Creator: KU Leuven DTAI, Method: Gradient-boosted trees, Window: Next 10 actions, Off-ball information: Possession history proxies

University of Michigan - Ann Arbor

It is common for skilled soccer teams to control the ball for an entire game but never score a goal. Possession models look to further quantify what kinds of possessions are actually valuable. A simple version is a location-based possession value, which estimates the probability of scoring based on where the ball is at that point in time. As the ball gets closer to the goal, that probability increases.

Soccer Possession Models

Creator: S. Rudd
Method: Markov chain
Window: One possession
Off-ball information: Defensive states tagged in event data

Markov Chains

Creator: N. Mackay, Method: Logistic regression and GAM, Window: One possession

Possession-Based Model

Creator: K. Singh, Method: Markov-line, Window: Next 5 actions

Expected Threat (xT)

Valuing Actions by Estimating Probabilities (VAEP)

Creator: J. Fernandez, Window: Next goal, Off-ball information: Full tracking data

Expected Possession Value (EPV)

Possession Value (PV) is a soccer possession-value model developed by Stats Perform. It uses gradient-boosted trees to estimate the value of a possession over a forward-looking window of the next 10 seconds, incorporating off-ball context through possession-history proxies as a substitute for full tracking data.

Possession Value (PV)

Creator: American Soccer Analysis, Method: Gradient-boosted trees, Window: Two possessions, Off-ball information: Possession history proxies 

Goals Added (g+)

Creator: StatsBomb, Method: Gradient-boosted trees, Window: Two possessions, Off-ball information: Broadcast freeze frame

Learn Before

Related