Maybe You Just Need a Change of Scenery: A (Hopefully) Unbiased Look at the Impact Changing Teams Mid-Season Has on Wins Produced

So, the trade deadline has passed.  Everyone knows that Wins Produced’s consistency is worse when players change teams.  In fact, I believe that this is one of the main criticisms of Wins Produced and advanced stats in general.  So, I have decided to do some investigation.

Now, I’m honestly not sure if this has been done before.  In fact, I’m sure that Arturo Galletti has done something about this at some point, but I can’t remember the exact details.  What I have done is looked at every player since the 2000-01 season who has played at least 250 minutes each with multiple teams (although the word “multiple” is somewhat misleading as no player over this time span has played more than 250 minutes with three or more teams in a single season) and compared their Wins Produced numbers between the two teams.  Here goes nothing:

In all, 287 seasons met the criteria.  First, I found the correlations between the two data sets.  Here is the data:

C0rrelations

  • 2011-12: 14 Data Points, 60.4% (I’m not sure if listing these as percentages is necessarily proper, but it is much easier to visualize this way.)
  • 2010-11: 34 Data Points, 56.2%
  • 2009-10: 26 Data Points, 72.3%
  • 2008-09: 27 Data Points, 64.7%
  • 2007-08: 29 Data Points, 59.6%
  • 2006-07: 15 Data Points. 20.0%
  • 2005-06: 22 Data Points, 63.5%
  • 2004-05: 28 Data Points, 37.4%
  • 2003-04: 40 Data Points, 18.1%
  • 2002-03: 16 Data Points, -8.7%
  • 2001-02: 14 Data Points, 45.8%
  • 2000-01: 20 Data Points, 60.0%
  • Overall Data: 286 Data Points, 46.6%

There seems to be a definite leap in consistency in 2005-06, with a drop perhaps caused by a smaller sample size in 2006-07, followed by continued years of consistency since then.  At this point in the post, I have no intention of making conclusions; I’ll save those for the end.  Well, while correlations are definitely useful, they only measure consistency; if the two data sets are parallel, the correlation would still be high even if there wasn’t much or any overlap.  (For example, the data sets {3 6 8} and {7 10 12} have perfect correlation.)  So, I decided to find the average distance both as the straight-up difference and absolute value, as well as the standard deviations for each season.  Note that for the average difference in absolute terms, a negative value actually means an increase when playing for the second team rather than the decrease that would be seemingly implied:

Other Data Points

  • 2011-12: 14 Data Points.  Average Absolute Difference: .082.  Average Unabsolute Difference: -.060.  Standard Deviation: .108.
  • 2010-11: 34 Data Points.  Absolute: .052.  Unabsolute: -.017.  Standard Deviation: .072.
  • 2009-10: 26 Data Points.  Absolute: .050.  Unabsolute: -.002.  Standard Deviation: .065.
  •  2008-09: 27 Data Points.  Absolute: .055.  Unabsolute: -.017.  Standard Deviation: .070.
  • 2007-08: 29 Data Points.  Absolute: .079.  Unabsolute: -.006.  Standard Deviation: .094.
  • 2006-07: 15 Data Points.  Absolute: .057.  Unabsolute: .042.  Standard Deviation: .081.
  • 2005-06: 22 Data Points.  Absolute: .073.  Unabsolute: -.041.  Standard Deviation: .086.
  • 2004-05: 28 Data Points.  Absolute: .073.  Unabsolute: -.024.  Standard Deviation: .091.
  • 2003-04: 40 Data Points.  Absolute: .088.  Unabsolute: .060.  Standard Deviation: .122.
  • 2002-03: 16 Data Points.  Absolute: .107.  Unabsolute: -.063.  Standard Deviation: .128.
  • 2001-02: 14 Data Points.  Absolute: .045.  Unabsolute: .013.  Standard Deviation: .065.
  • 2000-01: 20 Data Points.  Absolute: .076.  Unabsolute: -.036.  Standard Deviation: .097.
  • Overall Data: 286 Data Points.  Absolute: .070.  Unabsolute: .022.  Standard Deviation: .090

I also have some other miscellaneous statistics that help paint a picture:

Miscellaneous Statistics

  • Five-Number Summary: Minimum -.349, Q1 -.071, Median -.023, Q3 .032, Maximum .278
  • Interquartile Range: .103
  • % of Players Whose Wins Produced Rate Increased After Switching Teams: 59.8%
  • % of Players Whose Wins Produced Rate Decreased After Switching Teams: 39.2%
  • % of Players Whose Wins Produced Rate Remained Exactly the Same After Switching Teams: 1.0%
  • % of Players Whose Wins Produced Rate Increased More Than One Standard Deviation (roughly .090 Wins Per 48 Minutes): 20.3%
  • % of Players Whose Wins Produced Rate Decreased More Than One Standard Deviation: 9.4%
  • % of Players Whose Wins Produced Rate Did Not Change More Than One Standard Deviation: 70.3%
  • % of Players Whose Wins Produced Rate Would Be Considered Outliers Using the Formula 1.5*Interquartile Range From Mean: 7.7%
  • Outliers Because of Decreasing Productivity: 9
  • Outliers Because of Increasing Productivity: 13 (Note that if the mean had been 0, the numbers would have been 5 and 18 respectively using the same interquartile range.)
  • Data with the Removal of Outliers (Simplified to be a change of .15 Wins Per 48 Minutes): Correlation: 64.1%.  Mean Absolute Value: .057.  Mean Absolute Difference: -.013.  Standard Deviation: .070.

Now, it’s time for some conclusions.  First of all, it seems very clear that, at least from a correlations standpoint, players’ play has been more consistent when switching teams over the last five seasons.  While I have some theories as to why the numbers have become more consistent in recent years, I have nothing to corroborate them with.  One of my guesses is that the game has become more homogeneous over the past few years, but I have no idea where that would come from besides the recent advances in technology, nor do I have any idea about how that would be measured.  Furthermore, I have no glimmer of an idea as to how the game has become more homogeneous.

It also seems very clear to me that there seems to be some benefit to switching teams as the unabsolute difference is usually at least slightly negative, although it positive for the data set as a whole because of the strange 2003-04 season where there were lots of trades but little consistency among the halves of the season for traded players.  Also fueling the fire for the benefits of switching teams is the fact that the median value for the sample is negative, that there was a signficantly greater number of players whose productivity increased rather than decreased after the trade, and the fact that there were more players with drastic increases in productivity rather than drastic decreases, with “drastic” being defined as a change of +.090 Wins Produced Per 48 Minutes, which is approximately one standard deviation as measured by Josh Weil, and by outliers in my study found by taking the interquartile range.  However, there appears a great degree of randomness in the sample, which makes sense as all NBA players are individuals.

Part of the reason why I wrote this article was to investigate the claim that Wins Produced is inconsistent as players switch teams.  While this appears to be true in the first half of the sample, this becomes less apparent in recent years, and I think that it is fair to say that Wins Produced does “keep” well even when players are traded mid-season, which is arguably when one would expect it to be the least consistent.

The spreadsheet I used for the data can be found at TeamSwitchCorrleation.  Thank you for reading, please comment, and please come back.

Advertisements

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s