This.
If more players averaged over 15 last season than those that did not, it just simply means more players were better than the average guy a season ago.
Interdasting. I'd go:
1. Adjusted Plus Minus+
2. Win Shares
3. Wins Produced
4. Subjective Evaluation (in-depth, not limited scouting)
5. PER
6. Raw Plus Minus
Obviously subjective evaluation depends heavily on how in-depth we're talking. If it's as far as teams like the Spurs or Rockets take it, it can get moved all the way up to #1. Here, I'm defining it more in the traditional scouting role (I know that's vague).
Actually I would agree with you on one thing- Raw Plus Minus should be last. Was a mistake in my previous post.
I am not so keen on the "Wins produced" advanced stats. It suffers from serious limitations, which a lot of hoop experts have commented upon.
Agree with you on "Subjective Evaluation" as well.
Agree PER has a strong bias towards offense.
I like APM+, BUT only when you have large samples. This is something we discussed a while back.
Haha, okay, that was the one that stood out to me as weird and got me to post my order. I was pretty on board with everything else you'd had.
I'm a sucker for regression, hence my love for Wins Produced, but I have to admit I don't know much about it beyond what's on their website. Time for me to look up its shortcomings...
Edit: Just looked it up here. Wow. I'd just mistakenly assumed the correlations would be good enough -- in theory, determining the values of box score components via regression seemed great to me, and I thought the normal shortcomings of regression would be overmatched by its positives. Apparently not.
I wonder if a neural net model would work better? If you've got the data, it's not hard to run with R or SAS. Hm.
Last edited by Drz; 04-04-2012 at 09:50 PM.
Indeed.
I lost my love for PER after I found how badly it evaluated individual defense (and therefore Bruce Bowen). No one who looks at PER to evaluate players, lets say in posterity, would understand the importance of Bowen to the Spurs team. Hollinger was consistently wrong about the Spurs over a point of time because of using his PER as the only evaluating metric to guesstimate team performance.
Although TBH, he has corrected himself over the years a bit by acknowledging the problems with his advanced stats, the biggest drawback he suffers from is that he relies on subjective weights and box scores alone for his advanced stats. A recipe for wrong evaluation.
Pop also loves to throw a wrench on certain statistical models with stuff like sitting out players or constantly changing rotations
Especially hurtful to Hollinger's power ranking
I like how you think my 3 posts (including the one where I corrected myself to say that technically you weren't wrong but that your concern over inefficient chucking-induced PER inflation was utterly unwarranted) came around to [supporting] what you had postedI love how you used three posts to come around to what I had posted, but you still seem to think I'm wrong. At best you can nitpick the semantics of my post since production is an indicator of performance, but I stand by what I said. I also stand by the idea that "practically speaking," it's obvious (to people who've done even basic research on PER) that if the metric is inflated by taking more shots, which affects nearly every player in the NBA as has been covered in this thread, then that is a very practical issue with the metric.
That was cute when you talked about eFG% and advanced metrics, by the way.
You stand by that idea? Fine. That's your prerogative if you want to be wrong
Any player that "chucks" while maintaining a FG% > 30.9 will have an inflated PER. For the sake of the argument, let's define chucking as attempting > 10 FGA per game but maintaining a FG% < 45. How many players in the Top 50 PER meet this criteria? 6/50. A measly 6 out of 50 players in the top 50 have FGA per game > 10 and a FG% < 45 - Kobe, DWill, Lowry, Pierce, Stuckey, and Gooden.
If we set the bar for "chucking" even higher (> 15 FGA per game), only Kobe and Deron Williams meet the aforementioned criteria. Bottom line, practically speaking your (implied) concern over chuck-induced PER inflation was overstated to begin with. I corrected myself to say that - yes - hypothetically if you had a hordes of players chucking up large number of shots per game while maintaining the proper break-even FG%/3P%, then yes, this could potentially be a big flaw in PER. It's not. Practically speaking, this isn't the case at all. Looking at just the top 50 PER players, it's a problem for 6 of them (out of 50) if you have a really generous definition of chucking and FG% inefficiency (> 10 FGA per game and < 45 FG%, respectively) and only 2 (out of 50) if you increase the bar for what cons utes "chucking" to > 15 FGA per game.
By the way, I thought it was adorable how you brought up applying a "neural net model" and using R to analyze basketball dataBe sure to let us know how your undergraduate-level statistics class goes
I will bold exactly where you are going wrong.
This was already posted in the thread, but it looks like you missed it:
Given [that a field goal made is 1.65 points, and a 3 pt. field goal is 2.65 points], with a bit of math we can show that a player will break even on his two point field goal attempts if he hits on 30.4% of these shots. On three pointers the break-even point is 21.4%. If a player exceeds these thresholds, and virtually every NBA player does so with respect to two-point shots, the more he shoots the higher his value in PERs. So a player can be an inefficient scorer and simply inflate his value by taking a large number of shots."
You are arbitrarily defining chucking and say "look, based on these imaginary limits I set up, if they were true, only X players would be affected!"
The unfortunate truth is almost everyone is affected, no matter what fake chucking definition you throw out there.
I'm in my early 30s, my career is in statistics. No graduate degree, but I did take grad-level stat classes, albeit 10 years ago. Because of my career, I still regularly read up on statistical modeling. Obviously none of that alone makes me any more qualified to speak on this topic, but that's my background. What's yours?By the way, I thought it was adorable how you brought up applying a "neural net model" and using R to analyze basketball data
Sorry, I hadn't read your "edited post".
I don't know anything about the "neural net model" and my understanding of statistics is limited to a very utilitarian use of basic regressions for electoral data that I analyse as part of my work sometimes.
But thanks for the links you provided. They explain the shortcomings of various advanced stats very well.
I still find APM+, warts and all, to be the most robust followed by win shares.
Just out of curiosity, Spursfanfromafar, what is appealing to you about Win Shares? When I've broken down the formula, it basically is like taking PER but then adding in a bias for players on a good team. I've really never put much stock in it due to that bias. For example, according to WS/48, Rose, Noah, Boozer, Gibson, Jimmy Butler and Kyle Korver are all top 50 players in the NBA ahead of very good players on bad teams like Greg Monroe, Anderson Varejao and even Steve Nash.
And then even with that strong bias to players on good teams, WS still doesn't give much credit to a player like Bowen. While he wasn't last (like with PER), he was always on the bottom 2 or 3 on the team.
I also find that the formula leans much too heavily on steals, blocks, defensive rebounds and eFG%. But perhaps you can give me some insight on what I'm missing.
Tbh, WS is only interesting to me in situations when it explains how teams devoid of talent were still able to win a lot of games ... like with David Robinson.
"flashes" ain't gonna cut it in the playoffs, sorry.
My point is WS is more appealing to me than PER, only in relation to it. I can try explaining it, but I think it is like reinventing the wheel about the subject. The best explanation that I have managed to search out in limited time is this -
http://www.backpicks.com/2011/01/24/...in-basketball/
As I mentioned in my gradings of various advanced stats, PER and WS both rank below APM+ as they both fail the Bruce Bowen test and both are limited to box score stats. But WS seems to be relatively better than PER. The weighting mechanisms for various stats that result in "points produced" and ulatively help in delivering wins - the "shares" are (albeit linear), more systematically obtained than PER which seems to be too dependent on arbitrary weights (the 1.65 pts for 2FGMs and 2.65 pts for 3FGMs)
Specific criticisms of PER -
http://wagesofwins.com/2012/03/04/wa...mplifies-pers/
You yourself had used plus/minus to check out player pair performances and found out the utility value of Bruce Bowen using point differentials. His value relative to others in the team was established. I don't have readymade links to data on APM for Bowen in 2007 and before, but I am sure, even the adjusted APMs will yield high positives for him as compared to other advanced stats.
WS vs. PER is an interesting comparison. Even though WS correlates better with what actually produces wins in the NBA, I just don't find it very valuable when comparing players on different teams within a single season. IMO, if you are going to account for team play, you might as well switch to a plus/minus model rather than use a PER hybrid. PER, even though it is flawed and it's probably not even technically an advanced statistic due to its arbitrary nature, I do like that Hollinger didn't even attempt to correlate PER with team wins. Due to that, it's a better standalone stat than WS or WP. But it is splitting hairs, tbh.
And, yes, I think the holy grail statistic for basketball will be plus/minus related. I just don't think box score stats offer enough information; statisticians have probably reached the limit of what they can squeeze out of those traditional box score stats. As you've probably noticed, I'm a big fan of plus/minus as long as it has a relatively large sample size. I've never understood the criticism of plus/minus since the game of basketball itself is judged on a type of plus/minus.
One day, there will probably be a type of plus/minus that becomes the go-to stat for judging NBA players. We're not there yet but we're a lot closer than even a half decade ago.
As for APM, I like it but I like RAPM even better. Any thoughts on RAPM?
For reference, here are the rankings of the Spurs players in each of the systems talked about in this thread:
Raw Plus/Minus
Bonner +12.56
Ginobili +11.09
Splitter +9.15
Parker +7.36
Duncan +5.33
Green +5.28
Neal +5.14
Leonard +3.73
Blair -0.47
On-Off Plus/Minus
Bonner +10.72
Ginobili +6.11
Splitter +5.59
Parker +3.07
Neal -1.97
Green -2.02
Duncan -2.39
Leonard -5.00
Blair -12.07
Adjusted Plus/Minus
Bonner +11.37
Parker +9.36
Duncan +7.60
Ginobili +5.29
Splitter +2.68
Green -0.17
Leonard -5.94
Blair -6.29
Neal -6.70
Regularized Adjusted Plus/Minus
Bonner +3.3
Parker +3.0
Ginobili +1.5
Splitter +1.4
Duncan +1.1
Green +0.6
Leonard -0.2
Neal -0.2
Blair -1.9
PER
Ginobili 22.5
Parker 21.8
Duncan 21.3
Splitter 19.7
Leonard 16.8
Blair 17.1
Bonner 14.1
Green 14.0
Neal 13.2
WS/48
Ginobili .226
Splitter .177
Parker .169
Leonard .168
Duncan .152
Bonner .151
Blair .128
Green .114
Neal .079
WP/48
Leonard .288
Ginobili .271
Splitter .160
Green .143
Bonner .139
Parker .126
Duncan .111
Blair .076
Neal .049
I'm going to throw my 2c on this one even though the question wasn't addressed at me.
Ridge regression (Regularization) is a good way to fight off the noise that's the major component of APM when the samples aren't high enough. To put it in very simplistic terms, it allows to "cap off" values into known limits as you calculate your regressions.
The positive is that it can make APM with a little samples to look much more accurate. Values "check off" because they're within the boundaries that you would expect. The negative is that it can hide truly exceptional events or glaring matchup advantages.
Ultimately, if you have a really high sample APM, you're going to get better quality numbers. Unfortunately, as we all know, sometimes getting high enough samples of APM can be troublesome (ie: playoffs only), so RAPM is probably the best next thing.
BY itself, PER or another metric that uses Boxscores like Points Created or TENDEX, is a number that rates the positives more than the negatives. Since something highly valued by the Spurs, POSITION DEFENSE, doesn't show up in a Box Score, then it can't be used to quantify a player like Bowen.
However, if you took it game by game and looked at what a player like say, Bean for LA was averaging in PER and then ran his PER for that game alone against Bruce, you would often find that Bowen would hold Bean under his normal output.
In that case, then PER does show that a guy like Bowen has an effect.
Exactly. It would also show that someone like Blair gives up a high PER to the player he defends.BY itself, PER or another metric that uses Boxscores like Points Created or TENDEX, is a number that rates the positives more than the negatives. Since something highly valued by the Spurs, POSITION DEFENSE, doesn't show up in a Box Score, then it can't be used to quantify a player like Bowen.
However, if you took it game by game and looked at what a player like say, Bean for LA was averaging in PER and then ran his PER for that game alone against Bruce, you would often find that Bowen would hold Bean under his normal output.
In that case, then PER does show that a guy like Bowen has an effect.
Exactly. Who wins in the playoffs often comes down to who makes the fewest mistakes. The teams that advance make the fewest mistakes. That's why Pop dropped Blair from the rotation against the Grizzlies. He was making too many mistakes.
I won't even try to get into a debate over the validity of APM methodology, when there's such a huge disciple present. Easier just to stick to results. I just threw darts and pulled up the APM for 2006-2007. It shows Tony as 145th in the NBA that year, as far as the impact he had while on the floor.
But that's not the best part. The numeric score of his APM for that season was -3.06. In other words, substantially below that of an "average" NBA player, whatever the that is.
But that's still not the best part. At 127 on the list is... wait for it... Juan Dixon. And his APM rating for that season was -2.21. Substantially better than Parker. Perhaps better still was Jannero Pargo coming in at 108 on the list, with an APM of -.67 - much, much better than the lowly Tony Parker.
Does anyone remember watching Tony Parker and Jannero Parker in '06-'07? Does anyone remember considering that Pargo actually had more impact when he was on the floor than Parker? Does anyone remember thinking, "Ya know... if we didn't have Parker, I'd sure love to see that Pargo guy come here"?
How bad does a tool have to be before you recognize that there are insurmountable problems? And what else is there to go by but results? Does anyone really need a tool to tell them that Tim Duncan and Kevin Garnett are good? APM produces a lot of totally expected results. And then it produces a bunch of "surprises" that have to be sorted out by other methods (like common sense). But always, within the list of surprises, there are some players that turn out to be pretty decent. Those are the "success stories" for APM. The ones that are wildly off base just get tossed. Believers always like to say it's because of sample size. But the same things happen no matter how much data you feed into it. Regularly, and without fail.
Edit: I just noticed Mikki Moore in 60th place on the list, with an APM of +1.85. He was almost as far above "average" as Tony Parker was below. I think we all know how much impact he had on the game. Moore did get an $11M contract out of it, so I guess that makes it right.
Some posters in this thread seem to imply that Head Coaches are following their players stats to the very last number. That simply ain't correct.
They don't work in detail, reading stats daily, weekly, and monthly, while considering traditional box scores, Adjusted Plus Minus, Raw Plus Minus, On-Off Plus/Minus, Regularized Adjusted Plus/Minus, PER, Win Shares, Wins Produced, etc.. There are also other stats which aren't widely circulated among fans, like shot charts for every player in different situations and its defensive matchups (e.g. Tim Duncan shooting from the left of the low post after a Gary Neal pass in a determined offensive set vs a determined defensive system vs a particular defender, in the 3rd quarter, and so on). There are also computer-recored game physical stats, like distance covered, speed, amount of jumps, etc. for players and team. There's also gym data, appart from the obvious how many lbs and how many times a certain player lifts them, like resistance levels, strength measurements, wingspans, lactic acid levels, heart rate, etc. All those stats are available about your own team and all the others in the league. And there are also millions of game tape hours.
The ones that are really sunk into that amount of data (and many more which are only for NBA insiders) are the General Manager and the President of Basketball Operations, to decide a players value, trades, how much to offer for a contract, if it's worthy to re-sign a particular player, etc..
But as far as coaches, it's other people in the coaching staff who are processing all that information and giving the head coach a briefing. No single person can process that raw amount of data and transform it into a working system and then into a short speech to their players in a pre-game scouting report or even worse, a timeout. There are tendencies, which then could back up or refute direct observation. Not everything can be measured in stats and that's the beauty of it. Players like Robert Horry, Bruce Bowen, heck even Manu, can't be measured by numbers. And there are players whose stats may look impressive but their on-court performance don't.
"Ok, Melo, reading your PER and chart shot indicates that when you shoot three pointers after a pass from Lin in the left corner in the 2nd quarter your % isn't as good as when you receive in the high post from Chandler in the 3rd, but only when playing versus Toronto away, and only if Calderon comes to double team you, because if it's Bargnani, your WP/48 indicates a slight drop, which doesn't happen against the Bucks, but only at home on February the 6th on even years. Also your lactic acid level is low, have some more yogurt"
RAPM is based on a regression technique that is non-linear and is more sophisticated (I had to do a course on Linear Analysis and Quan ative studies, which I found too complicated and unnecessary for my research interests in grad school and so I basically never pursued it further... but this is the technique used - Tikhonov Regularisation )
A website that uses RAPM to rank NBA players is available. It is quite plain vanilla though -
http://stats-for-the-nba.appspot.com/
Joe Sill who pioneered this technique used to host numbers on a site called "hoopnumbers" but APBR forums say that he has since become a consultant for some NBA team and has removed the site.
An article that explains the various linear, ridge and other regression techniques used to calculate APMs is available here -
GSH, may I ask where did you get the APM information for 2006-07? For what sites I refer to for APMs (82games and basketballvalue)..the data is available only from 2007-08 (although there is downloadable data that needs to be processed).
There are currently 1 users browsing this thread. (0 members and 1 guests)
Posting Permissions
Reply With Quote
