Worst performing 2012 free agents

[DSMok1]

Just to be clear, do we think that RAPM (even eight years of it) is reality? I've had the feeling that some people implicitly think that, but it can't be true, right?

Depends on what you want to measure. Theoretically, that is a big enough sample size to get very nearly an exact average value for each player. Note--that ignores fit/synergies.

To use for the basis of a regression, I would say it is certainly very close to precisely accurate, if looking to get the values of different box score stats. Individual issues should balance out.

For individual players, there are still some issues--primarily related to the effects of aging. And if a player only played a few minutes, well, there still isn't sample size enough to resolve anything.

But for high-minutes players it is very good.

[Crow]

Based on reported declining estimated errors for new versions and other data & observation, my rough take is that is at reasonably close in at least 2/3rds of the bigger minute cases. I tend to talk like I trust it more but I try to cautious and am willing to consider the view that it is off in particular cases or off by enough in many or most cases to the extent that it shouldn't be used as the sole rating source. A consistent or fairly consistent patter in RAPM over many years for a big minute player is something I tend to generally respect / believe in.

[xkonk]

This is the same RAPM that a number of people have asked to have aging curves applied to, correct? And that JE has recently reported is improved by adding box score information? How can any existing version of it be 'reality'? I'm on-board with saying it's a good shot, or perhaps the best measure we have right now, but to call any statistical measure of basketball that we have 'reality' seems like quite a stretch to me. Reality is what happened out on the basketball court, and we are approximating it at varying levels of coarseness and accuracy.

[talkingpractice]

Everything I have done/tested makes me think that, at least for players with reasonably consistent performance (and/or matching normal aging patterns), a RAPM model with a well-specified informed prior is pretty darned close to reality, in much less than 8 years.

[v-zero]

But we should never forget that one important limitation of RAPM (and any such metric) is that basketball teams are not truly expressible as a linear sum of individuals. All models are an abstraction, even in Physics, let alone in basketball statistics.

I like RAPM, but I don't trust it.

[xkonk]

I've enjoyed reading the qualifications in every defense of the "RAPM is reality" statement so far. It's the "60% of the time, it works every time" of sports statistics.

[DSMok1]

This is the same RAPM that a number of people have asked to have aging curves applied to, correct? And that JE has recently reported is improved by adding box score information? How can any existing version of it be 'reality'? I'm on-board with saying it's a good shot, or perhaps the best measure we have right now, but to call any statistical measure of basketball that we have 'reality' seems like quite a stretch to me. Reality is what happened out on the basketball court, and we are approximating it at varying levels of coarseness and accuracy.

The article I wrote here should sum up exactly what the benefits and issues with APM/RAPM are: http://godismyjudgeok.com/DStats/2011/n ... ilization/

[Crow]

What are the current season big minute xRAPM player values that folks (especially serious, heavy doubters or non-believers or non-users of RAPM) feel are the most wrong or most likely to be significantly wrong? I understand and accept general criticism and caution with RAPM and folks are free to not use it or reject it if they feel they should / want to but I would like to hear what 3-10 players are the best cases that they feel support their belief that RAPM should not be used because it is considered seriously flawed, unreliable or wildly unreliable as a measure of current overall value. In bringing such names up, it would be helpful to state the main resources used to come to this belief / judgment about the perceived error of the RAPM value. It is pretty easy to criticize RAPM generally, it has been done a lot and it does deserve critique and caution. I think it could be helpful though to hear some more specific player case critiques. One year errors would be the easiest to spot and bring forward and perhaps be acknowledged. Are there any big minute players with at least 3 prior seasons in the league that any one wants to assert are far better or worse in "true impact" than their XRAPM value history?

Among the top 30 player values this season, the main 2 that I quickly identified having some concerns about are for Ilyasova and Gallinari. Both current XRAPM values seems high by their current boxscore data. Both histories show substantial improvement over time. I tend to believe they have improved but maybe they are not as good right now as currently shown by the prior-informed value and perhaps they were not as bad as their early career values, influenced by the way rookie values are intiially set. I still think it very likely that they are at least meutral in true impact, if not a modest positive. If RAPM is going to be challenged more fully it would be better to find / hear about cases of players with long histories and longer RAPM valuation histories.

Among those worse than -3 on the xRAPM estimate, the first name that seems worth a bit of review is Ben Gordon. Is he having one of the very worst impacts this season? It looks that way from extremely bad raw +/-. Maybe he isn't really -5.8 but xRAPM has him worse than -2 for four years running, down from what he generallly had before that (near neutral to mildy negative). I'll generally believe that general story. That xRAPM story is basically consistent with the boxscore metric story (at least winshares / 48).

Barbosa has been estimated on xRAPM to be worse than -2 basically his entire career except one year with a neutral rating. He has been worse in recent years including this season. I tend to accept that valuation history. Anybody think he is and has been significantly better in overall impact? The GMs in the free agent period genrally didn't seem to think so. Boston liked something or had the need at the minumum price. But the raw team +/- and win % for Barbosa is extremely bad and by far the worst of the rotation.

[v-zero]

Big names? Kobe, my favourite punching bag. I am of the opinion that he was once good, and that now he really isn't, mostly because he's so desperate to chase down MJ (which nobody apparently has the heart to tell him is hilariously unrealistic). He has suffered a big boost over the years from collinearity, and continues to do so as I see it. He is pitiful defending, and his offensive production has been negative (in my opinion) for the last season and a half.

[DSMok1]

Big names? Kobe, my favourite punching bag. I am of the opinion that he was once good, and that now he really isn't, mostly because he's so desperate to chase down MJ (which nobody apparently has the heart to tell him is hilariously unrealistic). He has suffered a big boost over the years from collinearity, and continues to do so as I see it. He is pitiful defending, and his offensive production has been negative (in my opinion) for the last season and a half.

ASPM is not based on +/- data; it's based purely on box-score data, and it still likes Kobe quite a bit. In fact, it has Kobe as the second-best offensive player in the league so far this year.

[Crow]

Thanks for the name nomination for public consideration.

I see that you had Kobe as below average in the construction of your team predictions at your site. http://thenbamodeller.wordpress.com/201 ... edictions/ Do you still have him negative now just for this partial season? Is there any other metric than has him below average / negative for this season? I am not immediately aware of one. Anyone let me know if you know of another such rating. xRAPM has him in 43rd place with a +2.3 estimate overall, +3.8 on offense and -1.3 on defense. That is only slighly better than last season, a bit of a surprise to me (is xRAPM heavily affected by the prior?). And this is way too high for you? Low to way too low to many Kobe fans and admirers. I am pretty comfortable with xRAPM's rating of him on quick glance around. If anything I wonder if it might be a bit too low. it might drift up if he sustains the early performance and the prior diminshes somewhat in impact.


I'd be surprised if "Kobe is rated too high by XRAPM" is the best, most widely accepted criticism of xRAPM but I appreciate the feedback.

[v-zero]

ASPM is not based on +/- data; it's based purely on box-score data, and it still likes Kobe quite a bit. In fact, it has Kobe as the second-best offensive player in the league so far this year.

I know, I also don't agree with that. ASPM is formed using a regression against RAPM, so it's not like they're inseparable. ASPM has a usage adjustment that I can't really say I like.

EDIT: For the year I have Kobe as a +0.3 player, when there's no prior.

[Crow]

Reviewing Daniel's articles on over and underrated for RAPM compared to PER at his site, I generally support the RAPM values as more accurate than the PER values for these cases due to 1) PER's lack of full consideration of player's offensive impact on teammates offense and perhaps overcrediting for high usage and of course the lack of shot defense and 2) my subjective opinion about the players overall value. I haven't systematically checked other metrics for these outliers but probably will if any make the assertion that specific names on Daniel's list are strong cases for where RAPM over or under rates these players in the long run compared to true value rather than compared to PER.

[Crow]

EDIT: For the year I have Kobe as a +0.3 player, when there's no prior.

Thanks for the reply. Is an offense / defense split available for that overall value? Are there other big name offensive players that your metric rates significantly lower than other metrics including xRAPM? Would you say this is primarily happening to Kobe (and possibly others) because you have a significantly different analysis of and weighting for usage or usage in a dynamic handling? Based on your reply to Daniel, I currently assume it is but would listen if you had more to say on the topic.

[v-zero]

Sadly a split isn't available because of the construction of the model at the moment, but I may try to change that in future. The model primarily adjusts for usage by considering the interaction effects between assists and usage, and as such doesn't reward guys with high usage who just chuck it up without ever looking around for open shots. I have tried making adjustments for usage outside of this but they don't produce superior predictive ability so I discarded them. The model uses the final score difference of each game pro-rated to minutes played to give each player a base rating and then separates out team mates by considering each player's marginal production of box score stats above or below the average of his team mates over the course of the game. The reason I do this is to remove the problems caused by using the box score in a vacuum. If a team has a load of guys who look good on the box score, but who don't get good results, then they won't have good ratings.

I don't have a list of guys that it appears to over/under rate relative to RAPM because I am moving on from that model as soon as I complete my work on parsing through play-by-play data to build a better box score (like JE has recently done).