Saturday, April 19, 2014

2013-14 Rookie Player Impact Estimates

Among the many measures aimed at capturing a player's overall value to a team (i.e., PER and WARP), NBA Stats uses Player Impact Estimate (PIE)*. The statistic combines a range of conventional stats and standardizes them per game. With the regular season complete I wanted to take a retrospective look at rookie performance and do some simple forecasting to see which players might excel in their sophomore season. I pulled PIE data from NBA Stats for all rookies that played at least 10 minutes per game. The first chart presents the results. Interestingly, based on PIE, top overall pick Anthony Bennett was actually the worst rookie in the 2013-14 season, and seventh overall pick Ben McLemore wasn't much better. The top performing rookies included Gorgui Dieng, Mason Plumlee, Michael Carter-Williams and Victor Oladipo. The latter two were effective consistently throughout their rookie years, Diegn and Plumlee came on strong during the last two months of the season.


Next, I wanted to get an idea of how PIE might predict performance among the 2013-14 rookies in their 2014-15 sophomore season. I collected data on the April PIE for 2012-13 rookies (i.e., the Anthony Davis class) and compared it with their full season PIE from their 2013-14 sophomore season. I captured only players from the 2012 draft class that played registered a PIE in April 2013 and registered a PIE for 2013-14, which resulted in 56 players (actually 57, but I dropped Robert Sacre and his -200% PIE for April 2013, which reflected 1.6 of the worst minutes of basketball in the 2012-13 season). 

My thinking is that by April, the last month of the regular season, rookies have become accustomed to the NBA game. As a result, it represents a good focal point to project performance into their sophomore year. (I'm hesitant to project anything over the long-term based on rookie performance as their are players, like Steve Nash, that need a few years to adjust to the NBA game.) I included a linear regression line, which indicates a positive correlation between rookie year April PIE and sophomore season PIE. (The anomalous point that is about 15% for his rookie season and 16% for last April is, of course, Anthony Davis.)


Finally, I pulled April PIE data for this season's rookies. The top twenty are listed below. If April's performance holds, we can expect that Carter-Williams, Plumlee, Olynyk, Diegn and Oladipo to lead next year's sophomore class. (In April, Dieng averaged a double-double in points and rebounds plus two blocks per game; Olynyk averaged almost 17 points and 7 rebounds a game while shooting 55%.)


Friday, April 4, 2014

Conventional Stats that Matter

With a growing focus on advanced statistics and measures, some analysis disregards conventional playing statistics, like steals, points or field goal attempts. But many "advanced" stats are based largely on conventional measures. For example, effective field goal %; (eFG%) simply places greater weight on three point made baskets to account for their value. More complex advanced stats, like PER, combine a range of individual and team conventional statistics, like assists, rebounds, steals, FG%, etc.

Now this isn't to say that advanced stats are deficient in any way. In many cases they are superior measures of performance--because of their ability to account for more than a single measure of play, like just rebounds, or just points--but conventional statistics are of intrinsic value.

I collected team data from the last three full NBA seasons (2011-12 was cut short 16 games by a collective bargaining agreement work stoppage) and included all conventional team stats from basketball-reference.com (for example . . . ). Next, I cheated and ran a correlation matrix to determine those conventional statistics most correlated with winning and most correlated with losing. Below are charts that show the two conventional statistics most correlated with winning and the one conventional statistic most correlated with losing. (Wins over these three seasons are plotted on the horizontal axis for the first two charts; losses are plotted on the horizontal line for the third chart.)

For each chart I included a linear regression line (blue line) with a 95% confidence region (grey area). The line and grey area give some indication of the relationship between wins and each conventional statistic. The upward slant of the regression line indicates a positive correlation between blocks (first plot) and assists (second plot) with wins. In other words, blocks and assists, rather than FG%, points, or rebounds might represent the most important indicator of a team's success (as measured by regular season wins) 



When I ran the correlation matrix, the conventional statistic most highly correlated with losing was fouls. While the association between blocks and assists with winning and fouls with losing does not necessarily hold with individual players, there is some evidence that at the team-level these are critical conventional statistics.