Predicting soccer game evolution through AI-based tracking data analysis

Nowadays, technology is increasingly used in soccer. An open challenge is how to use the massive data produced by technology to create a framework to simulate different match situations and help trainers understand the dynamics on the field better. This thesis aims to extrapolate logical patterns that describe how the ball moves on the field in different game situations. We use tracking and event data of several matches to extract players and ball positions on the field. Then, we build two machine learning approaches. The first approach involves the use of handmade features passed to a Random Forest classifier. The second approach is a Convolutional Neural Network that automatically highlights valuable features to make a prediction. We show that the Random Forest provides a better understanding of the rules governing the movement of the ball than the Convolutional Neural Network. This result emphasizes that conditional control statements based on the position of the object on the field alongside handmade features work better than an automated feature extraction method based on deep learning.