1 of 8

Expected Value Models

2 of 8

Soccer

  • Model expected value of possession given location ball, players, defenders, etc.
  • Assign players value based on successful passes/movements of the ball/converted shots

https://arxiv.org/pdf/2011.09426.pdf

3 of 8

Soccer

  • For example, can break pitch up into boxes, calculate the probability of the possession resulting in a goal for each box
  • Treat soccer as a Markov chain, an idea introduced by Sarah Rudd (http://nessis.org/nessis11/rudd.pdf)
  • If a player dribbles from one box to the other, or completes a pass into a different box, they are rewarded with that value
  • Exepected Threat takes this one step further (https://karun.in/blog/expected-threat.html)

https://statsbomb.com/2019/02/attacking-contributions-markov-models-for-football/

4 of 8

Markov Chains

  • Finite number of possible states
  • Probability matrix tells you the probability you go from state B given that you’re in state A
  • Markov assumption: these probabilities only depend on your current state, no the history of the system/how you got there

https://qph.fs.quoracdn.net/main-qimg-feb4de879cdea1e7eafab76c4c55d120.webp

5 of 8

Basketball

  • Expected possession value computed from spatial tracking data
  • Use EPV at start of player’s possession minus EPV at end of player’s possession to evaluate player actions

https://arxiv.org/pdf/1408.0777.pdf

6 of 8

Basketball

  • Diagnostic tool could help indicate bad decisions:
    • Shoot/not shoot?
    • Who to pass to?
  • Easier to improve decisions than improve skill

https://arxiv.org/pdf/1408.0777.pdf

7 of 8

Volleyball

  • Each hit in rally as a state in a Markov model

  • UTSPAN project: obtained data from UofT women’s volleyball team with each hit in a rally and the quality
  • Use expected point value before/after a player’s touch to assign credit
  • E.g. If opponent just did a “good” attack, we have P(win point) = 0.3, if a player responds with a “good” dig, maybe that goes up to P(win) = 0.5, so that player gets credited 0.2

https://docs.lib.purdue.edu/dissertations/AAI10607243/

https://arxiv.org/pdf/2402.01083.pdf

8 of 8

Football

  • What is the equivalent of EPV in football?
  • Can we model football with a Markov chain?
  • Who will win Super Bowl LVIII