RealGM Top 100 List #4

[ceiling raiser]

We're at an impasse because what my quite says already addresses your concern.

If you wanted to say that the sample size if basketball is just not sufficient in your assessment to glean enough signal from the noise that would be a valid concern. however when you literally just point out that two players' impacts can get confused in theory by the mechanism all you're doing is saying correlation is not causation.

Doc, my issue is with the fundamental methodology of RAPM itself. The very samples used to calculate it, have no distinctions between players. It's not about the size of the sample, but instead the misinterpretation of what they entail.

How do we attribute individual impact to a stat that does nothing to account for the individual? RAPM is a value that represents the relative success of various rotations, nothing more. At no point has the "individual impact" element been explained. RAPM backers tend to say it quantifies what the box score can't....but of course, RAPm is actually a box score stat itself. All of its data comes from the same place as PER or WS. It's just manipulated in a vastly different manner. And that data represents an entire lineup, instead of an individual.

Jason Collins in 2005 had the best DRAPM at 6.3, what can we glean from this. Is it noise? No, it's simply how the rotations he was in performed, nothing more really. How does that 6.3 speak to his individual impact? There has never been an explanation for this. Because again, how does the 2009 DPOY carrying a non-defensive Orlando sqaud to the #1 DRtg...only get a 2, and fall behind Rashard Lewis. RAPM has never shown actual correlation to impact.

Huh? That's news to me. The data used in calculating RAPM comes from parsing play-by-play data. Last I checked, box scores do not contain complete play-by-play logs.

PER and WS are calculated by manipulating numbers in box scores. No attention is paid to the play-by-plays.

[ronnymac2]

Maybe I'm wrong and thats why it got ignored, and if so then I will just shut up, but for me a guy who even at his peak couldn't shoot better than 43% from outside 3 feet (one natural step for him) of the basket doesn't have a whole lot of true basketball skills. Maybe he could make up for that somewhere else, but he didn't- never really a great passer, on/off defense but again can be attributed to his size, terrible ft shooter… essentially any skill that his size couldn't dominate he was okay at best. If this is true for Wilt, please tell me. I have not been able to find it but again, as I said before, he could be a damn good passer when he wanted to be and had the athleticism to do other basketball things than just dunk.

1. Shaq was a great passer.

2. You are penalizing the man for having a physical advantage over the opponent, which he used correctly for the betterment of his team.

3. If all Shaq did was dunk and still produced what he did, what does it matter? Nevermind the fact that he did not just dunk all the time, but instead had a light touch around the rim, a left hand, a turnaround jumper that at his peak was effective from 13ish feet and in, effective even if simple footwork, and a reliable jump hook.

[An Unbiased Fan]

We're at an impasse because what my quite says already addresses your concern.

If you wanted to say that the sample size if basketball is just not sufficient in your assessment to glean enough signal from the noise that would be a valid concern. however when you literally just point out that two players' impacts can get confused in theory by the mechanism all you're doing is saying correlation is not causation.

Doc, my issue is with the fundamental methodology of RAPM itself. The very samples used to calculate it, have no distinctions between players. It's not about the size of the sample, but instead the misinterpretation of what they entail.

How do we attribute individual impact to a stat that does nothing to account for the individual? RAPM is a value that represents the relative success of various rotations, nothing more. At no point has the "individual impact" element been explained. RAPM backers tend to say it quantifies what the box score can't....but of course, RAPm is actually a box score stat itself. All of its data comes from the same place as PER or WS. It's just manipulated in a vastly different manner. And that data represents an entire lineup, instead of an individual.

Jason Collins in 2005 had the best DRAPM at 6.3, what can we glean from this. Is it noise? No, it's simply how the rotations he was in performed, nothing more really. How does that 6.3 speak to his individual impact? There has never been an explanation for this. Because again, how does the 2009 DPOY carrying a non-defensive Orlando sqaud to the #1 DRtg...only get a 2, and fall behind Rashard Lewis. RAPM has never shown actual correlation to impact.

Huh? That's news to me. The data used in calculating RAPM comes from parsing play-by-play data. Last I checked, box scores do not contain complete play-by-play logs.

PER and WS are calculated by manipulating numbers in box scores. No attention is paid to the play-by-plays.

Play by Play data only tells you how the box score numbers accumulated over the course of a game, there's nothing really magical about it. What RAPM does is parse out lineup data sets from it.

[Basketballefan]

Runoff Vote: Wilt

As mentioned the attitude/work ethic thing is a wash so i go with wilt because i think he was a better defender, rebounder, and passer even though Shaq might've been more reliable in the playoffs. Shaq's competition as far as centers was questionable when he actually was winning, Hakeem, Ewing and Drob were all well past their prime, so what great center did he have to compete with? The amount of records Wilt Broke was crazy. i will list some.


MINUTES
Most seasons leading league: 8
Most consecutive seasons leading league: 5
Highest average per game, career: 45.8
Highest average per game, season: 48.5

SCORING
Most cons. seasons leading league: *7
Most points, season: 4,029
Highest average, season: 50.4
Most points, rookie, season: 2,707
Highest average, rookie, season: 37.6
Most points, game: 100
Most points, rookie, game: 58
Most games, 50 or more, career: 118


REBOUNDS
Most seasons leading league: 11
Most, career: 23,924
Highest average, career: 22.9
Most, season: 2,149
Most, rookie, season: 1,941
Most seasons, 1,000 or more: 13
Highest average, season: 27.2
Most, game: 55
Most, rookie, game: 45
Most games, 50 or more, season: 45
Most cons.e games, 50 or more points: 7
Most games, 40 or more, career: 271
Most games, 40 or more, season: 63
Most cons. games, 40 or more points: 14
Most cons. games, 30 or more points: 65
Most cons. games, 20 or more points: 126
Most points, one half: 59

REBOUNDS, PLAYOFFS
Most, 5-game series: 160
Most, 6-game series: 171
Most, 7-game series: 220
Highest average, per game, series: 32.0
Most, game: 41
Most, half: 26
Most by rookie, game: 35

And there are many others. Let's not forget 50 ppg and 26 rpg in a season, obviously that wouldn't happen today but still a great feat nonetheless. Even though shaq has an edge in longevity i think Wilt actually aged better and was the better rebounder and defender in his twighlight, shaq was the better scorer at that point but his D was average and his rebounding wasn't elite anymore. Shaq's last few seasons might as well have not even happened. Wilt was more durable throughout his career as well.

[ceiling raiser]

Play by Play data only tells you how the box score numbers accumulated over the course of a game, there's nothing really magical about it. What RAPM does is parse out lineup data sets from it.

Play-by-play tells you "how" everything was achieved during a game. All a box score does is tell you "what" was achieved during a game. Colossal difference between the two.

This is the entirety of the input data for the calculations of PER/WS:

http://bkref.com/tiny/8oO81

4 or 5 pages worth. Probably a few kb worth of data.

Here are the input files for the same year for RAPM:

phase 1 (total RAPM): http://www34.zippyshare.com/v/85123463/file.html
phase 2 (off/def splits): http://www34.zippyshare.com/v/82178248/file.html

Each file is a~25MB csv. Massive, massive, difference in the volume of data used (and to find prior-informed RAPM, you use a vector of the coefficients from the previous season's results, so the size of the input is astronomical compared to that of PER/WS...note that the SportsVU dataset is much, much bigger still ).

RAPM doesn't parse out lineup data sets. The parsed lineup data IS the input for RAPM. RAPM seeks to solve a system of equations for coefficients (i.e. the RAPM values) that minimize norm in the L2 sense, though since the number of players involved doesn't match the number of equations, ridge regression is used to simulate the OLS solving mechanism.

Yes, you can extract box score numbers from play-by-play. But "what" is light years different than "how".

[90sAllDecade]

This was a little tougher than I initially thought as both were similar players with similar weaknesses and issues.

Shaq went against modern competition and performed versus Wilt and although Wilt was likely better defensively, he had distinct rule and competition advantages over Shaq inflating this impact. The floor spacing and other arguments goes both ways and is a disadvantage to modern center's defensively.

Wilt had eye popping numbers and could dominate whatever statistical category he chose, be it assists, FG%, offense or defense which I have to appreciate. But all those categories got worse under pressure, with the exception of rebounding even his FT% got worse in the playoffs; it was clearly mental toughness with Wilt.

His hallmark and biggest credit, raw individual dominance, was consistently reduced under pressure. Who could truly stop him in that era? No one could really stop Wilt back then except himself imo, even Russell and Thurmond could only contain him individually.

I value mental toughness and although imo Shaq did the same by also underachieving at times and lack of effort in discipline, Wilt would stop himself mentally more often.

I voted Hakeem as better than both and after holding a microscope to these players I still truly believe that, but if I had to choose only between these two; my runoff vote is Shaq.

[An Unbiased Fan]

Play by Play data only tells you how the box score numbers accumulated over the course of a game, there's nothing really magical about it. What RAPM does is parse out lineup data sets from it.

Play-by-play tells you "how" everything was achieved during a game. All a box score does is tell you "what" was achieved during a game. Colossal difference between the two.

That's the same thing I said, it tells you how the box score number accumulated over the course of the game. What RAPM is gleaning from PbP is how the +/- was for various lineups.

The very base of data used is no different.

RAPM doesn't parse out lineup data sets. The parsed lineup data IS the input for RAPM.

We're saying the same thing here. RAPM parses out lineup data from the PbP, and then uses it as input.

The parsed lineup data IS the input for RAPM. RAPM seeks to solve a system of equations for coefficients (i.e. the RAPM values) that minimize norm in the L2 sense, though since the number of players involved doesn't match the number of equations, ridge regression is used to simulate the OLS solving mechanism.

Yes, you can extract box score numbers from play-by-play. But "what" is light years different than "how".

RAPM never touches the "how". It's all about +/- manipulation. This goes back to lineup data being used as the base input. There is no mechanism within RAPM calculation that separates the individual from the lineup. There is no "how", only "what" all 5 guys did on court in relation to the opposing 5 guys.

There is no model that will glean impact from lineup data, that's the problem, and that's why the results don't correlate to actual impact. Lineup data is collective, there is no mathematically way possible to seperate the group. You would need individual data samples, (like something from synergy) as a base instead. Or extract a whole different data set from the PbP data that wasn't lineup based.

[ceiling raiser]

RAPM never touches the "how". It's all about +/- manipulation. This goes back to lineup data being used as the base input. There is no mechanism within RAPM calculation that separates the individual from the lineup. There is no "how", only "what" all 5 guys did on court in relation to the opposing 5 guys.

There is no model that will glean impact from lineup data, that's the problem, and that's why the results don't correlate to actual impact. Lineup data is collective, there is no mathematically way possible to seperate the group. You would need individual data samples, (like something from synergy) as a base instead. Or extract a whole different data set from the PbP data that wasn't lineup based.

1) That's exactly what RAPM is though...you're given a system of equations (the equations being the lineups for both teams), and ridge regression finds the which coefficients for the individual variables (which correspond to the player) find the optimal solution to the problem. The only issue is when you have players who are on the court together very often, but using a prior can add possessions during which they are not on the floor together. If your criticism is collinearity, I can respect that 100%.

2) How can the results correlate or not correlate to impact, when it's a fuzzy term? There is no "impact" stat, by Merriam-Webster's definition, to have "impact" is "to have a strong and often bad effect on (something or someone)". Finding the individual coefficients that solve the group problem (with lineup data as an input) is isolating an individual's effect on the scoring margin. You don't have to call that impact, but I think it approximates the concept very closely.

3) That's exactly what RAPM is. You're solving a system of equations (groups, or lineups) for a vector of coefficients (individual) that minimizes the problem. Again though, there's no reason to compare players who are doing different things on the floor. Compare guys who are playing in similar roles (use things such as the eye test and box scores to ascertain roles).

I know I've said this a few times, but I don't want to derail the thread any further. I feel like we're going back-and-forth about the same points. You're certainly entitled to your opinion and I can respect that, but I felt the need to respond one last time.

[Dr Positivity]

My run off vote is for Shaquille O'Neal, I feel his game is more simplified and "knows what it is" and I value that. He also has the insane Finals performances in his favor.

Extremely minimal analysis. I will count this one but we are looking for more. Remember that this project is about the discussion, not just vote counting.

[An Unbiased Fan]

RAPM never touches the "how". It's all about +/- manipulation. This goes back to lineup data being used as the base input. There is no mechanism within RAPM calculation that separates the individual from the lineup. There is no "how", only "what" all 5 guys did on court in relation to the opposing 5 guys.

There is no model that will glean impact from lineup data, that's the problem, and that's why the results don't correlate to actual impact. Lineup data is collective, there is no mathematically way possible to seperate the group. You would need individual data samples, (like something from synergy) as a base instead. Or extract a whole different data set from the PbP data that wasn't lineup based.

1) That's exactly what RAPM is though...you're given a system of equations (the equations being the lineups for both teams), and ridge regression finds the which coefficients for the individual variables (which correspond to the player) find the optimal solution to the problem. The only issue is when you have players who are on the court together very often, but using a prior can add possessions during which they are not on the floor together. If your criticism is collinearity, I can respect that 100%.

2) How can the results correlate or not correlate to impact, when it's a fuzzy term? There is no "impact" stat, by Merriam-Webster's definition, to have "impact" is "to have a strong and often bad effect on (something or someone)". Finding the individual coefficients that solve the group problem (with lineup data as an input) is isolating an individual's effect on the scoring margin. You don't have to call that impact, but I think it approximates the concept very closely.

3) That's exactly what RAPM is. You're solving a system of equations (groups, or lineups) for a vector of coefficients (individual) that minimizes the problem. Again though, there's no reason to compare players who are doing different things on the floor. Compare guys who are playing in similar roles (use things such as the eye test and box scores to ascertain roles).

I know I've said this a few times, but I don't want to derail the thread any further. I feel like we're going back-and-forth about the same points. You're certainly entitled to your opinion and I can respect that, but I felt the need to respond one last time.

Fair enough, I think we've said enough for this thread.

I will say though, that ridge regression is incapable of separating the individual from the lineup data. There in lies the rub, its just not mathematically possible. RAPM results correlate to rotations, and does so quite nicely. But at no point has a correlation to individual impact ever been shown..because it's not possible with that data set.

[john248]

This project would benefit with a discussion thread. This is mainly in regards to the current advanced stats discussion that is going on while the runoff vote is happening. Just thinking out loud.

[HeartBreakKid]

I believe Wilt is up by six votes since the run-off has started. 20-14 from a quick count.

[Basketballefan]

I believe Wilt is up by six votes since the run-off has started. 20-14 from a quick count.

So would Oneal have to surpass the accumulated total to win i'm not sure how that works?

[HeartBreakKid]

I believe Wilt is up by six votes since the run-off has started. 20-14 from a quick count.

So would Oneal have to surpass the accumulated total to win i'm not sure how that works?

Pretty sure at this point it is just head to head between these two, who ever has the most votes totaled in a day wins. Ie, if Shaq has just one more vote than Wilt, he would win.

I could be wrong, but pretty sure that's how it went before and the only way it makes any sense to me.

[therealbig3]

I have it 19-15 Wilt so far.

GC Pantalones, Texas Chuck, Baller2014, Doctor MJ, ronnymac2, drza, 90sAllDecade, and Dr Positivity changed their votes to Shaq...7+8 = 15.

Clyde Frazier and Basketballefan changed their votes to Wilt...17+2 = 19.

[Dr Positivity]

I believe Wilt is up by six votes since the run-off has started. 20-14 from a quick count.

So would Oneal have to surpass the accumulated total to win i'm not sure how that works?

Pretty sure at this point it is just head to head between these two, who ever has the most votes totaled in a day wins. Ie, if Shaq has just one more vote than Wilt, he would win.

I could be wrong, but pretty sure that's how it went before and the only way it makes any sense to me.

Just to point out the old votes still count, original Wilt and Shaq supporters don't have to reaffirm.

[Dr Positivity]

I believe there was 40 original votes by the way

[Ryoga Hibiki]

So in the case of the 2009 Magic, it does seem strange to see Lewis ranked higher than Howard. But again: RAPM traces team scoring trends, NOT the players goodness at something. And as you look closer, you see that the Van Gundy defense in Orlando relied a lot on having a long, mobile PF that could press out to the 3-point line. Plus, with Howard and Gortat as the two centers, the PF didn't have to be that good at defensive rebounding...his mobility, length and effort would be more important. This role played right to Lewis' strengths, so it's not shocking that Lewis might have been a person of interest in that particular defense. However, Lewis clearly never replicated that level of impact in any other system, so it is obvious that his positive results were tied very specifically to that team situation.

Howard, on the other hand, had never measured out as a big impact defender before 2009 anyway (which makes sense for a young big that relied on his athleticism and may not have understood defense yet). Howard's defensive RAPM in 2009 was a high-water mark at the time, but his defensive RAPM got progressively larger from 2009 to 2010 to 2011 to 2012. In those last two years he measured out as among the best marks in the league. Not coincidentally, Lewis was gone in those two years to be replaced minutes-wise by Brandon Bass (2011) and Ryan Anderson (2012), neither of whom had the same length/mobility that Lewis had. Perhaps this required Howard to have to do more on defense to make up for slower-footed 4s, and this reflected in the defense tracking more closely with Howard? Or, maybe Howard was just getting better as a defender each year as he approached his peak, which also makes sense.

Anyway, Howard clearly measures out on the whole as an excellent defender using RAPM (though never really to the level of a multiple DPoY...this could be worth further discussion when Howard really enters the conversation) and Lewis clearly doesn't, outside of this one specific case that played right to his strengths. Plus, finding one year counterintuitive results doesn't in any way address Kobe's ultra-consistent lack of measured defensive impact over a 15 year sample. I mean, was he NEVER in a position to make a measurable defensive mark? While the vast majority of the actually elite defensive wings (that he was consistently beating out for defensive honors) did, repeatedly? This doesn't strike you as an odd occurrence that goes well beyond coincidence? Lightening happened to strike the same place 15 times in a row?

Be careful using RAPM with Howard in his last years in Orlando. It was more true in 2010, but he ws never playing with Gortat and Gortat was almost always his backup. That makes the regression useless to compare them with other players.

[lorak]

This is completely correct.

There has not been an "abstract call for rebellion" against impact oriented analysis. I'm not even sure what that phrase means. Impact on this forum has come to mean figuring at how much a player has improved their team. Essentially everyone is interested in that. The majority of posts I've read in multiple threads have been good faith attempts to partly answer this question.

What is rejected by many, including myself, is the assertion that the on/off family of stats are where you go to measure impact and that the counting category of stats are a distraction.There has been a subterranean, and occasionally overt, assertion that once you sufficiently measure on/off you have determined how much impact a player has.

There are two main family of stats: on/off and counting. IT has not been conclusively established that on/off comes close to completely measuring impact and accordingly it is inaccurate to label on/off as impact.

No need for you to explain all of that, though. It's why, despite already having access to terabytes of on/off data, the NBA has spent boatloads of cash on technology that seeks to count and quantify every single action that takes place on the court.

Actually that's not why they do it. Single metric like RAPM describes impact, but doesn't say HOW it was achieved. That's why teams want more detailed information, because when you are building a roster, try to improve Xs & Os, you need to know such details. But if you just want to know how valuable player is, and you don't care why he is so valuable, then there's no need to use tracking data. Of course we rather should care about it, because it gives us better description of players.

[90sAllDecade]

Hakeem is the only player in NBA history to win a championship without an all star, HOF talent or elite/GOAT level coach.

So Rick Barry 1975 - was al attles a hof coach?

I enjoyed researching this and I have a new respect for Rick Barry.

It's a semantic debate, as he didn't have a HOF coach or an all star, but he did have a HOF talent.

Eventual HOFer Jamal Wilkes won NBA Rookie of the Year in 1975 playing with Rick Barry, All Rookie 1st team and was a very valuable contributor in both the RS and PO. Arguably GSW second best player and offensive threat next to Barry that run.

Although he struggled offensively in the finals, he still had a great RS for a rookie and great playoff run in the rounds before that.
http://www.basketball-reference.com/pla ... eja01.html

Hakeem had All Rookie 2nd team Horry, no HOF talent, no elite/HOF coach or player who made the all star game that year.

Olajuwon was also the clear #1 offensive and defensive anchor. I may be wrong, but it looks like Rick Barry wasn't known as a dominant defender and Clifford Ray or George Johnson were the likely defensive anchors on that team.

So, I still have to say Hakeem is the only player in NBA history to ever win a championship without that team support, Jordan and all our other top players included.