Modelling team performance in soccer using tactical features derived from position tracking data

Abstract

Abstract Decision-makers in soccer routinely assess the tactical behaviour of a team and its opponents both during and after the game to optimize performance. Currently, this assessment is typically driven by notational analysis and observation. Therefore, potential high-impact decisions are often made based on limited or even biased information. With the current study, we aimed to quantitatively assess tactical performance by abstracting a set of spatiotemporal features from the general offensive principles of play in soccer using position tracking data, and to train a machine learning classifier to predict match outcome based on these features computed over the full game as well as only parts of the game. Based on the results of these analyses, we describe a proof of concept of a decision support system for coaches and managers. In an analysis of 302 professional Dutch Eredivisie matches, we were able to train a Linear Discriminant Analysis model to predict match outcome with fair to good (74.1%) accuracy with features computed over the full match, and 67.9% accuracy with features computed over only 1/4th of the match. We therefore conclude that using only position tracking data, we can provide valuable feedback to coaches about how their team is executing the various principles of play, and how these principles are contributing to overall performance.