1993 MVP and best player

[fpliii]

let's compare dominique to jordan... they're both in similar roles as wing players that are the primary go-to option for their teams.

the rapm data says dominique had a bigger offensive impact for his team in 1993, than jordan had for his team.

which may or may not be true... but does that mean dominique would have a bigger impact on the bulls than jordan had?

does that mean dominique is a better offensive player?

see, for this reason, the eye test is more reliable and will yield the better result when making decisions than rapm would.

That's not real RAPM data in 1993. It's a regression based on RAPM, from box score stats. The first season for which we have real RAPM is 96-97. The data doesn't exist to compute it beforehand.

Also, as I said before, there are separate calculations for total RAPM, and splits. The totals are calculated first, splits after. If guys are playing similar offensive roles, then from 96-97 on we can compare them in normalized offensive RAPM. If they're playing similar two-way roles, we can compare them in normalized total RAPM.

[fpliii]

Here is the thread on APBRmetrics where J.E. (the owner of the site stats-for-the-nba.appspot.com, and same guy who does RPM for ESPN) discusses fake RAPM:

http://apbr.org/metrics/viewtopic.php?t=8067

He does not have any digitalized play-by-play (which you need to compute RAPM) prior to 00-01, and NOBODY has it before 96-97.

[3ball]

rapm seems like it could work well for gambling purposes perhaps, or to help give guidance on optimal lineups for a coach to use, but i don't see how it has any value in determining if player A is better than player B.

the only way to determine whether one player is better than another by using rapm, is to have lineup
data for BOTH players on EACH team.

we'd need data for dominque as a bull, and jordan as a hawk.

so we'd need data for dominique in all the various bulls lineups against all the lineups from around the league, and data for jordan in all the various hawk lineups against all the lineups from around the league.
.

[fpliii]

rapm seems like it could work well for gambling purposes perhaps, or to help give guidance on optimal lineups for a coach to use, but i don't see how it has any value in determining if player A is better than player B.

the only way to determine whether one player is better than another by using rapm, is to have lineup data for BOTH players on EACH team.

so we'd need data for dominique as a bull, and data for jordan as a hawk.

Well, it depends how you define better. It tells us how good guys are at impacting scoring margin in their roles for a season.

We don't need data for guys in other teams, since it uses a regression to isolate players from teammates, and opponents faced. Again though as I said, we do not have actual RAPM before 96-97. The numbers linked are fake RAPM from the 90s, which are just a box score calculation. EDIT: I would advise ignoring them completely (which I do).

[3ball]

Well, it depends how you define better. It tells us how good guys are at impacting scoring margin in their roles for a season.

We don't need data for guys in other teams, since it uses a regression to isolate players from teammates, and opponents faced. Again though as I said, we do not have actual RAPM before 96-97. The numbers linked are fake RAPM from the 90s, which are just a box score calculation. EDIT: I would advise ignoring them completely (which I do).

i'm looking at the screamingacrossthecourt RAPM data for 1997, and it says christian laettner is the best player in the league, shaq had half the offensive impact of jordan, and terry mills had a GREATER offensive impact than jordan.

there are all kinds of problems with this data, and i think it all comes back to sample size... it is impossible to get reliable sample sizes with this methodology.

[fpliii]

i'm looking at the screamingacrossthecourt RAPM data for 1997, and it says christian laettner is the best player in the league, shaq had half the offensive impact of jordan, and terry mills had a GREATER offensive impact than jordan.

there are all kinds of problems with this data, and i think it all comes back to sample size... it is impossible to get reliable sample sizes with this methodology.

1) 1997 is NPI (non-prior informed), probably better to look at his numbers for 98, 99, 00, and make sure it's the prior-informed list. NPI is okay, but doesn't have the same predictive value.

2) You need to compare players in similar roles. If a guy shows up towards the top and isn't a superstar, all it means is that he's doing a lot of things that don't show up in the box score that help his team. Laettner played around 3500 minutes combined between the season and playoffs which is enough to stabilize for a single season (2500), though ideally you want a guy who played 5000 minutes combined in a given year plus the previous year.

3) As I noted, there are separate regressions. The initial regression finds total RAPM, and a second regression can give you RAPM offense/defense splits.

[fpliii]

Anyhow 3ball I hope these answers help. You can't compare MJ to Mills or Shaq to Laettner directly because they played different roles.

Have to run now, I'll check the thread later.

[3ball]

as I said, we do not have actual RAPM before 96-97. The numbers linked are fake RAPM from the 90s, which are just a box score calculation. EDIT: I would advise ignoring them completely (which I do).

if the 1993 rapm data should be ignored completely, why are people itt using it to say hakeem was better than jordan?

We don't need data for guys in other teams, since it uses a regression to isolate players from teammates, and opponents faced.

how can you isolate a player, or determine a player's "stand-alone ability" (if you will), if a player's production (the data you track) is always based on playing alongside his teammates?

like, if a player averages 20ppg, that is based on him playing alongside the specific teammates that he played alongside... how can you somehow remove the teammates and still use the 20ppg?... the 20ppg was based on playing alongside those teammates.

can you go into a little more detail on how a regression solves this issue... because this is one of the main flaws i see (other than sample size).

[3ball]

You can't compare MJ to Mills or Shaq to Laettner directly because they played different roles.

so when i see that 1997 rapm says scottie had a bigger offensive impact than jordan, i can't compare the two because they play "different roles".

and therein lies the rub... you can't compare ANYBODY with rapm, because everyone plays "different roles" on their respective teams.

as i said, rapm might have some use in gambling, or giving coaches guidance on what lineups to use, but it can't be used to say "player A is better than player B because he has a higher RAPM."

and for the record, i think using rapm to determine optimal lineups is a bad idea, and would indicate a coach doesn't understand basketball - but i can see the how RAPM would help lineup decisions from a theoretical standpoint.. i wouldn't hate a coach for looking at rapm-based lineups that an assistant prepared or something, only as guidance though.
.

[fpliii]

[QUOTE=3ball]if the 1993 rapm data should be ignored completely, why are people itt using it to say hakeem was better than jordan?[quote]
I ignore it completely, but I also ignore PER and win shares. Like those two, it's based on box scores. It has no use to me, but others are free to use it. Note that on that site (stats-for-the-nba.appspot.com), numbers through 1999-2000 are fake RAPM. From 00-01, those are xRAPM, which is similar to RPM on ESPN's site.

how can you isolate a player, or determine a player's "stand-alone ability" (if you will), if a player's production (the data you track) is always based on playing alongside his teammates?

can you go into a little more detail on how this is done... because this is one of the main flaws i see (other than sample size).

I broke down this specific process here:

http://www.insidehoops.com/forum/sho...3&postcount=79

Basically though, each player is treated as a variable, and the regression finds the best fit. This means you're not solving for one player, but all simultaneously.

Say you have two lineups:

x1+x2+x3+x4+x5-y1-y2-y3-y4-y5=z1
x6+x2+x3+x4+x7-y1-y6-y3-y7-y8=z2

x's are home players, y's are away players, z's are the points/possessions when they're on the floor. What you're doing is finding which values of x and y best predict the values of z, when testing in out-of-sample lineups. If various lineups/matchups are on the floor for more time, those observations are weighted more. There are generally 30-35 thousand between a season and playoffs.

We use matrix algebra to find the best estimate for the x's and y's to make all of the equations fit best. Using ridge regression, we'll end up with a very good fit, with scaled numbers differing by less than one standard deviation on average.

These x's and y's only tell us how well a player is playing in his given role. For instance, Kevin Garnett was asked to do different things in Minnesota than he was in Boston, especially after he got hurt. But some of his late-career seasons, from 2010-2013, rate favorably when compared to his peak seasons from say 2003-2005. Does this mean he was better when playing more of a specialist role? No, of course not. It just means that in that role, he was helping his teams as much or more than he was earlier in his career when he was a two-way superstar, anchoring offense and defense.

If you get five specialists, who are all RAPM leaders, and put them on the floor together, you're gonna be ****ed. None of them can create a shot most likely, and none of them will demand defensive attention.

This is why you have to compare guys in similar roles to one another in RAPM.

Compare two way superstars to superstars.

Compare big man defensive specialists to big man defensive specialists.

Compare 3-and-d guys to 3-and-d guys.

Compare pass-first PGs to pass-first PGs.

This is why the eye test is important. Not only will it give you an idea of how well a guy plays, but it will help you identify his role. RAPM is the only stat I use when evaluating players, but it's coupled with watching AS MUCH TAPE AS POSSIBLE. If you have know clue of what a guy is asked to do on a team, then the RAPM numbers are useless.

[fpliii]

so when i see that 1997 rapm says scottie had a bigger offensive impact than jordan, i can't compare the two because they play "different roles".

and therein lies the rub... you can't compare ANYBODY with rapm, because everyone plays "different roles" on their respective teams... also, why can't we compare laettner to shaq?... i thought rapm isolated the player from teammates and opponents?

as i said, rapm might have some use in gambling, or giving coaches guidance on what lineups to use, but it can't be used to say "player A is better than player B because he has a higher RAPM."

and for the record, i think using rapm to determine optimal lineups is a bad idea, and would indicate a coach doesn't understand basketball - but i can see the how RAPM would help lineup decisions from a theoretical standpoint.. i wouldn't hate a coach for looking at rapm-based lineups that an assistant prepared or something, only as guidance though.
.

1) Well that's also an issue of it being NPI. If it was prior-informed, it would be more useful.

I think they are somewhat similar in terms of roles, but the total RAPM calculation is more reliable. I think Scottie rated pretty poorly in defensive RAPM, while Jordan looked awesome. Look at the totals and it makes more sense.

2) It isolates players from teammates, but you can't compare players in different roles. Look at Wilt with the Lakers under Sharman. He was playing a super defensive specialist role. But it doesn't mean that's all he was capable of those seasons. So you can't compare him to Kareem directly who was asked to anchor both his offense and his defense. Would Wilt have been able to do the same thing? Maybe, but if RAPM existed, it wouldn't help us here. It would just tell us that Wilt was playing his role better than Kareem was his.

3) RAPM doesn't tell us who is more talented, it just tells us who is better at impacting scoring margin, with both guys in their current roles. Gamblers use it a ton, front offices do in putting teams together. A coach can't look at it and just pick the best 5 RAPM guys, because you need to have different functional requirements n your lineup. i.e. you can't have 5 defensive specialists, but no ball-handlers.

RAPM is a valuable tool in predicting outcomes, player evaluations, and lineup optimization, as long as role is taken into account. Without roles, RAPM is useless.

[3ball]

[QUOTE=fpliii][QUOTE=3ball]if the 1993 rapm data should be ignored completely, why are people itt using it to say hakeem was better than jordan?

I ignore it completely, but I also ignore PER and win shares. Like those two, it's based on box scores. It has no use to me, but others are free to use it. Note that on that site (stats-for-the-nba.appspot.com), numbers through 1999-2000 are fake RAPM. From 00-01, those are xRAPM, which is similar to RPM on ESPN's site.


I broke down this specific process here:

http://www.insidehoops.com/forum/sho...3&postcount=79

Basically though, each player is treated as a variable, and the regression finds the best fit. This means you're not solving for one player, but all simultaneously.

Say you have two lineups:

x1+x2+x3+x4+x5-y1-y2-y3-y4-y5=z1
x6+x2+x3+x4+x7-y1-y6-y3-y7-y8=z2

x's are home players, y's are away players, z's are the points/possessions when they're on the floor. What you're doing is finding which values of x and y best predict the values of z, when testing in out-of-sample lineups. If various lineups/matchups are on the floor for more time, those observations are weighted more. There are generally 30-35 thousand between a season and playoffs.

We use matrix algebra to find the best estimate for the x's and y's to make all of the equations fit best. Using ridge regression, we'll end up with a very good fit, with scaled numbers differing by less than one standard deviation on average.

These x's and y's only tell us how well a player is playing in his given role. For instance, Kevin Garnett was asked to do different things in Minnesota than he was in Boston, especially after he got hurt. But some of his late-career seasons, from 2010-2013, rate favorably when compared to his peak seasons from say 2003-2005. Does this mean he was better when playing more of a specialist role? No, of course not. It just means that in that role, he was helping his teams as much or more than he was earlier in his career when he was a two-way superstar, anchoring offense and defense.

If you get five specialists, who are all RAPM leaders, and put them on the floor together, you're gonna be ****ed. None of them can create a shot most likely, and none of them will demand defensive attention.

This is why you have to compare guys in similar roles to one another in RAPM.

Compare two way superstars to superstars.

Compare big man defensive specialists to big man defensive specialists.

Compare 3-and-d guys to 3-and-d guys.

Compare pass-first PGs to pass-first PGs.

This is why the eye test is important. Not only will it give you an idea of how well a guy plays, but it will help you identify his role. RAPM is the only stat I use when evaluating players, but it's coupled with watching AS MUCH TAPE AS POSSIBLE. If you have know clue of what a guy is asked to do on a team, then the RAPM numbers are useless.

i appreciate you explaining rapm itt.

but if effective comparison of the rapm stats of two players depends on those two players playing the same role, then rapm can't be used to compare two players, since no two players play the same role for their teams.

also, garnett's predictive values when playing with teammates on OTHER teams (like the bobcats or something) are based on the values he put up alongside his actual teammates... so i don't see how the rapm data could be representative of what his impact would be if he were on ANY team - the predictive values are still based on what he did alongside his own teammates.

and if rapm predicts what garnett would do on ANY team, then why is this a yearly stat like an income statement, instead of an ongoing stat that can be measured at any time like a temperature or a balance sheet?... like, why does 1997 rapm show shaq as having half the offensive impact he had in 1998?... shouldn't such a stat remain pretty stable over the years?.. this goes back to the sample size problem from what i can see.
.

[fpliii]

1) Role doesn't need to be identical, just similar. The more similar the roles two players have, the more valid the comparison is.

2) It works to project him in other lineups assuming he plays the same role. RAPM removes the effect of teammates, but it won't tell you how a guy would play if a different team needs him to play a different role.

3) Single season NPI RAPM (which is all we have for 97) is pretty stable, but the offense defense splits are not. As I mentioned, it's a separate regression. The first season for which prior informed (NPI means non-prior informed, or meaning it doesn't use previous season data to smooth out results) RAPM exists is 98. So when comparing 98 to 99 with priors (which means instead of telling the regression to try fitting everyone around 0, it fits players based on the previous season first, in hopes of smaller error), if there is a big difference, it's probably valid.

[3ball]

Without roles, RAPM is useless.

"similar" roles is very subjective criteria that will differ a lot among whoever is analyzing it

and if "similar roles" are crucial to comparing rapm's of players, then rapm is useless when it comes to comparing players, since no two players play the same role.

i.e. you could never compare jordan's rapm to lebron's, because those guys play completely different roles for their teams... one is a ball-dominating point guard, and the other is an off-ball assassin.... two completely different roles.. i don't see how rapm could apply in comparing these two players.

how can the stat be a measure of the scoring margin impact of a player on ANY team, if the specific role that the player has within his OWN team is crucial to any comparison?
.

[fpliii]

"similar" roles is very subjective criteria will differ a lot among whoever is analyzing it

and if "similar roles" are crucial to comparing rapm's of players, then rapm is useless when it comes to comparing players, since no two players play the same role.

i.e. you could never compare jordan's rapm to lebron's, because those guys play completely different roles for their teams... one is a ball-dominating point guard, and the other is an off-ball assassin.... two completely different roles.. i don't see how rapm could apply in comparing these two players.

how can the stat be a measure of the scoring margin impact of a player on ANY team, if the specific role that the player has within his own team is crucial to any comparison?

That's why there's a lot of research being put into classifying players by their roles on the floor. We don't have any data during MJ's prime anyway, but a comparison to Kobe might make more sense.

Problem is, if you have too many classification groups, you'll overfit to the data, and eliminate the use of the regression. But grouping two-way wing superstars isn't a huge problem.