Injury forecasting in soccer utilizing machine learning and multivariate time series

Injuries have a great impact on professional soccer due to their influence on team performance and considerable costs of rehabilitation for players. In this thesis, we use injury records and workload data describing the training sessions of players in a professional soccer club, spanning two entire seasons, to train and compare three classes of approaches to injury forecasting, i.e., predicting whether or not a player will get injured in next matches or training sessions. The first class of approaches is based on traditional techniques used in sports science and industry, such as the Acute Chronic Workload Ratio. The second class is based on machine learning tools such as decision tree and k-nearest neighbor classifier. The third class of approaches extends the second class by fully exploiting the temporal information present in the data through the usage of a multivariate time series representation of a player's workload history. We demonstrate that machine learning approaches significantly outperform traditional techniques still used in sports industry, moving accuracy prediction from 4% up to 50%, paving the way to a more accurate monitoring of the health status of soccer players.