How Inconsistent is Jeff Green, really?

[PickNRoll]

Player	PPG	Avg Deviation (pts)	Avg Deviation %
2013-14 Jeff Green	17	6.15	36.15%
87-88 Larry Bird	29.9	7	23.30%
90-91 Jordan	31.40	5.6	17.84%
2011-12 Rondo	11.88	5.98	50.35%

Put simply, Green scores 17 +/- 6.  Bird scored 30 +/- 7.  Rondo had slightly less absolute variance (in points), but extremely high variance as a percentage of his scoring output.  Somehow I expected that. 

These comps are completely arbitrary and not the most useful.  Just a sniff test.

First, Tommy Point for actually making a legit table. I never have the patience.

But one thing you should factor in with this is % of ppg with the average deviation. Bird's % is 23%, Greens is higher, and therefor has a more pronounced impact .

Yeah, tables are a pain.  Here's what I did.  Put it in a spreadsheet.  Copy-paste it into this website:  http://tableizer.journalistopia.com/.  It converts it to an html table.  Now put it in a text editor and replace the angle brackets with square brackets.

[Fafnir]

, but the 40% threshold is arbitrary.

And punishes Green for playing in a era that understands how to use the 3 ball. eFG% or bust!

Man I'd love to see Larry with the sort of offensive structure that they put around Dirk. He shredded the league anyways but still....

[mmmmm]

Player	PPG	Avg Deviation (pts)	Avg Deviation %
2013-14 Jeff Green	17	6.15	36.15%
87-88 Larry Bird	29.9	7	23.30%
90-91 Jordan	31.40	5.6	17.84%
2011-12 Rondo	11.88	5.98	50.35%

Put simply, Green scores 17 +/- 6.  Bird scored 30 +/- 7.  Rondo had slightly less absolute variance (in points), but extremely high variance as a percentage of his scoring output.  Somehow I expected that. 

These comps are completely arbitrary and not the most useful.  Just a sniff test.

Just from your data set, it looks like the higher the average points scored, the lower the variation as a percentage of scoring output.

That would stand to reason.

Avg Deviation % = Avg Deviation / Avg PPG

It's a very limited stat.  It punishes low scorers like Rondo and forgives high scorers like Bird. A more valid measure might be "for all scorers who average between 15 and 20ppg,how does Green rank?".  You could also looks at Z scores for points or deviations.  I think Roy's analysis is actually a better measure than what I presented, but the 40% threshold is arbitrary.

What about using USG rates and minutes played as the category?  I.E. all players with USG between 20-23%, playing 30-36 mpg, or whatever?   That might seem to better normalize their 'role' in the offense.

[kozlodoev]

Put simply, Green scores 17 +/- 6.  Bird scored 30 +/- 7.

That's not really what it is, assuming that by "average deviation" you mean the calculated standard deviation of whatever sample of games you chose to work with.

If so, what the data shows that in roughly 95% of his games, Jeff Green will score between 5 and 29 points, and Bird will score between 16 and 44.

The coefficient of variation (that's what you call "average deviation %") is somewhat useful to the extent that we usually expect players that score more to have higher absolute volatility.

Overall, I'm not sure what this tells us, exactly. Numeric analysis is only useful if it's answering a very specific question, and in this case this thread is pretty meaningless unless we have established criteria by which to evaluate "consistency".

By "average deviation" I mean average deviation about the mean, or Median Absolute Deviation, not standard deviation.  Absolute deviation is less sensitive to outliers and does answer the question, "on average Jeff Greens scores n +/- y".  Here are the standard deviations if you're curious.

Green: 8.12
Bird: 8.77
Jordan: 7.07
Rondo: 7.74

e.g. If a guy scores 0,20,20,20,20,40 , standard deviation is high (12.6).  MAD is low (6.66).  That guys looks fairly consistent to me.  SD and MAD are both limited, just in different ways.

"Median absolute deviation" is a measure of dispersion around the median value, not around the mean. Therefore it cannot be used the way you're using it here.

Moreover, the only reason to use MAD is if you have a small sample that can be sensitive to a handful of extreme points. With sample sizes of 30+ observations (such as the usual NBA season), there's typically no practical reason to use MAD instead of SD.

Also:

e.g. If a guy scores 0,20,20,20,20,40 , standard deviation is high (12.6).  MAD is low (6.66).

The MAD (the median value of the absolute deviations from the median) of this sample is 0 (deviations are 0, 0, 0, 0, 20, 20; their median value is 0). Calculating a SD of a sample of 5 observations is statistically pretty close to meaningless.

[bdm860]

While I always like to see people pick apart stats and get a detailed look to support their thoughts, criticism, praise, etc., I think using average and deviation to look at consistency is not the best, because consistency isn’t necessarily a good thing.  In the examples given by the OP, if Jordan or Bird is -1 deviation away from their average, they’re still in the top 10 in scoring in the league.  I wouldn’t consider Jordan or Bird inconsistent because they put in 20 some nights and 40 on others.

If we wanted to look at just points, I think a better way to look at it is to establish some kind of “minimum acceptable threshold” that you need your star player to score.  How often will player X give us Y or more points?  For Jordan and Bird, I’d probably say anything under 20 points is likely a bad game for them.  Jordan scored 20 or more 95% of the time, Bird 87% of the time in the seasons the OP picked.

Jeff Green on the other hand, what’s the minimum acceptable threshold for him? 15? 12? 10?  He only scores 15 or more  61% of the time, 12 or more 77% of the time, 10 or more 83% of the time.  You're more likely to get 20 in a game from Bird or Jordan or ’06 Pierce, then you would get 10 in a game from Jeff Green.

Since numbers don't format too well here, I threw it in an image adding several other players I thought we might expect similar scoring from.  (I tried to pick players around age 27 or between 17-20ppg, or both, but I wasn't too consistent with this so feel free to complain).  Personally I don't think Green compares too well to these players.  I could count on Ricky Davis or Ron Artest or Antoine Walker more than I could count on Jeff Green to consistently score what I would expect from them.  (Though this is only taking into account points scored and nothing else).





EDIT - Correction, Walker's stats are for '05 not '06.

[PhoSita]

Consistency aside, he's just not a terribly efficient scorer or a particularly productive player in anything but points.

[PickNRoll]

Put simply, Green scores 17 +/- 6.  Bird scored 30 +/- 7.

That's not really what it is, assuming that by "average deviation" you mean the calculated standard deviation of whatever sample of games you chose to work with.

If so, what the data shows that in roughly 95% of his games, Jeff Green will score between 5 and 29 points, and Bird will score between 16 and 44.

The coefficient of variation (that's what you call "average deviation %") is somewhat useful to the extent that we usually expect players that score more to have higher absolute volatility.

Overall, I'm not sure what this tells us, exactly. Numeric analysis is only useful if it's answering a very specific question, and in this case this thread is pretty meaningless unless we have established criteria by which to evaluate "consistency".

By "average deviation" I mean average deviation about the mean, or Median Absolute Deviation, not standard deviation.  Absolute deviation is less sensitive to outliers and does answer the question, "on average Jeff Greens scores n +/- y".  Here are the standard deviations if you're curious.

Green: 8.12
Bird: 8.77
Jordan: 7.07
Rondo: 7.74

e.g. If a guy scores 0,20,20,20,20,40 , standard deviation is high (12.6).  MAD is low (6.66).  That guys looks fairly consistent to me.  SD and MAD are both limited, just in different ways.

"Median absolute deviation" is a measure of dispersion around the median value, not around the mean. Therefore it cannot be used the way you're using it here.

Moreover, the only reason to use MAD is if you have a small sample that can be sensitive to a handful of extreme points. With sample sizes of 30+ observations (such as the usual NBA season), there's typically no practical reason to use MAD instead of SD.

You are correct.  I meant "Mean absolute deviation" where I said "Median absolute deviation".  My mistake.

[kozlodoev]

You are correct.  I meant "Mean absolute deviation" where I said "Median absolute deviation".  My mistake.

Sure, but the point remains that absolute deviations are a very rough approximation which is not necessary with 82 data points.

[mmmmm]

I'm not sure how useful that is without knowing what each player's USG rate was, how many minutes they were on the floor each game, how many games they played that season and how many shots they were taking.   I know that Marion was a similarly modest USG player and his numbers in that chart look very similar to Green's.  Bird, Jordan & Pierce were all 30+% USG players.   

[PickNRoll]

You are correct.  I meant "Mean absolute deviation" where I said "Median absolute deviation".  My mistake.

Sure, but the point remains that absolute deviations are a very rough approximation which is not necessary with 82 data points.

Not sure what you mean by "rough".  Take a collection of 82 deviations.  Now square them to create another set.  Which set is more "rough"? 

I understand the reasons why SD is the more often used measure, but one isn't more precise than the other.  SD just assigns more weight to outliers.  BTW, I'm not a math guy.  I took 3 years of math in college, but I'm open to the possibility that I'm wrong. 

[PickNRoll]

@bdm - TP for those insights!
@mmmmm - I think you're trying to introduce efficiency where bdm is only measuring consistency.

[mmmmm]

Consistency aside, he's just not a terribly efficient scorer or a particularly productive player in anything but points.

Well, his efficiency has tanked since Jan 13.  It has somewhat recovering over the last couple of months, but still sub-par.

Going into the Jan 13 game, Green's eFG was 49.8%, his 3PT% was 37.5% and he was scoring 1.25 points per FGA.  That's not at all inefficient.

If you want a larger sample, from coming back from his heart surgery up until that date, (3764 minutes of play) his eFG was 50.5%, his 3PT% was 38.5% and he was scoring a very efficient 1.28 points per shot.

I don't see how you can characterize that as 'terribly inefficient'.

However, since injuring his shoulder in the Rockets game on Jan 13, Green's eFG is 43.9%, his 3PT% is 32.8% and his pts/shot has plummetted to a mundane 1.13 per FGA.   (Actually, the last value is not as terrible as the shooting efficiencies.  Its just way below where he has been before that).

Most of the damage to his averages during that period comes from the small span of games from Jan 13 through Feb 2.  During that time, his eFG plummeted to the floor at just 39%.

I think it's pretty obvious from the data that the shoulder injury had a dramatic affect on his shooting immediately and continues to bother him.

So, which Jeff Green are you talking about?  The healthy one or the obviously injured one?

[ManUp]

I think when people say that Jeff Green is inconsistent, they mean he has a lot of really bad games.

2014 Jeff Green -- Games below 40% FG%:  27 out of 69
1988 Larry Bird -- Games below 40% FG%: 10 out of 76

The other part of the "consistency" Green lacks (that stars usually possess) is the ability to impact the game in other ways when their shot isn't falling.  When Larry wasn't shooting well, he usually impacted the games through rebounding or passing.  Green doesn't do that.

(One rough estimate of impact on a game is John Hollinger's "game score" (http://www.basketball-reference.com/about/glossary.html).  In 1988, Larry only had 4 games where his game score was below 10.0; so far in 2014, Jeff Green has had 36 such games, including 5 where his game score was negative.  I know you weren't comparing them as players, but in terms of ability to consistently impact the game, Larry was at a much higher and more consistent level.)

The thread should have ended after this.

TP Hobbs.

[kozlodoev]

I understand the reasons why SD is the more often used measure, but one isn't more precise than the other.  SD just assigns more weight to outliers.  BTW, I'm not a math guy.  I took 3 years of math in college, but I'm open to the possibility that I'm wrong. 

It's a bit simplistic to say that "SD assigns greater weight to outliers".

SD "penalizes" observations that are further away from the mean. Not all observations that are far from the mean are "outliers", some variations are just naturally volatile

The need to correct for outliers in small samples arises from the fact that, (a) it's considerably harder to determine whether an observation is an outlier, and (b) a single outlier willmay result in completely invalid SDs.

And yes, SD and confidence intervals are absolutely more precise. They give you value intervals and an actual percent chance for the observed variable to fall in said ranges. MAD is inherently unable to provide confidence intervals, and is thus less precise.

[mmmmm]

@bdm - TP for those insights!
@mmmmm - I think you're trying to introduce efficiency where bdm is only measuring consistency.

Well, I think that 'consistency' is in large part a function of role in context.

The simple fact is, players have different roles on teams.   That affects how they are utilized in games and will affect how many shots they will take.   I think it is simply not useful to compare players with dramatically different roles playing in dramatically different contexts.

If you are measuring consistency of production, efficiency is important, but pales in comparison to utilization.  A guy shooting 45% eFG, taking 20 shots, will score a lot more than a guy shooting 50% eFG taking 10.   One is not necessarily better for the team than the other.