Proposal of offense-defense WPA metrics in B.LEAGUE

Presentation sildes for ASCC2022 (https://acss.ear.com.sg/)

Absgtract: Classical stats recorded in basketball box scores of matches attribute contributions to the player who handled the ball last in the play. Therefore, it is difficult to quantify each player’s defensive contribution using classical stats.
Play-by-play data is recorded and disclosed in several professional basketball leagues. Advanced metrics based on this data have been extensively investigated to quantify each player’s contributions to their team’s victory. This study focuses on win probability added (WPA). In basketball, WPA can
be defined as the difference between the predicted winning ratio before and after a single play.
The probability difference is equally distributed among all players on the court. WPA can be calculated separately in offense and defense situations.
In this study, play-by-play data disclosed by B.LEAGUE, a men’s professional basketball league in Japan, was collected and analyzed. In the first step, we developed a predicted win probability model that guarantees monotonicity on the score difference. Next, we calculated the offense/defense WPAs
and their time averages. These metrics quantified each player’s contribution to winning. Moreover, they expressed player features such as the pace of their plays.

Transcript

1. Proposal of offense/defense WPA metrics
   in B.LEAGUE
   Koji SUGIE*, Eiji KONAKA (Meijo Univ. Japan)
   

   View Slide

2. Outline
   Background
   STATS in basketball
   Impact metrics
   Proposed method
   Construction of consistent win probability model
   Main proposal: Offense/Defense Win Probability Added
   Numerical experiment
   Summary and future work
   

   View Slide

3. Background: stats in basketball
    (Classical) stats: counts and ratios of events.
    Attributed to the player who touched the ball last.
   

   View Slide

4. Background: stats in basketball
    Detailed play-by-play records
    Available in Japanese
   professional B.LEAGUE
   Stats and records can be used to evaluate each player’s
   contribution
   

   View Slide

5. Background: contribution evaluation from
   stats
   How to evaluate the
   contributions of the
   players who did not touch
   the ball?
   https://www.youtube.com/watch?v=qKde9M1gKvY&t=1311s
   2022/07/18 accessed
   

   View Slide

6. Background: impact metrics
   “Impact metrics (IM)” tries to evaluate each player’s
   contribution in single value.
   Example: +/-, EFF, PIPM,LEBRON,RPM,RAPTOR, …
   Problem:
   Simple IM: do not include situation (time, score diff, …)
   Advanced IM: calculation methods are not fully disclosed in academic
   context (“magic recipe”)
   

   View Slide

7. Background: win probability model
    How to include situation?
    Time and score difference
    Win probability model
    Problem
    Consistency
   Large score diff<->Large win prob.
    Previous studies: Inconsistent win
   probability model (Deshpande[2016])
   Deshpande[2016]
   Time[s]
   Score difference]
   

   View Slide

8. Research flow and main objective
   • Construct consistent win probability
   model
   • Contributions to win of all players is
   evaluated
   • Numerical evaluation
   • Japanese B.LEAGUE
   

   View Slide

9. Outline
   Background
   STATS in basketball
   Impact metrics
   Proposed method
   Construction of consistent win probability model
   Main proposal: Offense/Defense Win Probability Added
   Numerical experiment
   Summary and future work
   

   View Slide

10. Consistent win probability model with
    score difference
     Win probability model ෝ
    𝑤 𝑡, Δ𝑠
     𝑡: time[s] from tipoff
     Δ𝑠: score difference
     Dataset
     B1 League, 1736 matches
     2016/9-2019/10 (3 regular seasons)
    

    View Slide

11. What is “consistency”
     Consistent win probability model
    ෝ
    𝑤 𝑡, Δ𝑠
     ෝ
    𝑤 𝑡, Δ𝑠 is consistent ⇔ ෝ
    𝑤 is
    monotonic as Δ𝑠
     Logistic regression is done for 𝑡 =
    0, ⋯ , 2400
    

    View Slide

12. WPA originates in baseball (MLB)
    Offense WPA(WPA𝑂): sum of the change of ෝ
    𝑤 in offense situation
    Offense situation=ball possession
    Defense WPA(WPA𝐷): sum of the change of ෝ
    𝑤 in defense situation
    Defense situation=non-ball possession
    Total WPA：WPA = WPA𝑂 + WPA𝐷 for each player
    Definition of WPA (Win Probability Added)
    Can evaluate the contributions of the players
    who did not touch the ball
    

    View Slide

13. Example of WPA
    O/D ෝ
    𝒘 𝒔𝒊
    𝒕 𝒔𝒋
    ෝ
    𝒘 O/D
    O(Start) 0.65 18 702 9 0.35 D(Start)
    O(end) 0.72 21 722 9 0.28 D(end)
    D(start) 0.72 21 722 9 0.28 O(Start)
    ・
    ・
    ・
    WPA𝐷 in Team 𝑗:
    0.28 − 0.35
    5
    = −0.014
    WPA𝑂 in Team 𝑖:
    0.72 − 0.65
    5
    = 0.014
    ・
    ・
    ・
    Team 𝑖 Team 𝑗
    ෝ
    𝑤 722,12 = 0.72
    

    View Slide

14. Outline
    Background
    STATS in basketball
    Impact metrics
    Proposed method
    Construction of reasonable win probability model
    Main proposal: Offense/Defense Win Probability Added
    Numerical experiment
    Summary and future work
    

    View Slide

15. Numerical experiment
    268 players in B1 League, 2018-19 regular season
    Following example
    Teams: Chiba (52 wins) and Tochigi (49 wins)
    Played more than 10000[s] (about 3[min/match])
    

    View Slide

16. Numerical experiment
    Team: Chiba Team: Tochigi
    

    View Slide

17. Discussion and analysis: pace
     Team characteristics
     WPA/s ~ Points/s
     Tochigi has larger WPA𝑂/𝑠 and
    WPA𝐷/𝑠
     Tochigi favored high-paced
    transition game
    WPA/s (Pink: Tochigi, Green: Chiba)
    

    View Slide

18. Discussion and analysis: PGs
    Team: Chiba Team: Tochigi
    Togashi
    Nishimura Watanabe
    Ikaruga
    

    View Slide

19. Discussion and analysis: PGs
    Team: Chiba Team: Tochigi
    high
    Low
    high
    high
    

    View Slide

20. Discussion and analysis: weakness of
    bench players
    ONO Ryumo
    Bench player
    Position: SF/PF
    Similar WPA𝐷 , but small
    𝑊𝑃𝐴𝑂 compared with the
    regular SF and PF
    Team: Chiba
    Ono
    

    View Slide

21. Summary
    ☑
    • Construct consistent win probability model
    ☑
    • Contributions to win of all players is
    evaluated
    ☑
    • Numerical evaluation in Japanese
    B.LEAGUE
    

    View Slide

22. Future work
    ☑
    • Construct consistent win probability model
    New!
    • Contributions to win of all lineup will be evaluated
    • Pairwise comparison model
    New!
    • Numerical evaluation in Japanese B.LEAGUE
    

    View Slide

23. Many thanks! And welcome
    questions and comments!
    Sugie, Konaka. “Proposal of offense/defense WPA metrics in B.LEAGUE”
    

    View Slide