Forecasting extremes of football players' performance in matches.

Abstract

Based on training history, accurate predictive modelling of an athlete's performance in competition can be a cornerstone for prior optimal planning of exercise mix and intensity. While universal in many sports, such a goal is challenging in football due to the complexity of factors leading to the final score and the complexity of the preceding training process. We developed and tested a range of models, the best of which were to forecast selected play performance indices with an accuracy of 10-20%. Such score applies to models run on raw player location data and aggregating performance indices developed with expert knowledge in the football training domain. Results show that individual player models perform better than collective ones and that more recent training data are better predictors. While we consider the accuracy of the models still of limited reliability, their transparency and present quality make them useful in the daily planning of training activities that impact player performance in the coming match. Additionally, observations of training parameters generated in short-term intervals are more effective and correlated with extreme match results than long-term dates. Specific training parameters may be key in predicting exceptional football player performance, but they may also vary from person to person.