Extended Model for Expected Threat in Football
Conference presentation co-authored with Ville-Pekka Inkilä and given at the New England Symposium on Statistics in Sports.
In soccer, expected goals (xG) models are used to estimate the probability of scoring when a shot is taken from a given location. Expected threat (xT) models combine xG models with soccer dynamics by introducing a stochastic model for the next ball event (move the ball or shoot). In xT models, moving the ball to another location by dribbling or passing is modeled as a Markov Chain. If a shot is taken, the scoring probability is computed as in xG models. Once the probabilities for both the Markov and xG model have been estimated from event data, the xT model can be used to compute an xT value (i.e., the probability of the team scoring during the next few events) for each field location.
A crucial shortcoming of existing xT models is that they exclude the possibility of turnovers, whereby the analysis is focused entirely on scoring goals. Moreover, these models are limited to actions taking place during only a few subsequent events and the part of the field from which a goal can be expected to be scored during those events.
We present an improved xT model which accommodates turnovers and subsequent ball movement as well as the negative xT posed by the opponent coming into the possession of the ball. Our model also makes it possible to consider all the events up to the next shot (towards either of the goals), thereby covering the entire field. We illustrate the model using real data from Stats Perform.
