View previous topic :: View next topic |
Author |
Message |
deepak_e
Joined: 26 Apr 2006 Posts: 324
|
Posted: Tue Aug 14, 2007 3:00 pm Post subject: |
|
|
I suspect John wants to keep his PER formula relatively easy to calculate from the official box score stats you'd find on NBA.com or ESPN.com. |
|
Back to top |
|
|
Don Babbitt
Joined: 01 Feb 2007 Posts: 9 Location: California
|
Posted: Tue Aug 14, 2007 3:23 pm Post subject: |
|
|
John Hollinger wrote: | Interesting point on the defense, because we've all seen cases where player's blocks and steals numbers are hugely deceiving. |
We need to assemble a mass of volunteers to break down games similar to what the Football Outsiders do, re: their game charting project.
Otherwise it's all based on the subjective evaluation of the scorekeepers.
Unfortunately there are 82 games in the regular season and not 16. |
|
Back to top |
|
|
jkubatko
Joined: 05 Jan 2005 Posts: 508 Location: Columbus, OH
|
Posted: Tue Aug 14, 2007 9:58 pm Post subject: |
|
|
gabefarkas wrote: | One big thought that's been lingering inside me for a while: PER is essentially a per-minute stat (or per 40 or whatever). How about a gross, overall assessment of production, normalized for the overall production of all players? |
This can be done fairly easily. For example, I recently added two new stats to the Player Stats Search that I'm calling Value Above Average (VAA) and Value Above Replacement (VAR). These measures just take a player's PER and turn it into a number that can be thought of as a representation of a player's overall value. VAA is set up such that the league average player (i.e., one with a PER of 15) has a value of 0:
Code: |
VAA = (PER - 15)/15 * MP
|
For those who dislike saying that an average player has a value of zero, you can use VAR:
Code: |
VAR = (PER - 9)/15 * MP
|
where the 9 in the equation above is just an estimate of replacement-level PER.
2006-07 leaders for VAA and VAR can be found here and here, respectively. _________________ Regards,
Justin Kubatko
Basketball Stats! |
|
Back to top |
|
|
Mountain
Joined: 13 Mar 2007 Posts: 216
|
Posted: Tue Aug 14, 2007 10:41 pm Post subject: |
|
|
Nice additions.
For VAR top 100, average age is 26.2 and average height is almost 6-7. It is comprised of 10 listed centers, 46 forwards and 44 guards.Mean VAR of top 100 is 1619. Average WinShares for this group is just a bit over 20.
VAR/WS varies pretty widely from max 67% above average ratio to 44% below. (Or the ratio could be flipped.) Perhaps someone else better with advanced analysis could look at and comment further about this relationship?
I decided to look at top 100 by Winshares. Average about 1/2 a year older and about 1/2 an inch taller. It is comprised of 11 listed centers, 49 forwards and 40 guards.Mean VAR is lower 1553. Average Winshares for this group is a bit over 21.
Small differences between the top 100s can prompt observations about players (Vets and bigs doing better on Winshares than VAR) but also highlight the different approaches. Winshares seeks to account for defensive impact with equal weight as offense, though as shown in team stats. PER admittedly to date emphasizes offense more and reaches to credit nonscoring contributions to offense. Defense is boxscore individual defense contributions only due to lack of reliable shot defense stats and consensus credit/blame rules. Each method's results are informative, looking at both better, a hybrid always possible.
Last edited by Mountain on Wed Aug 15, 2007 11:39 am; edited 1 time in total |
|
Back to top |
|
|
gabefarkas
Joined: 31 Dec 2004 Posts: 879 Location: Durham, NC
|
Posted: Wed Aug 15, 2007 8:16 am Post subject: |
|
|
There were 95 players that were in at least 2 of the top 100 lists for VAA, VAR, and WS.
Based on these, I obtained correlation coefficients as follows:
Code: | Compare Correl
VAA-VAR 0.9663
VAR-WS 0.8019
WS-VAA 0.7552 |
|
|
Back to top |
|
|
Mountain
Joined: 13 Mar 2007 Posts: 216
|
Posted: Wed Aug 15, 2007 11:41 am Post subject: |
|
|
Thanks Gabe. |
|
Back to top |
|
|
John Hollinger
Joined: 14 Feb 2005 Posts: 77
|
Posted: Thu Aug 16, 2007 8:32 pm Post subject: |
|
|
I'm still figuring out where to take it next. It's a little bit like turning around an ocean liner at this point. Because so many people use it and emulate it, I can't just screw around with it like I could when it was just an excel formula on my computer that nobody else had ever heard of.
John |
|
Back to top |
|
|
Mike G
Joined: 14 Jan 2005 Posts: 1506 Location: Delphi, Indiana
|
Posted: Fri Aug 17, 2007 6:44 am Post subject: |
|
|
jkubatko wrote: | ..
For those who dislike saying that an average player has a value of zero, you can use VAR:
Code: |
VAR = (PER - 9)/15 * MP
|
where the 9 in the equation above is just an estimate of replacement-level PER.
|
Justin, this is no doubt a move in the right direction regarding 'value' assessment. I'm curious about:
-- Could the 'replacement' PER be lower than 9 ? Does that figure have significance?
-- When you total team VAR's (or VAA's), are teams then ranked close to their rank in their actual or expected W-L record? Could you find the best-fitting 'replacement' level this way? _________________ 40% of all statistics are wrong. |
|
Back to top |
|
|
jkubatko
Joined: 05 Jan 2005 Posts: 508 Location: Columbus, OH
|
Posted: Fri Aug 17, 2007 8:19 am Post subject: |
|
|
Mike G wrote: | Could the 'replacement' PER be lower than 9? Does that figure have significance? |
In any given season, if you calculate the percentiles for PER (using player minutes played as weights), 9 is usually right around the 5th percentile. I thought that was a reasonable definition of "replacement." Below is a summary of the last 28 years:
Code: |
Season P5
1979-80 9.2
1980-81 8.8
1981-82 9.3
1982-83 8.8
1983-84 9.3
1984-85 8.5
1985-86 8.7
1986-87 9.1
1987-88 8.5
1988-89 8.8
1989-90 8.5
1990-91 8.7
1991-92 9.1
1992-93 9.2
1993-94 9.2
1994-95 9.6
1995-96 8.8
1996-97 9.0
1997-98 9.1
1998-99 8.9
1999-00 8.8
2000-01 8.5
2001-02 8.6
2002-03 8.3
2003-04 8.7
2004-05 8.6
2005-06 8.1
2006-07 8.4
|
I'm sure you can figure it out, but "P5" above stands for the 5th percentile.
Quote: | When you total team VAR's (or VAA's), are teams then ranked close to their rank in their actual or expected W-L record? Could you find the best-fitting 'replacement' level this way? |
I haven't looked at this yet. I'll get to it. _________________ Regards,
Justin Kubatko
Basketball Stats! |
|
Back to top |
|
|
jkubatko
Joined: 05 Jan 2005 Posts: 508 Location: Columbus, OH
|
Posted: Fri Aug 17, 2007 9:05 am Post subject: |
|
|
Mike G wrote: | When you total team VAR's (or VAA's), are teams then ranked close to their rank in their actual or expected W-L record? Could you find the best-fitting 'replacement' level this way? |
All teams since the 1979-80 season (EW is expected or Pythagorean wins):
Code: |
Pearson Correlation Coefficients, N = 744
W EW VAR VAA
W 1.00000 0.97158 0.82861 0.80178
EW 0.97158 1.00000 0.84096 0.81122
VAR 0.82861 0.84096 1.00000 0.90663
VAA 0.80178 0.81122 0.90663 1.00000
|
_________________ Regards,
Justin Kubatko
Basketball Stats! |
|
Back to top |
|
|
KnickerBlogger
Joined: 30 Dec 2004 Posts: 166
|
Posted: Sat Aug 18, 2007 6:49 am Post subject: |
|
|
John Hollinger wrote: | I'm still figuring out where to take it next. It's a little bit like turning around an ocean liner at this point. Because so many people use it and emulate it, I can't just screw around with it like I could when it was just an excel formula on my computer that nobody else had ever heard of.
John |
This is a good point. PER has made a lot of people aware of statistics, and I think it's fairly well known, at least in certain internet spheres (blogs & certain columnists). Of course there is some resistance to it, but I think at this point most people use it for it's intended purpose: to give you a rough estimate of a player's worth. No one thinks that a player with .5 advantage in PER is truly better, which is a good thing.
However a lot of people use it for a starting point in the discussion, and quickly move away from it, mainly because of the issues regarding it. Or rather I should say the issues with mainstream stats. Since most of these deficiencies are on the defensive end, I think splitting PER into two components (offense and defense) is a good idea. This way we can see a player's value on offense, and then use another tool like 82games data to better evaluate defense. Of course having a dPER to incorporate the pbp data would be ideal, but if you're limiting yourself to traditional stats then that's out of the question. _________________ KnickerBlogger.Net - now indispensable! |
|
Back to top |
|
|
UGA Hayes
Joined: 30 Jun 2007 Posts: 2
|
Posted: Sat Aug 18, 2007 9:03 am Post subject: |
|
|
Even if you had to get the info from a second party I'd find a way to factor charges taken. Also I too am in favor of splitting up offense defense PER.
One criticism I have is that you don't seem to correlate PER with team success all the time.
For instance on the question of Nash, it seems that his teams are among the least turnover proned, even if he himself has around average number of turnovers. Some would say that having Nash decreases team turnovers. He has a very high ball possesion to turnover ratio. If we were keeping a stat it would be called turnovers that didn't happen. Does PER capture this? Could there be way to reward/punish a ballhandler for team efficiency even if its not the PG |
|
Back to top |
|
|
KnickerBlogger
Joined: 30 Dec 2004 Posts: 166
|
Posted: Sat Aug 18, 2007 10:28 am Post subject: |
|
|
UGA Hayes wrote: | Even if you had to get the info from a second party I'd find a way to factor charges taken. |
Is this information available online? _________________ KnickerBlogger.Net - now indispensable! |
|
Back to top |
|
|
Mountain
Joined: 13 Mar 2007 Posts: 216
|
Posted: Sat Aug 18, 2007 12:50 pm Post subject: |
|
|
PER may continue to stay classic PER. Though it would be possible to keep it and still add a PER 2.0 or advanced PER that was split offense and defense, tweaking offense and building defense fully.
Ideally one formula had appeal but no popular metric handles defense fully and satisfactorily. Separate may be more easily digested and accepted. People can back defense out of existing broad metrics, continue to refine existing alternatives or create new offensive and defensive metrics.
Offensive PER might have the lead on offense but Mike G's work is a worthy challenger that attempts more in some areas and Nick S's NPV is coming on the scene too. VAR has appeal and could perhaps be customized for PGs, wings and bigs.
DCS is a new start but in my view probably needs further changes in weighting. Defensive tendex had appeal and could have been a more valuable toool if the data for the whole league were available. While new metrics tacking on shot defense are possible with assumptions about how to score points from own man and rest of opponents it would be simpler and perhaps more solid if adjusted +/- became real-time and split for offense and defense. Then I'd give adjusted defensive +/- my vote for best defensive metric.
Popularizing metrics is tough but the times might be easier than in the past, especially if APBR attempted further leadership on popularizing a next wave. The thought of building on PER recognized PER's important place in the broader market. But maybe the next stage will come from new sources.
Last edited by Mountain on Sun Aug 19, 2007 10:27 pm; edited 1 time in total |
|
Back to top |
|
|
basketballvalue
Joined: 07 Mar 2006 Posts: 51
|
Posted: Sun Aug 19, 2007 8:49 pm Post subject: |
|
|
John Hollinger wrote: | I'm still figuring out where to take it next. It's a little bit like turning around an ocean liner at this point. Because so many people use it and emulate it, I can't just screw around with it like I could when it was just an excel formula on my computer that nobody else had ever heard of.
John |
A nice problem to have John. I'll look forward to an update when you release it. Count me as a vote to give it a new name for the new version, though, so we can all continue to use the well-established defintion for PER you've developed. Revising it and continuing to call it PER would likely cause confusion....
jkubatko wrote: |
In any given season, if you calculate the percentiles for PER (using player minutes played as weights), 9 is usually right around the 5th percentile.
|
Justin, this makes sense, but I don't have a feeling for whether 5th percentile is the right place to make the cutoff (as opposed to 10th, say). Could you look at who the actual replacement players were in the last couple of years using the transaction wire archives on NBA.com and find the average of their PER? I'd be interested to see if it's really 5th percentile.
Thanks,
Aaron
www.basketballvalue.com |
|
Back to top |
|
|
|