- Joined
- Jul 6, 2009
- Messages
- 11,107
- Reaction score
- 21,301
- Points
- 135
For those of you who are interested,
Cranjis McBasketball (Lakers analyst who is big into quantitative data) has created a new NBA stats website: bball-index.com.
http://bball-index.com/18-pipm/
http://bball-index.com/about/about-the-data/
He has a few cool stats on there -->
Cranjis McBasketball (Lakers analyst who is big into quantitative data) has created a new NBA stats website: bball-index.com.
http://bball-index.com/18-pipm/
http://bball-index.com/about/about-the-data/
He has a few cool stats on there -->
PLAYER IMPACT PLUS-MINUS
Player Impact Plus-Minus is a metric that combines traditional boxscore value with luck-adjusted on/off player data to estimate how much value a player adds to their team.
Luck-adjusted data, developed by Nathan Walker, is used to adjusted for factors that are out of an individual team or player’s control. For instance, free throw shooting and three point shooting can cause wide variance in the specific ratings, but in studies it has been shown that teams and players have limited control over makes or misses. Another example is adjusting for rebounding and turnovers to attempt to limit the noise from the final values.
The boxscore component is calculated off a regression from a 15-year RAPM sample. Especially on offense, there is real value to be found in the traditional boxscore. Combining that with more advanced play by play data, PIPM is able to see who is adding value that the boxscore is unable to capture.
POINTS OVER EXPECTATION
Points Over Expectation (POE) is a metric that tries to explain the following question: How well does a player perform compared to how we’d expect an average player to perform?
To do this, POE turns to data on the NBA’s website about Synergy play types and uses a regression or two along with some more basic math to calculate values. League average points per possession data is used to set the bar for expected performance.
POE = Offense (positive values good) – Defense (positive values bad)
Offense – Created Points Over Expectation
The offensive part of points over expectation is simple, and is called Created Points Over Expectation. It looks just as scoring. Positive values are good.
Here’s the 30 second version: If a Spot Up possession yields 1 point per possession on average, and Kevin Durant has 5 spot up possessions in a game, he’d be expected to score 5 points. If he scores 7, he has a +2 spotting up CPOE. If he scores 4 points in those 5 possessions, his spotting up CPOE is -1.
To get your total game CPOE, do that for every play type and add them together. That’s it.
CPOE is simple and adjusts for role in a way that overall points per possession does not. 1 point per possession could be great or horrible efficiency compared to expectations, depending on role. If it’s on all putbacks, you’re slacking. If it’s all in the pick and roll as the ball handler, you’re doing well. CPOE accounts for role in that way.
It’s not precise, as there are many factors that would ideally go into expectation other than just play type, but it’s what can be done with that simple data and does a decent job at evaluating scoring performance vs expectation.
Defense – Defensive Points Over Expectation
On the defensive end, we use total Defensive Points Over Expectation (tDPOE). It’s similar to CPOE but with a slight difference in calculation. It also different from CPOE in that positive values are indicative of below average performance.
Synergy play type defensive data accounts for primary defense, which allows primary defense to be calculated the same as CPOE. This is Defensive Points Over Expectation.
Help defense can’t be captured in that data, and thus a regression with PIPM is used to help identify residual values by player, which are used as inputs to calculate help defense. This component of defense was recently added, and may be referred to as help defensive DPOE.
Throw DPOE and help DPOE together and you get total defensive points over expectation. Just like CPOE, tDPOE adjusts for role, allowing for a more detailed look at performance by adding an expectation.
The most important thing to remember regarding these grades is that they are attempting to use every publicly available statistic to describe specific skills that a player has and to measure how good or bad the player is at that skill. These grades aim to take all of the added layers of information (coaching, scheme, teammates, etc.) and strip them back to just the underlying skills that a player has.
Players are graded on seven different offensive skills: Perimeter Shooting, Off-Ball Movement, One on One, Finishing, Roll Gravity, Playmaking, and Post Play.
Perimeter Shooting
A player’s perimeter shooting is grades on two factors: how good a player is at three point shooting compared to the openness of their shots and at what volume a player attempts three point shots. Openness is calculated using data from NBA.com/stats, effectively calculating an expected 3P% based off shot openness to compare to actual results. Volume is accounted for using a sigmoid function to regress low volume players. A sigmoid function was selected because it does not impact high volume players, while creating an S curve for regressing lower volume players towards 0. This category is aiming to grade what players best converted their three point attempts relative to how easy their attempts were.
Off-Ball Movement
Off-ball movement is graded using player tracking average velocity data and Synergy play type data. Average offensive velocity is a small portion of the final grade, but it’s included to improve accuracy players who were constantly moving well on offense. Cutting and off-screen Synergy play type data is adjusted for efficiency and volume on the play type. Again, a sigmoid function is utlizied to regress low volume players for Synergy data and low playing type players for movement data. This category is aiming to grade what players moved well without the ball and used that movement to create off-ball scoring opportunities.
One on One
One on one is graded using NBA driving data along with Synergy play type data. Driving data from tracking cameras is adjusted for efficiency and frequency. Isolation and handoff Synergy play type data is adjusted for efficiency and frequency as well. This category is aiming to calculate what players best used the ball in their hands to apply pressure on the defense and create individual scoring opportunities.
Finishing
Finishing is graded using rim shooting data along with Synergy play type data. Shots within three feet of the rim and any dunk or layup attempts that occur outside of that range are adjusted for how often they are assisted versus unassisted and the volume at which a player shoots. Putback play type data from Synergy is used with an adjustment for volume, but none for efficiency. The rationale behind not adjusted putbacks for efficiency is that they are effectively free attempts at points and thus the expectation on those attempts is effectively 0 points. This category is aiming to calculate what players best made use of their opportunities around the rim, accounting for how those shots are created.
Roll Gravity
Roll Gravity is a newer concept that we have introduced as a part of the grades we put together. The idea behind them is simple: in the same way that a great shooter can pull defenders out of the paint when they’re off-ball, a great roll man can suck defenses into the paint to cover them. It’s an attempt to quantify a player’s vertical spacing.
Roll Gravity is graded using Alley-Oop data, NBA screen assist data and Synergy play type data. Alley-oop data is adjusted for volume and efficiency, and the specific play type is a great indicator of how much verticality a player is offensively able to provide in the middle of the court. Screen assist data, adjusted for volume to regress small samples, is used to indicate what players are best creating space for the ball handler in a pick and roll. Screen assists help clarify what players are creating opportunities for their teammates by setting a good screen that creates space for a teammate to score. Roll man Synergy data is used to measure an individual player’s efficiency when they receive the ball back from a pick and roll. It is adjusted for both volume and efficiency. This category is aiming to calculate what players do the best job creating vertical spacing for their team.
Playmaking
Playmaking is graded using NBA passing data, Synergy data and PBPStats passing data. From NBA passing data, and along with offensive role data calculated using this methoddeveloped by Todd Whitehead, an expectation for assists can be calculated using teammate efficiency, potential assists and passes. This expectation is compared to actual assists, allowing to measure how good a player was at creating assists within their offense. The other main factor is box creation, developed by Ben Taylor, which attempts to measure a player’s ability to create scoring opportunities for teammates outside of using just assists. This category is aiming to calculate how good a player is at creating opportunities for teammates with their passing, with role and teammates accounted for.
Post Play
Post play is graded using Synergy play type data. Post up data from Synergy is compared to expected efficiency of the play and adjusted for volume attempted. Putback data is adjusted the same way it is for Finishing, accounting for the fact that they are effectively free points for the offense if they can convert. This category is aiming to calculate how good a player is at post activities.
Perimeter Defense
Perimeter Defense is graded using Synergy play type data, NBA contest data, NBA movement tracking data, NBA hustle data and luck-adjusted on/off data. Perimeter defense is by far the most complex category to attempt to grade on an individual level. So much of perimeter scheme is about bending, not breaking and forcing opponents to shoot the shots they want. Furthermore, not all steals are equal. Some players gamble for steals, and while they may end up with two per game, they also may end up giving up three easy scores to the opponent because they gambled their way out of position. Despite these difficulties, the process for grading perimeter defense is as follows.
Three point shot contest data from the NBA is used to measure a player’s ability to get in position to make a shot more difficult. A Synergy blend consisting of pick and roll ball handler, isolation, handoff and offscreen is used to compared to expectation and volume in each play type is used to look at specific defensive actions, though this blend is a small portion of the final perimeter defense grade because of the well known errors in Synergy defensive data. NBA defensive movement tracking data is used as a proxy for players who are active on the defensive side of the court, however it is used as a small factor in the overall grade because activity does not equal ability.
The three other components of perimeter defense are steals per 75 possessions, regressed for small samples, deflections per 75 possessions, also regressed, and luck-adjusted defensive on/off data. The largest amount of weight goes to steals, deflections and three point shot contests.
Interior Defense
Interior Defense is graded using NBA defensive rim tracking data and luck-adjusted on/off data. With interior defense, there are strong statistics for measuring rim defensive field goal attempts at the team level and individual level. Rim attempts contested are adjusted to account for how often each team allows a shot at the rim. Points at the rim allowed is adjusted for this frequency of rim attempts allowed and compared to the expected value of those shots if there was no contest at the rim. These are then adjusted another time for volume of attempts, to pull small samples down towards zero, and the expected value of a contest on the final total of points saved. The process is a slightly more complex version of what Seth Partnow, now with the Bucks, outlines here. Points saved from attempts at the rim is then combined with luck-adjusted defensive on/off data to incorporate a player’s defensive impact on defense overall with their specific impact for shots around the rim.
Offensive and Defensive Rebounding
The process for both offensive and defensive rebounding are nearly identical with one exception: the inclusion of putback Synergy data in offensive rebounding grades. Both use rebounding on/off data and NBA tracking rebounding data. NBA adjusts rebounding chance for deferred opportunities to create an efficiency of a player when they’re in a position to secure a rebound. This is than adjusted for volume of rebounds secured with double weight being given to contested rebounds. After regressing small sample players, this is blended with putback points per possession and rebounding on/off data to create a rebounding talent grade.
