how they perform against them because it'd be informed that they were a weaker unit. Make sense?

Next: prior data, outliers, etc. Another key thing to remember is that the APM calculation, if it had sufficient data, would return a more accurate value than the RAPM calculation. But, a single year isn't really enough for APM accuracy...and small minute players can cause big changes in the regression, which isn't good. So, ridge regression is used (making it now RAPM) to smooth out those potential outliers. The method inherently distrusts outliers, and attempts to bring them back to the mean. This is usually a positive, for the reason I mentioned of the small minute guy outliers. However, it does allow for the possibility of a main player actually being a huge outlier that the regression minimizes. That's where repetitive measurements come in, to help see a better signal emerge over time.

As far as prior data, yes, it is important. The way I think of it, is, the prior information gives the algorithm a bit of a head start towards solving these thousands of simultaneous problems. A prior of essentially zero, setting everyone at initially the same level (npi RAPM) isn't quite enough, and the algorithm can struggle. Using previous year RAPM as a prior gives the algorithm reasonable neighborhoods for starting points of different players, and the algorithm is able to use that to come to more stable answers.

This leads to potential trade-offs in approaches:

*Some want to know only what a player did in ONE season, with no other info. They tend to want to use npi RAPM, but as I said, it doesn't give good enough results for me to have any confidence in it.

*The PI-RAPM could be argued to be letting info from one season influence the next, but I think it's still reasonable. There's still plenty of room for the algorithm to change values from year to year. And I think it's rare for player value to change SO much from year to year that it'd be more accurate to start from zero and have all types of noise. So, for me, the PI-RAPM approach is better.

Doc MJ's scaled RAPM approach: Before (I believe) the last Top 100 project, Doc MJ collected the best PI-RAPM studies he could find for every year that was published and put it in one place. That was the main value of his work, for me...taking info that was found on several different sites, some of which that closed down shop (as their owners were hired, often by NBA front offices or ESPN), and putting it all in one place.

But, he went further. Remember above, when I pointed out that APM (if there were enough data) is actually more accurate than RAPM? The upshot of that, is, an APM calculation gives an actual estimate for how much a player changes a given team's scoring margin. RAPM doesn't give exactly that...it's a bit scaled in each calculation. Thus, looking at RAPM scores from one year's calculation doesn't necessarily operate on the same scale as RAPM scores from a different year's calculation. Doc MJ took that concept, then, and wanted to convert RAPM data into an estimate of how much a player changes the scoring margin like APM. To that end, he:

*Used a multi-year APM model as a template to estimate the relationship between standard deviation and APM score
*He then took every single-year, RAPM study for every player, and calculated the standard deviation for that year and converted each RAPM score to a standard deviation
*Finally, he multiplied the standard deviation score for each single-year RAPM entry by the APM multiplier he found in part 1

His result, as he styled it, put each RAPM study on the same scale. So, you could look at a 2000 RAPM study and a 2010 study, and compare the scaled values directly and have them be on the same scale, as an estimate of how much an individual player was actually changing the team's scoring margins.

Sorry was kind of long. Hope it helps

[rebirthoftheM]

^ Thanks for clarifying that.. the Scaled RAPM stuff now makes a lot of sense. I also didn't really get the real distinction between APM and RAPM (well I knew RAPM adjusts in many ways) but yeah very information:

And re the clarification re RAPM: I generally things are adjusted for the 'bad team' effect (although little knowledge of how this works out). What I was moreso trying to get as is, does it account for the fact that the same team can be elite on one end, below average to bad on another, and still be great in terms of SRS/Win Record (take Nash's Suns for e.g.)?

So for example... if X player on Team Y missed a game v Nash's Suns. Because the Suns are horrible defensively, Team Y still has a strong chance of producing above average offense. Does RAPM take this into account?

[enko]

Great stuff OP.