Latent-state modelling of momentum in sports based on events and their informative time points
Conference
65th ISI World Statistics Congress 2025
Format: IPS Abstract - WSC 2025
Keywords: latent, markov, maximum likelihood, modelling, poisson process, sport, states
Session: IPS 813 - Sports Analytics (2 of 2)
Thursday 9 October 2 p.m. - 3:40 p.m. (Europe/Amsterdam)
Abstract
In many sports, momentum shifts – i.e. changes in the match dynamics over time – are characterised by, for example, teams switching between phases of pressure/attack with many goal attempts as opposed to phases of defence with few attempts. To investigate such momentum shifts in sports, latent-state models can be used, which assume the current match dynamics (e.g. the level of pressure a team exerts on its opponent) to be an unobserved state underlying the observed match events (e.g. goal attempts). Crucially, the time differences between consecutive events provide information on the state process, as events like goal attempts occur more frequently when a team exerts pressure, leading to clustered observations. For modelling these so-called informative event times driven by underlying latent states, the class of Markov-modulated Poisson processes (MMPPs) provides a flexible framework. Additionally, event outcomes (so-called marks; e.g. whether a goal attempt was successful or not) and their informative time points can be modelled jointly by extending MMPPs to Markov-modulated marked Poisson processes (MMMPPs). In our contribution, we apply MMMPPs to analyse sports data and detect momentum shifts within matches.