Today I decided to modify David Berri's Win Score (WS) and Dan Rosenbaum's Alternate Win Score (AWS), with a couple SportVu derived metrics. Specifically, I replaced traditional rebounds and assists, with contested rebounds and potential assists, so as to get a different Vu (I had to) of player production.
To this point in the season, about half of all potential assists become an assist. Though I would rather try to derive the value of a potential assist from "scratch", I decided to just make the potential assist value +.25 (half of the WS/AWS value of an assist), since this study is Win Score focused (WS/AWS places the assist value at +.5).
Since possession based linear metrics like WS/AWS are all about gaining possessions, it made sense to count all contested rebounds as +1. Non-contested rebounds were worth 0, and there was no penalty for missed rebounds. That's because a player trying for and missing a rebound, likely didn't prevent a teammate from trying for and grabbing the rebound (which is the case for missed shots).
To be "safe" I added a position adjustment to both WS and AWS, as well as their SportVu derivatives, so I could compare metrics and players without fear of an overt position bias (which I suspect of the SportVu WS/AWS derivatives). I made all scores per-48 minutes, and filtered out players with less then 10-games played, and 20 minutes per game, which left me with a pool of about 210 players. Observe the key.
WS48 - Position Adjusted Win Score
AWS48 - Position Adjusted Alternate Win Score
WS-VU - SportVu modified Win Score
AWS-VU - SportsVu modified Alternate Win Score
Below are a few correlations (R2 value displayed).
To this point in the season, about half of all potential assists become an assist. Though I would rather try to derive the value of a potential assist from "scratch", I decided to just make the potential assist value +.25 (half of the WS/AWS value of an assist), since this study is Win Score focused (WS/AWS places the assist value at +.5).
Since possession based linear metrics like WS/AWS are all about gaining possessions, it made sense to count all contested rebounds as +1. Non-contested rebounds were worth 0, and there was no penalty for missed rebounds. That's because a player trying for and missing a rebound, likely didn't prevent a teammate from trying for and grabbing the rebound (which is the case for missed shots).
To be "safe" I added a position adjustment to both WS and AWS, as well as their SportVu derivatives, so I could compare metrics and players without fear of an overt position bias (which I suspect of the SportVu WS/AWS derivatives). I made all scores per-48 minutes, and filtered out players with less then 10-games played, and 20 minutes per game, which left me with a pool of about 210 players. Observe the key.
WS48 - Position Adjusted Win Score
AWS48 - Position Adjusted Alternate Win Score
WS-VU - SportVu modified Win Score
AWS-VU - SportsVu modified Alternate Win Score
Below are a few correlations (R2 value displayed).
Here are the top 30 players (who qualify) in terms of WS-VU, with their AWS-VU shown on the left.
- The correlation between the WS-VU ad AWS-VU was 91%, which didn't seem low enough to warrant a separate table sorting by AWS-VU.
- The correlations between WS/AWS and their SportVu derivatives were all pretty high. The position adjustment in everything probably helped.
- Derrick Rose ranked last among both WS-VU and AWS-VU.
That's it. Questions and recommendations are always welcome, as I fine tune and try to get this in optimal shape by the All-Star break. Then I'll start seeing how well the Sport-Vu WS/AWS derivatives explain wins, and what type of defensive adjustment I could install to get a Wins-Produced like correlation.
- The correlations between WS/AWS and their SportVu derivatives were all pretty high. The position adjustment in everything probably helped.
- Derrick Rose ranked last among both WS-VU and AWS-VU.
That's it. Questions and recommendations are always welcome, as I fine tune and try to get this in optimal shape by the All-Star break. Then I'll start seeing how well the Sport-Vu WS/AWS derivatives explain wins, and what type of defensive adjustment I could install to get a Wins-Produced like correlation.