Edit Note: Sorry for the long delay in posts; work and personal life got better of the previous two months. But let’s do an in-depth NBA MVP analysis to make up for it.
With the NBA season wrapping up with a surprising Cleveland Cavaliers seven game victory over the Golden State Warriors, questions resurfaced about the most valuable player in the NBA. This year, it was the first ever unanimous vote for Steph Curry. But do some simple analytics back that up? If we look at raw numbers such as points, steals, and so on, it’s very hard to deny Curry his status as an MVP. However, such simple totals don’t paint the whole picture. For instance, Steph Curry led the league in steals for the second year in a row with 169 steals; a rate of slightly over 2 per game (2.14 to be exact), so the natural extension is that Curry is labeled as one of the best on-ball defenders. Which, if you watched majority of the Golden State Warriors games, the Western Conference Finals against Oklahoma City, and the NBA Finals against the Cavaliers, you would find out that Curry’s on-ball defense is one of the worst on the Golden State team.
In fact, in the Oklahoma City series, Russell Westbrook scored on over half of his possessions with Curry on ball. Klay Thompson? less than thirty percent of possessions. Now, you will find this statistic is different than ESPN’s statistics as they do not include free throws due to fouls and do not include switches on screens. Which is a disappointment, as the Oklahoma City (and later the Cavaliers’) game plan was to screen Iguodala, Thompson, and Livingston off their scorer and onto Curry. But Curry led the league in steals…
Curry has established himself as one of the best off-ball defenders and gamblers. Over 100 steals (123 of 169) were obtained by shooting passing lanes. That is, if we take location data into account, the steals that were obtained by over six feet from the player who committed the turnover, we find that Curry decimates the league by and large. For steals within six feet, Curry is sub-par. So combining the fact that Curry gives up points roughly half the time he is on-ball defender on a shooter and steals less than average starting point guards in the league, we would find that Curry is a rather mediocre on ball defender.
So why go into this diatribe on Curry’s on-ball defense? It’s because it cannot be used to say Curry is a premier defender. Curry is still an elite team defender when off-ball defense is included. And he is the top offensive player in the league.
So let’s use a metric that identifies the impact of a player on the court. To do this, let’s consider the role of every player on the team. For instance, does a rebounder fare better than a shot-blocker? Does a three point shooter who is a sieve on defense better than a player who gets five assists a game and fouls out? More specifically, does LeBron James contribute more to his team overall than Stephen Curry?
Here, we break down every possession in the 2015-16 NBA season. For this season, a total of 476 NBA players logged a minimum of one of the 758,759 total NBA possessions. The maximum total number of possessions a player could have played this season was 13,119 (Milwaukee Bucks). The maximum attained was by James Harden (Houston Rockets) with 9,863 possessions.
For each possession, we measure the offensive and defensive contribution of every player on the court. To do this, we use the offense as a positive interaction. This means that if a player gets an offensive rebound, they get a point. If they make a basket, they get a point. If they get an assist, they get a point. If the basket is scored, the points are also added. If the inverse happens: missed shot, defensive rebound, turnover, etc… then a negative point is scored.
For example, if James Harden drives the lane and tosses an oop to Dwight Howard, who misses the dunk, but Harden grabs the rebound and lays it in; the possession is scored as a miss (-1), offensive rebound (+1), made basket (+1), and (+2) points scored for a total possession of 3 points. Now if Harden drives the lane and oops to Howard who is blocked and the ball is rebounded by the defense, the possession is score as a block (-1), a missed field goal (-1), and a defensive rebound (-1) for -3 points. Since the offense is positive, the first possession is 3 points for the offense and -3 points for the defense. In the second situation, the score is -3 points for the offense and 3 points for the defense.
Now, using this scoring system, we calculate the projection of possession scores onto the players on the court. This will result in a heavily correlated scoring system. By adjusting for correlations between players by whitening the data, we obtain a per possession score.
That’s a lot done so far. But basically, this projection just states how many possession points, we expect a player to get on offense and stop on defense. If a player expects a basket and a stop every possession, we expect a score of approximately 10. If they score a basket every two possessions more than than their opponent, we expect a score about 5 points.
Now, if a player averages say 5 points a possession, then we have to identify whether they are a garbage time champion or a key contributor. To distinguish this, we scale the points per possession by the percentage of possessions played. This is a slightly double-edged sword as key players who spend significant time on the disabled list might look worse than their actual value. But then again… we are looking for the MVP for a season. And they are not contributing in games they do not play in.
So with that set-up in mind, let’s take a look at the top 25 players in the NBA:
|Rank||Player||Score||Poss. Played||Total Poss.||Weighted Score|
Here, this suggests that Klay Thompson is the top player in the league. There is a definite argument to be made as he set a record high in three-pointers made for a second-option that would have broken the three-point made record a couple seasons ago for all first-options. That combined with his on-ball defense being tops in the league… it’s no wonder why Thompson tops the list.
The only surprise name in the top 10 is Kentavious Caldwell-Pope who puts up pedestrian starting numbers: 14 PPG, 1.4 SPG, 3.7 RPG. However, Caldwell-Pope works effectively well with Andre Drummond on defense, helping lead the Pistons to make the playoffs (despite a sweep to Cleveland).
Here, we find Curry at #8 on our list, which is one of the lowest numbers for our list to win MVP over the past 10 years. The lowest to date is Derrick Rose’s 2011 NBA MVP ranking #13 on our list. Typically the MVP lands on our top five list; three times being the #1 player in the past 10 years.
So let’s look at an example between the Golden State Warriors and the Cleveland Cavaliers. In a possession between each starting five: Curry-Thompson-Green-Bogut-Barnes versus James-Thompson-Love-Irving-Smith; we have an expected possession of::
Golden State: 3.368526 + 5.899544 + 3.661577 + 3.421157 + 2.231349 = 18.582153
Cleveland: 3.549414 + 1.751189 + 1.541738 + 2.397047 + 2.726413 = 11.965801
This would suggest that after 10 possessions on both ends of the court, the score should be a difference of about 10 points. Taking the coefficient variances, we find that the probability of being up 10 (despite being expected) is only 53%. However, based on the regular season data, the Warriors had a 84% chance of leading after 10 possessions. The fact that the Cavaliers led or were tied after the first ten possessions of each game serves to show how impressive the Cavaliers played in the NBA Finals.
So let’s take a look at the most valuable player on each team:
|Golden State||Klay Thompson||4.311245|
|New Orleans||Ryan Anderson||1.785153|
|New York||Carmelo Anthony||1.780164|
|Oklahoma City||Russell Westbrook||2.093660|
|San Antonio||Kawhi Leonard||2.025130|
There are a couple surprises, such as Lillard nor McCollum were tops for the Trail Blazers. They were 78 and 80th overall just behind 53rd Crabbe. Similarly, Ryan Anderson scored highest on the Pelicans.
So what do you think of the rankings system? Here, we considered overall effort on the court instead of individual statistics. How would you improve the model?