2013 RAPM/IPV/etc.

[Doctor MJ]

explain 1963 for me in regards to Wilt. That year he had the best PER in history yet his team went 31-49. How is that possible?

Figure it's really helpful bringing some of ElGee's estimate in here. Here's how things look from '62 to '64 for the Warriors:

All numbers relative to median:

'62: ORtg: +2.0 DRtg: +1.1
'63: ORtg: +0.0 DRtg: -0.1
'64: ORtg: -2.2 DRtg: +8.0

Kind of amazing I think. This is saying that really the difference between '62 & '63 isn't really even that big of a deal. Basically everything you see here is your typical Wilt Warrior numbers - the stuff you expected to get with Wilt doing his big number stuff - nothing all that extraordinary...until you get to the '64 defense. THAT was astronomical and powerful enough to slaughter every other team in the game easily, except the Celtics whose defense was even better.

This goes toward what I say over and over again: It's not that Wilt didn't have huge impact, it's that you really can't take his numbers at all seriously as evidence of that impact. Wilt proved you could "hack" statistics on an epic scale. He didn't do this because he didn't want to have a big impact of course, but he felt he was safe guarding himself as long as he put up those numbers...and guess what? He pretty much did. To this day, most refuse to even consider that Wilt could get those numbers while accomplishing very little, but how else to explain '65?

'65: ORtg: -6.9 DRtg: 0.0

Massive falloff, Wilt's health a big part of it, but Wilt's stats? Basically the same as they ever were.

[Doctor MJ]

At some point people will realize that RAPM is simply not a useful metric.

What do you think about the fact that guys keep getting hired to do this analysis for NBA teams? As in, we literally are lacking good data for '12-13 because several iterations of these guys have been hired into silence.

What is it that you're realizing that NBA teams still fail to realize more than a decade into them using this data?

It's because he's looking at stats all wrong. If his guy isn't looked at favorably... you throw it out. I love seeing the differences in the top 100 over the years with more data and analysis becoming available. Proves that for the most part, we've been able to admit when we were wrong and correct ourselves.

Still have the same people that come in and dismiss stats that don't work for their guy, etc, but we're making progress.

What's really odd is that I feel like I've watched as much Garnett footage as anyone over the years, and I still underrated his defensive impact (though I did think he was the best defender of the 2000's, even before he got to Boston).

Well let me be more charitable than that:

A big issue people have with advanced stats is that they are basically looking for a reason to justify their distrust for new things they don't understand. So they find a piece of data that looks crazy and say "See you can't rely on it, it's not useful!". The hypocrisy is there in the sense that they can be swayed if the results look keen enough to them, and that they don't actually hold basic stats up to the same scrutiny, but fundamentally the issue is in the approach to assessment.

I go into every stat asking: How can I use this to help my understanding? And not surprisingly I get more use out of stats than people who go into it essentially asking: Is there a way I could jump to a crazy conclusion if I start by looking at the data?

[Dipper 13]

To this day, most refuse to even consider that Wilt could get those numbers while accomplishing very little, but how else to explain '65?

'65: ORtg: -6.9 DRtg: 0.0

Massive falloff, Wilt's health a big part of it, but Wilt's stats? Basically the same as they ever were.

Do you know of any sources that have complete boxscores from that era? By complete I mean including FGA. I am looking to plug in some the numbers into ElGee's formulas to see how both teams did before and after the trade. While the simple in/out doesn't indicate much, based on what I have read the Sixers slowed the pace down once Wilt arrived to work more in the half court whereas before the trade they were a running team keyed by Costello and Greer, two of the fastest guards in the league. He also (along with Luke Jackson) filled a major weakness they had on the boards. On average they made over two more field goals per game (41.4 to 43.6) as a team after the trade. The question is how many shot attempts did it take? We already know their free throw attempts didn't change much at all (37.5 to 37.8). Also I am interested to see how the Warriors offense was affected. I wouldn't expect a huge difference in the defense due to Thurmond and his shifting back to the center position.

Though I don't have the fixed percentages for the defensive formula, seeing as ElGee's website is down.

http://www.backpicks.com/pre-1974-estimations/

[colts18]

How is dirk in the top 5 if he has a negative on/off this year?

[E-Balla]

Its the bias. Dirk usually rates high and he plays a lot of minutes with OJ Mayo (a "bad" performer) and Collison. He gets a lot of the credit they deserve for what the team does. Personally I think in a one season context (and a little beyond that) these numbers are almost worthless.

[SideshowBob]

How is dirk in the top 5 if he has a negative on/off this year?

Dirk at 5 is hilarious

Same issue with the priors. Dirk's impact was big-time in 2011 and 2012, and he hasn't piled up as many possessions this season, so that prior's got a relatively larger influence. Still, it's down from last year (he was +8.0 in the latest 2012 update) and he's down to a +6.6 already this year with only 1014 minutes played. In this case, more time might actually help.

No, a prior does not make it a multi-year study. A prior is used to give the algorithm a distribution of the values in order to make the result more stable. In that way outliers in a smaller sample are easier "eliminated" and you basically needs just a 25% of a season in order to get a comparable error (even though the results are biased and a standard error makes not much sense for ridge regression) as OLS on a big multi-year sample. Obviously, in order to get good enough results for every player, the player also needs to play that amount of games. For Nowitzki for example there are a lot of missed games until ASB, which means his prior value isn't much effected. That should look different now, even though he improved his playing level significantly in recent games.

Anyway, RAPM is not meant to be a "best player list", it is meant to give you a predictive tool based on the result of different game snippets; using those values gives you a better estimation of future game results. Obviously, it doesn't take production, efficiency and fit into account. Things, which would need to be addressed differently in order to really get a "best player list". So, everyone who wants to critize the values should keep that in mind, because I already saw some posts which made it rather obvious that the people really don't understand the method and how to interpret the values.

[SideshowBob]

Its the bias. Dirk usually rates high and he plays a lot of minutes with OJ Mayo (a "bad" performer) and Collison. He gets a lot of the credit they deserve for what the team does. Personally I think in a one season context (and a little beyond that) these numbers are almost worthless.

No, not at all. RAPM is designed to work for single seasons, and relatively small samples in general. It loses relative power in multi-year studies, so that's a clear false statement. You can't make judgement like that without a full understanding of what's going on.

Also, see above response to colts.

[E-Balla]

Its the bias. Dirk usually rates high and he plays a lot of minutes with OJ Mayo (a "bad" performer) and Collison. He gets a lot of the credit they deserve for what the team does. Personally I think in a one season context (and a little beyond that) these numbers are almost worthless.

No, not at all. RAPM is designed to work for single seasons, and relatively small samples in general. It loses relative power in multi-year studies, so that's a clear false statement. You can't make judgement like that without a full understanding of what's going on.

Also, see above response to colts.

The multi-year studies are great to use. They're not going to tell you who's better now but it can tell you who's been better over a few years. If 25% of the season is enough to wade through the bias why are these numbers so far off from other raw +/- numbers?

[mysticbb]

The multi-year studies are great to use.

Multi-year studies have a big issue with aging and developing.

If 25% of the season is enough to wade through the bias why are these numbers so far off from other raw +/- numbers?

The standard error for RAPM after about 25 games is as big as the SE for 2yr APM studies. Additionally, it is mathematically proven that Ridge Regression is giving a better prediction than OLS. SO, for all intend and purposes RAPM is the way to go.

For Nowitzki: He played a lot of minutes with Marion (-3) and Collison (-4), where Marion had -3 and Collison -2 in the 2012 prior informed version. So, both had a lower prior than Nowitzki. Also, with Nowitzki+Marion, the Mavericks are at -10, with Nowitzki and without Marion at +7, with Marion and without Nowitzki at -5; with Nowitzki+Collison at -6, with Nowitzki and without Collison at +6, with Collison and without Nowitzki at -3. So, both are dragging Nowitzki's raw +/- further down than vice versa. Given the fact that Nowitzki played about 40% of his minutes with those two, that might be part of the explanation why Nowitzki's OnCourt is so low (the other part is obviously, Nowitzki sucking for a lot of games when he came back).

[giberish]

well yeah and i think if kobe's team defense wasnt so god awful for like 80%-90% of the season he would be much higher.

This is the sort of thing RAPM is good for. Showing that Kobe's been enough of a defensive liability to give back a lot of what he gives the Lakers on offense.

Casual fans and sportswriters see 30 points and highlight shots and ignore team defense (or figure that if he plays D in the All-Star game he must be a great defender the rest of the time).

[An Unbiased Fan]

At some point people will realize that RAPM is simply not a useful metric.

What do you think about the fact that guys keep getting hired to do this analysis for NBA teams? As in, we literally are lacking good data for '12-13 because several iterations of these guys have been hired into silence.

What is it that you're realizing that NBA teams still fail to realize more than a decade into them using this data?

It's because he's looking at stats all wrong. If his guy isn't looked at favorably... you throw it out. I love seeing the differences in the top 100 over the years with more data and analysis becoming available. Proves that for the most part, we've been able to admit when we were wrong and correct ourselves.

Still have the same people that come in and dismiss stats that don't work for their guy, etc, but we're making progress.

What's really odd is that I feel like I've watched as much Garnett footage as anyone over the years, and I still underrated his defensive impact (though I did think he was the best defender of the 2000's, even before he got to Boston).

This is pretty funny. I have had a problem with RAPM since day 1, Doc can attest to this. My problem has always been the same, these numbers are good for rotation analysis, nothing else.

Sorry, but when you have guys playing close to 40 mins a night, that leaves 2 small windows of time in both halves where they're not on the court. So depending on the type of defense that 2nd unit plays against the other 2nd unit, the player will be heavily influenced. It's a ridiculous way to try and measure impact. Post players will usually be apart of "big" lineups coaches use, so they will skew to higher defensive numbers. Perimeter players will usually be apart of "small" lineups, so they will skew to high offensive numbers.

There really isn't any mystery to these numbers. Ask yourself....why is Metta Word Peace the #33 on those rankings? When, it's fairly obvious that's he's been garbage this season.

[Doormatt]

regarding metta, i think theres some important things to point out.

first: his offense for the most part is awful (hence why hes at 0.3 which is barely average). however he still has some value in that he will occasionally hit a 3 and is a very strong post player and can at times be a matchup problem.

second: hes the only perimeter defender worth mentioning on the lakers. literally everyone else on the team falls somewhere between bad and average. hes also the only player capable of creating turnovers on the perimeter with his pressure/very strong hands (hence why hes averaging a very nice 1.7 steals this year). while he cant really guard the stars of the league too well anymore because hes slowed down considerably, he can still wreck havoc against certain players and make it hard for them. he also plays pretty decent team defense and is very good at generating turnovers. unfortunately hes terrible in transition so the turnovers/steals arent that meaningful.

so its pretty much a combination of still being a strong defensive player+not the worst offensive player. although i personally think his offense is worse than 0.3. its really bad. he takes a lot of horrible shots, isnt a great shooter, cant function in transition and seems clueless at times in the offensive system. theres not many redeeming qualities as a role player for metta offensively. his defense is solid but hes not going to slow down the durant/lebrons of the world anymore, which really makes him much less valuable come playoff time potentially matching up against durant imo.

[mysticbb]

This is pretty funny. I have had a problem with RAPM since day 1, Doc can attest to this. My problem has always been the same, these numbers are good for rotation analysis, nothing else.

And you were never able to show that. In fact, you failed to even understand the math behind the method.

Sorry, but when you have guys playing close to 40 mins a night, that leaves 2 small windows of time in both halves where they're not on the court. So depending on the type of defense that 2nd unit plays against the other 2nd unit, the player will be heavily influenced.

That is WRONG. We are NOT talking about raw On/Off numbers, we are talking about a regression analysis on the data when a specific player was ON the court. How good or bad the team plays defense without that player is not that important, but how good they are in the possession with that player.

[Rapcity_11]

This is pretty funny. I have had a problem with RAPM since day 1, Doc can attest to this. My problem has always been the same, these numbers are good for rotation analysis, nothing else.

And you were never able to show that. In fact, you failed to even understand the math behind the method.

Sorry, but when you have guys playing close to 40 mins a night, that leaves 2 small windows of time in both halves where they're not on the court. So depending on the type of defense that 2nd unit plays against the other 2nd unit, the player will be heavily influenced.

That is WRONG. We are NOT talking about raw On/Off numbers, we are talking about a regression analysis on the data when a specific player was ON the court. How good or bad the team plays defense without that player is not that important, but how good they are in the possession with that player.

*Crickets*

[acrossthecourt]

Right but it does tend to be easier to be more productive per-possession with a lower minutes load, especially with minimal work in isolation. Food for thought when comparing him to players with a different role, just as with ORTG.

Why do people keep saying that? You're pulling that out of your butt. You have absolutely no basis in numbers to state that.

http://web.archive.org/web/200711221152 ... _doctrine/

[SideshowBob]

http://talkingpracticeblog.com/

As an indirect consequence of creating our wagering model, we developed a player evaluation metric that has proven in our testing to be significantly more predictive than other plus/minus or statistical plus/minus (SPM) metrics. Given that our metric is meant to be predictive and not descriptive, we refer to it as IPV (Individual Player Value), and its scale is the standard plus/minus scale (per 100 possessions). IPV begins with the creation of an informed prior for each player. On offense, this prior is taken from a proprietary SPM model to which we apply a non-linear aging/experience curve (with some tweaks and a strong dose of mean regression). On defense, the prior is based primarily on defensive performance in the previous season (even advanced boxscore statistics are pretty useless on defense). We then use these values as a prior for a regularized plus/minus model.

Another blended metric (sort of). Seems pretty solid to me.

[B_Creamy]

http://talkingpracticeblog.com/

As an indirect consequence of creating our wagering model, we developed a player evaluation metric that has proven in our testing to be significantly more predictive than other plus/minus or statistical plus/minus (SPM) metrics. Given that our metric is meant to be predictive and not descriptive, we refer to it as IPV (Individual Player Value), and its scale is the standard plus/minus scale (per 100 possessions). IPV begins with the creation of an informed prior for each player. On offense, this prior is taken from a proprietary SPM model to which we apply a non-linear aging/experience curve (with some tweaks and a strong dose of mean regression). On defense, the prior is based primarily on defensive performance in the previous season (even advanced boxscore statistics are pretty useless on defense). We then use these values as a prior for a regularized plus/minus model.

Another blended metric (sort of). Seems pretty solid to me.

Severe lack of Amir Johnson in Top 3.

[Dr Pepper]

http://talkingpracticeblog.com/

As an indirect consequence of creating our wagering model, we developed a player evaluation metric that has proven in our testing to be significantly more predictive than other plus/minus or statistical plus/minus (SPM) metrics. Given that our metric is meant to be predictive and not descriptive, we refer to it as IPV (Individual Player Value), and its scale is the standard plus/minus scale (per 100 possessions). IPV begins with the creation of an informed prior for each player. On offense, this prior is taken from a proprietary SPM model to which we apply a non-linear aging/experience curve (with some tweaks and a strong dose of mean regression). On defense, the prior is based primarily on defensive performance in the previous season (even advanced boxscore statistics are pretty useless on defense). We then use these values as a prior for a regularized plus/minus model.

Another blended metric (sort of). Seems pretty solid to me.

First glance those rankings make more sense than than the RAPM #'s, but these numbers are also more up to date and different. Kobe is more of a top 15 player than a top 100+

[SideshowBob]

Updated rankings here and on first page.

Here's what v-zero had to say:

EDIT 30/03/2013:

Updated with most up to date pbp data and a vast improvement on the priors.

The link in the OP will now take you to the most up to date version of RAPM for 2013 - significant improvements have been made in the implementation of the priors, data is up to date as of today on BBREF.

[Mr. Crowley]

Updated rankings here and on first page.

Here's what v-zero had to say:

EDIT 30/03/2013:

Updated with most up to date pbp data and a vast improvement on the priors.

The link in the OP will now take you to the most up to date version of RAPM for 2013 - significant improvements have been made in the implementation of the priors, data is up to date as of today on BBREF.

Deron has been pretty bad at defense in several advanced metrics this season, -2.3 for RAPM.

The 3.4 drating for Parker is insane for a PG.