The outcome of football games, as well as matches of most other popular team sports, depends on a combination of the skills of players and coaches and a number of external factors which, due to their complex nature, are presumably best viewed as random. Such parameters include the unpredictabilities of playing the ball, the players' shape of the day or environmental conditions such as the weather and the behavior of the audience. Under such circumstances, it appears worthwhile to analyze football score data with the toolbox of mathematical statistics in order to separate deterministic from stochastic effects and see what impact the cooperative and social nature of the "agents" of the system has on the resulting stochastic observables. Considering the probability distributions of scored goals for the home and away teams, it turns out that especially the tails of the distributions are not well described by the Poissonian or binomial model resulting from the assumption of uncorrelated random events. On the contrary, some more specific probability densities such as those discussed in the context of extreme-value statistics or the so-called negative binomial distribution fit these data rather well. There seemed to be no good argument to date, however, why the simplest Poissonian model fails and, instead, the latter distributions should be observed. To fill this gap, we introduced a number of microscopic models for the scoring behavior, resulting in a Bernoulli random process with a simple component of self-affirmation. These models allow us to represent the observed probability distributions surprisingly well, and the phenomenological distributions used earlier can be understood as special cases within this framework. We analyzed historical football score data from many leagues in Europe as well as from international tournaments, including data from all past tournaments of the "FIFA World Cup" series, and found the proposed models to be applicable in all cases. To complete the picture, we conducted a field study with visitors of a science showcase to collect additional data from matches of tabletop football. As it turns out, also the latter data are represented well with our feedback models, underscoring their apparently rather universal applicability.
Bibliographical noteThe full text is available from: http://dx.doi.org/10.5488/CMP.12.4.739
- sport statistics
- extreme-value distributions
- feedback models