Exploration of Different Time Series Models for Soccer Athlete Performance Prediction

Abstract

Professional sports achievements combine not only the individual physical abilities of athletes but also many modern technologies in areas such as medicine, equipment production, nutrition, and physical and mental health monitoring. In this work, we address the problem of predicting soccer players’ ability to perform, from subjective self-reported wellness parameters collected using a commercially deployed digital health monitoring system called PmSys. We use 2 years of data from two Norwegian female soccer teams, where players have reported daily ratings for their readiness-to-play, mood, stress, general muscle soreness, fatigue, sleep quality, and sleep duration. We explore various time series models with the goal of predicting readiness, employing both a univariate approach and a multivariate approach. We provide an experimental comparison of different time series models, such as purely recurrent models, models of mixed recursive convolutional types, ensemble of deep CNN models, and multivariate versions of the recurrent models, in terms of prediction performance, with a focus on detecting peaks. We use different input and prediction windows to compare the accuracy of next-day predictions and next-week predictions. We also investigate the potential of using models built on data from the whole team for making predictions about individual players, as compared to using models built on the data from the individual player only. We tackle the missing data problem by various methods, including the replacement of all gaps with zeros, filling in repeated values, as well as removing all gaps and concatenating arrays. Our case study on athlete monitoring shows that a number of time series analysis models are able to predict readiness with high accuracy in near real-time. Equipped with such insight, coaches and trainers can better plan individual and team training sessions, and perhaps avoid over training and injuries.