The Implementation of Data Science in Football

Abstract

The interaction between technical, physical, and tactical aspects of play determine success in association football (soccer). While each of these factors have been researched extensively, the explosion of sports analytics including the use of data science, has opened the door to new possibilities. On a play-by-play basis, football players must make rapid decisions that fit the paradigm of the coaches’ game plan. To successfully guide individual and team tactics, the coaching staff need to prepare for each opponent based on their strengths and weaknesses as well as their own. Historically, this has been done with notational analysis of rudimentary statistics known as event data to answer simple questions such as how many passes have been completed or of all the shots taken, how many of them have been on target. Nowadays, because of advances in technology, a new type of data knowns as tracking data, also referred to as positional data, has become increasingly available. Via local or global positioning systems (LPS/GPS), or through computer vision algorithms, positional tracking data captures the positions of all players and often the ball, at up to 25 Hz per second. Given the speed and complexity of tactical play, and the difficulty in quantifying these dynamics, positional tracking data provides an understanding of not only what occurred, but the process behind key events. Taking a dynamic systems approach to tactical analysis, this thesis aimed to increase the understanding of the capabilities of data science to evaluate and improve football performance. To fulfil this aim, a series of four studies were completed. Study 1, a narrative review of literature investigating machine learning assisted quantification of tactical play, showed that passing behaviour of football players has been examined extensively, but that most of the research lacks in practical application and data scientists keep focusing on new metrics instead of improving existing ones. As a result, there is not a clear understanding of how to integrate data driven metrics into practice during actual football training and match-play. Further, the reliance on data science and machine learning approaches for purposes of prediction has resulted in experimental designs that lack in practical application for coaches and analysts. This is problematic in that the machine learning algorithms need further refining and the gap between the aims of researchers and the needs of practitioners needs narrowing. The narrative review of literature showed that there is a need to simplify the metrics and make them more process orientated to improve transferability to the pitch. Study 2 surveyed staff members across various levels in multiple countries to determine the use and the value they find in various key performance indicators. This included an explicit assessment of twelve attacking KPIs. The findings indicate that the level of play determines how practitioners implement KPI and there was an obvious preference for simpler metrics related to shots. The low perceived value of positional tracking data driven KPIs was explained by low buy-in that can be improved with better education and collaboration between data scientists and practitioners. Study 3 expanded on Studies 1 and 2 by exploring off-ball behaviour, an area of football that could add value to the coaching process but is currently understudied. A defensive pressure model was adapted from an earlier on-ball pressure model to examine an offensive player’s ability to create separation from a defender using 1411 high-intensity off-ball actions including 988 Deep Runs (DRs) DRs and 423 Changes-of-Directions (CODs). The effectiveness of the pressure model was validated by discovering defensive pressure on the receiver at the moment of the pass was lower for completed passes than incomplete passes. Greater starting pressure on the attacker player generally led to greater subsequent decreases for DRs and CODs. There were also differences between offensive and defensive positions and the number of off-ball actions. Study 4 represented the first study to investigate the implementation of positional tracking data to improve football performance during 11v11 match-play. Using professional football players midseason, results showed the two chosen data driven metrics, D-Def and Number of Outplayed Opponent (NOO), did not significantly improve for the intervention team. However, the traditional metrics based on notational analysis such number of passes, penalty box entries, and shots on goal penalty box entries, did show greater numerical increases demonstrating a general positive outcome from the video intervention, these findings suggest future studies should aim to include a lengthier intervention that includes collaborative efforts with coaches to design training exercises that encourage behaviors that match the chosen metrics. Together, the findings supported the theoretical basis of the thesis, such that the use of positional tracking data can assist in the discovery and developmental process of tactical play. In addition, the findings provide some insight into the constraining factors on the use of positional tracking data in football. The findings have important implications for research methodology and applied practice. By quantifying a player’s effectiveness when they do not have the ball, coaches are equipped with a unique way of providing feedback to such an important part of the game. Further, the novel use of positional tracking data to quantify off- ball behaviour as well as passes throughout this thesis shows the value of these methods for future investigations, especially in the ability to quantify process orientated aspects of play. While this thesis contains limitations in the design, the theoretical underpinnings, methodology, and findings of this thesis provide a platform for future investigations involving positional tracking data in football.