View previous topic :: View next topic |
Author |
Message |
Eli W
Joined: 01 Feb 2005 Posts: 401
|
Posted: Thu Mar 06, 2008 5:05 pm Post subject: Usage vs. efficiency yet again |
|
|
I took the method from my last study on diminishing returns in rebounding and applied it to scoring, which is basically just another way of investigating the usage vs. efficiency tradeoff. I used some of Dean's stats from Basketball on Paper and lineup data from BasketballValue. I think the big advantage to the method I used is that it narrows things down to situations in which players were forced to increase or decrease their usage, which removes most of the usual confounds from players upping their usage in games/matchups when they were playing better (more efficiently) due to having a weak defender on them or just being "hot."
http://www.countthebasket.com/blog/2008/03/06/diminishing-returns-for-scoring-usage-vs-efficiency/ _________________ Eli W. (formerly John Quincy)
CountTheBasket.com |
|
Back to top |
|
|
Neil Paine
Joined: 13 Oct 2005 Posts: 774 Location: Atlanta, GA
|
Posted: Thu Mar 06, 2008 6:28 pm Post subject: |
|
|
That's some phenomenal work there, JQ. I'm not a huge fan of the low R^2, but I like the methodology a lot... In fact, changing the WARP formula to use the 1.25 tradeoff instead of 0.6 creates results that "feel" more believable (Calderon is no longer side-by-side with, say, Kobe -- JC dropped to 17th-most-valuable, and Kobe jumped up to 3rd).
Also, I'd love to hear Dean's thoughts on this new research, and whether it means we can now assume the typical player's usage-efficiency tradeoff is +/- 1.25, not +/- 0.6. |
|
Back to top |
|
|
Harold Almonte
Joined: 04 Aug 2006 Posts: 616
|
Posted: Thu Mar 06, 2008 7:29 pm Post subject: |
|
|
I'm not sure, but I feel that this +/-1.25ORtg. each -/+1% usage is not linearly proportional all the way at the player level. I also feel that the way a player's usage is built has something to do. A player like Calderon with an A/TO higher than 3, and a lot of his usage built by that, and a player like Kevin Martin with a lot of his usage built by FTA, won't change eff. vs. usage like an average player.
Last edited by Harold Almonte on Thu Mar 06, 2008 7:51 pm; edited 1 time in total |
|
Back to top |
|
|
HoopStudies
Joined: 30 Dec 2004 Posts: 705 Location: Near Philadelphia, PA
|
Posted: Thu Mar 06, 2008 7:38 pm Post subject: |
|
|
I need to read what Eli did, but I will say that the 0.6 number I generated was a quick and dirty thing. I never use that since I generate individual values and use those. If Eli finds a different number, it's not a big deal. I hope he puts some range on it so that we can be confident it's not 0....
I'll try to look at it soon. _________________ Dean Oliver
Author, Basketball on Paper
The postings are my own & don't necess represent positions, strategies or opinions of employers. |
|
Back to top |
|
|
Neil Paine
Joined: 13 Oct 2005 Posts: 774 Location: Atlanta, GA
|
Posted: Thu Mar 06, 2008 7:45 pm Post subject: |
|
|
So, I'm guessing you're not at liberty to disclose how you generate individual values... or are you?
(Sorry, I had to try! ) |
|
Back to top |
|
|
HoopStudies
Joined: 30 Dec 2004 Posts: 705 Location: Near Philadelphia, PA
|
Posted: Thu Mar 06, 2008 7:48 pm Post subject: |
|
|
Harold Almonte wrote: | I'm not sure, but I feel that this +/-1.25ORtg. each -/+1% usage is not linearly proportional all the way at the player level. I also feel that the way a player's usage is built has something to do. |
There is no doubt that usage vs efficiency is a simplification for something more complex. All linear weights methods assume a constant value of stats regardless of how many you rack up. Hell, my own stuff is only a little better by having weights depending on context (which is reflected in quantity). But dramatically changing contexts are hard. Usg v efficiency - skill curves - was the first step towards dealing with that. Going further opens up a can of worms. _________________ Dean Oliver
Author, Basketball on Paper
The postings are my own & don't necess represent positions, strategies or opinions of employers. |
|
Back to top |
|
|
Ben F.
Joined: 07 Mar 2005 Posts: 391
|
Posted: Thu Mar 06, 2008 10:09 pm Post subject: |
|
|
Really great job, Eli, as always.
One thing I wonder about the graph - how big is the sample size for the farthest right point on the graph? Because without that point the graph looks like it's saying adding usage to a lineup gets you very little, and it doesn't matter how much you add. Whether it's +10% usage or +2% usage, you only get around +1 ORTG. Yet +12% usage jumps you way up, to close to +3 ORTG. That seems odd, so I wonder if there's any sample issues with that. If so, what does it mean that adding a lot of usage above league average doesn't get you much? And if not, why the sudden jump? |
|
Back to top |
|
|
Eli W
Joined: 01 Feb 2005 Posts: 401
|
Posted: Thu Mar 06, 2008 10:42 pm Post subject: |
|
|
Ben F. wrote: | One thing I wonder about the graph - how big is the sample size for the farthest right point on the graph? Because without that point the graph looks like it's saying adding usage to a lineup gets you very little, and it doesn't matter how much you add. Whether it's +10% usage or +2% usage, you only get around +1 ORTG. Yet +12% usage jumps you way up, to close to +3 ORTG. That seems odd, so I wonder if there's any sample issues with that. If so, what does it mean that adding a lot of usage above league average doesn't get you much? And if not, why the sudden jump? |
The last point represents just 97 lineups with 2784 total possessions played. Generally I don't think the sample sizes of the bins are large enough to draw any specific conclusions about the ups and downs of the chart. I was mainly just trying to show the general trend. Another issue with the chart is that I weighted the diffORtg's of the lineups in each bin by the number of possessions each lineup played - I think that's an improvement over taking just a plain average, but it also has some disadvantages (lineups that played together a lot can have a large effect). With more years of data it would be easier to draw conclusions on smaller segments, but my guess is more data would just serve to smooth things out more. Though the suggestion that low-usage lineups lose more efficiency in ramping up their usage than high-usage lineups gain when decreasing their usage does fit with Dean's findings in his study. _________________ Eli W. (formerly John Quincy)
CountTheBasket.com |
|
Back to top |
|
|
Guy
Joined: 02 May 2007 Posts: 109
|
Posted: Thu Mar 06, 2008 10:44 pm Post subject: |
|
|
Nice work, Eli. A few random thoughts:
Is the sample big enough to look separately at lineups with high vs low projected efficiency? It would be especially interesting to know what happens to low-usage/high-projected-efficiency lineups (since these are the players who should be shooting more, in absence of diminishing returns).
I think your projORtg should probably reflect actual usage in that lineup (if you have it), not the season average. For example, if a .90 usage lineup gets to 1.00 by giving all the 'extra' shots to its most efficient shooter, then the no-diminishing-returns efficiency expectation is a bit higher. (However, don't think this would change your results materially.)
It seems to me it's the low usage part of your graph that matters more, which is where the tradeoff is even larger (looks like a coefficient of approx .35). The important question is not how much more efficient, if at all, Kobe would be at average usage, but how much better he makes his teammates by employing 31% of the possessions. Looking at your graph, it looks like reducing the usage of four players from .80 to .69 (.0275 per player) increases efficiency by about .05. That means an extra 2.5 to 3 points a game, if I've done the math right. Pretty impressive. The fact that reducing the usage of high-usage players seems to pay only a small dividend is further disincentive to shift more shooting to low-usage players. |
|
Back to top |
|
|
Eli W
Joined: 01 Feb 2005 Posts: 401
|
Posted: Thu Mar 06, 2008 10:49 pm Post subject: |
|
|
Harold Almonte wrote: | I'm not sure, but I feel that this +/-1.25ORtg. each -/+1% usage is not linearly proportional all the way at the player level. I also feel that the way a player's usage is built has something to do. A player like Calderon with an A/TO higher than 3, and a lot of his usage built by that, and a player like Kevin Martin with a lot of his usage built by FTA, won't change eff. vs. usage like an average player. |
Definitely. All I was looking at was a general average, which obviously lumps a lot of things together and obscures differences between players and lineups. And you're right that individual possessions is something of a blunt instrument for these purposes. It's worth investigating how the tradeoff affects different parts of usage, from shooting to foul drawing to assists to turnovers. _________________ Eli W. (formerly John Quincy)
CountTheBasket.com |
|
Back to top |
|
|
Eli W
Joined: 01 Feb 2005 Posts: 401
|
Posted: Thu Mar 06, 2008 10:55 pm Post subject: |
|
|
Guy wrote: | Is the sample big enough to look separately at lineups with high vs low projected efficiency? It would be especially interesting to know what happens to low-usage/high-projected-efficiency lineups (since these are the players who should be shooting more, in absence of diminishing returns). |
I tried two separate regressions - one with high-usage lineups and one with low-usage lineups, and unfortunately the sample size was a problem. I couldn't get a statistically significant coefficient for high-usage lineups. I think I did get one for low-usage lineups though. I'll go back and try to find that.
Guy wrote: | I think your projORtg should probably reflect actual usage in that lineup (if you have it), not the season average. For example, if a .90 usage lineup gets to 1.00 by giving all the 'extra' shots to its most efficient shooter, then the no-diminishing-returns efficiency expectation is a bit higher. (However, don't think this would change your results materially.) |
Unfortunately I don't have that information. It would be nice to have and investigate though. _________________ Eli W. (formerly John Quincy)
CountTheBasket.com |
|
Back to top |
|
|
Guy
Joined: 02 May 2007 Posts: 109
|
Posted: Thu Mar 06, 2008 11:01 pm Post subject: |
|
|
Quote: | I tried two separate regressions - one with high-usage lineups and one with low-usage lineups, and unfortunately the sample size was a problem. |
Starting with 4 simple bins, usage X projected efficiency, might be instructive:
low-usage/low-projected-efficiency
low-usage/high-projected-efficiency
high-usage/low-projected-efficiency
high-usage/high-projected-efficiency |
|
Back to top |
|
|
Eli W
Joined: 01 Feb 2005 Posts: 401
|
Posted: Thu Mar 06, 2008 11:19 pm Post subject: |
|
|
Ok, I found the separated regressions that I ran. I'll try the breakdown Guy suggested later.
For all 5387 low-usage lineups (csum%TmPoss < 1), the OLS coefficient was 0.56 with a SE of 0.12 and an R-squared of 0.004. Weighting by possessions, the coefficient was 0.32 with a SE of 0.08, R-squared of 0.003. Using only the 240 low-usage lineups that played at least 50 possessions together, the coefficient was -0.23 with a SE of 0.19 and an R-squared of 0.006.
All 2729 high-usage lineups: coefficient of 0.24, SE 0.24, R-squared 0.0004. Weighted by possessions: coefficient of 0.096, SE 0.103, R-squared 0.0003. The 315 high-usage lineups with at least 50 possessions: coefficient of 0.40, SE 0.17, R-squared 0.02. _________________ Eli W. (formerly John Quincy)
CountTheBasket.com |
|
Back to top |
|
|
DLew
Joined: 13 Nov 2006 Posts: 222
|
Posted: Thu Mar 06, 2008 11:41 pm Post subject: |
|
|
This is really great work Eli. I quite enjoyed it. I hope that Henry links it on Truehoop.
Could you do the same thing with multiple seasons worth of data to determine how robust the findings are? |
|
Back to top |
|
|
Eli W
Joined: 01 Feb 2005 Posts: 401
|
Posted: Thu Mar 06, 2008 11:55 pm Post subject: |
|
|
DLew wrote: | Could you do the same thing with multiple seasons worth of data to determine how robust the findings are? |
Yeah, I forgot that BasketballValue had lineup data from the last two seasons that could work for this study. For the rebounding study only this year's data worked (because that's when Aaron added lineup rebounding numbers), so when I started this study I unthinkingly again just used this season's data. I'll try to re-run things with the larger data set in the next few days. _________________ Eli W. (formerly John Quincy)
CountTheBasket.com |
|
Back to top |
|
|
|