The MVP Equation
“What I cannot create, I do not understand.”-On Richard Feynman blackboard at time of death in 1988
I spend an excessive amount of time each year analyzing and agonizing over just who the NBA’s best player truly is (see Parts 1 and 2 for this year here). For stat geeks, It almost like a scab that we cannot help but pick at. I believe it’s the unpredictability of the MVP voting that keeps drawing us in. As much as we tout rational, fact based approaches for picking the best players, ultimately it seems like an arbitrary choice.
This time I’m going to go at it a completely different way. What if instead of trying to build a rational stat based evaluation system for the best player, I instead build a statistics based model for predicting MVP voter behavior?
Respect to Nate Silver for the idea (image courtesy of http://xkcd.com/)
Warning, here comes the science.
Predicted MVP Share =
Team Win %* 30*(A*Player%TotalPts+B*Player%TotalRebounds+
C*Player%TotalAssists+D*Player%TotalBlocks+
E*Player%TotalSteals+F*Player%TotalBlocks+
G*Player%TotalFGMiss+H*Player%TotalFouls+I*Player%TotalTurnovers)
+ DefensiveWinShares/GamesPlayed – Constant
Take that, miss with some stats and the end result looks like so:
That’s what happens when I get bored while my wife watches the most exciting Rose ceremony ever (no, really) and break out the regression tools to try to quantify a statistics based equation for the MVP voting. Now this has gone thru more than a few iterations as I’ve built up the model based on trial and error and obvious observations. Some easy ones are:
- Which stats should I use? Let regression decide.
- Do records matter? Yes.
- Should I adjust for pace? No, hell you shouldn’t even adjust for team production. Pundits have a hard time with scale and adjusting for it.
As you shall see, it’s not perfect but I think it does a great job at illustrating what is actually going on.
For the purposes of this build I used the last ten years of MVP. If I make a graph showing actual MVP vote % versus predicted vote percentages I get:
Or If I focus on the actual vote getters and guys with predicted scores of greater than 20%:
The system has the following top ten mvp seasons for the last decade:
| Players Getting MVP Votes by Year (2003 -2012) | ||||||
| Year | Player | Share of MVP Vote | MVP Rank | MVP Share Score | MVP Share Rank | Delta |
| 2009 | LeBron James | 96.9% | 1 | 100% | 1 | |
| 2010 | LeBron James | 98.0% | 1 | 96% | 1 | |
| 2004 | Kevin Garnett | 99.1% | 1 | 76% | 1 | |
| 2012 | LeBron James | 88.8% | 1 | 75% | 1 | |
| 2003 | Tim Duncan | 80.8% | 1 | 73% | 1 | |
| 2007 | Steve Nash | 78.5% | 2 | 66% | 1 | 1 |
| 2006 | Dirk Nowitzki | 43.5% | 3 | 66% | 1 | 2 |
| 2006 | LeBron James | 55.0% | 2 | 65% | 2 | |
| 2011 | LeBron James | 43.1% | 3 | 65% | 1 | 2 |
| 2012 | Kevin Durant | 73.5% | 2 | 65% | 2 | |
With Lebron’s 2009 season setting the standard and having 4 of the top ten scores (I’m totally ok with both of these statements). If we look at winners for the sample:
| Players Getting MVP Votes by Year (2003 -2012) | ||||||
| Year | Player | Share of MVP Vote | MVP Rank | MVP Share Score | MVP Share Rank | Delta |
| 2009 | LeBron James | 96.9% | 1 | 100% | 1 | |
| 2010 | LeBron James | 98.0% | 1 | 96% | 1 | |
| 2004 | Kevin Garnett | 99.1% | 1 | 76% | 1 | |
| 2012 | LeBron James | 88.8% | 1 | 75% | 1 | |
| 2003 | Tim Duncan | 80.8% | 1 | 73% | 1 | |
| 2011 | Derrick Rose | 97.7% | 1 | 62% | 2 | -1 |
| 2007 | Dirk Nowitzki | 88.2% | 1 | 62% | 2 | -1 |
| 2006 | Steve Nash | 73.9% | 1 | 59% | 3 | -2 |
| 2005 | Steve Nash | 83.9% | 1 | 52% | 1 | |
| 2008 | Kobe Bryant | 87.7% | 1 | 42% | 5 | -4 |
We see that six of the winners led their year. For those who didn’t lead their year we have:
- Rose getting one over Lebron over the Descision in 2011.
- Nash winning a tight one on momentum in 2006 over Dirk.
- Dirk stealing it back in 2007 over Nash.
- And 2008 where it was a tight bunch (the model acutually like Iverson for MVP who got no votes, yay stat nerds) with no favorite and Kobe snuck in and got a media lifetime achievement award.
So it looks like the model provides a pretty good indicator of what will happen with some pretty reasonable sources of variation (lack of differentiation at the top or external factors leading to an incumbent, a media favorite or a bridesmaid sneaking in). I suspect media market size and fanbase play a part (let’s call this the Kobe Corollary). I also suspect having multiple candidates on a team plays a part (let’s call this one Pau’s Dilemma). Also playing on a say a 50 win 8th seed in a loaded year can kill your chances (the Denver Rule). Finally, it does look like advanced stats do have some sort of effect (let’s name this one the Berri-Oliver-Iverson Hypothesis)
Let’s take a look at the candidates for 2013:
While Lebron James is having himself another kingly year at 87% for a score which would make him sixth since 2003 and easily MVP worthy, Kevin Durant is going out and resetting the standard this year. Durant’s 109% score means that this is the most dominant MVP season that I’ve measured in this period (I will, at some point take a look further back). The fly in the ointment for KD is the possibility of Lebron winning on momentum as the incumbent or the possibility of Westbrook stealing votes Ralph Nader style from his own team.
But really, It looks like the King is dead, long live the Prince.
-Arturo