Contribution Ratings: How to assess the impact of individual player actions?
Assessing the impact of the actions that players perform in matches is a crucial aspect of the player recruitment process. Unfortunately, most traditional football statistics fall short as they focus on rare actions like shots or fail to account for the context in which the actions occurred. To address this shortcoming, the newly introduced Contribution Ratings module in Insight values any type of player action based on its impact on the scoreline.
How will a football player’s actions impact his or her team’s performances in matches? Although this question is relevant for a variety of tasks within a football club, the task of objectively quantifying the impact of the individual actions that players perform in a match remains largely unexplored to date. This task is further complicated by the low-scoring and highly dynamic nature of football matches. While most actions do not impact the scoreline directly, they often do have important longer-term effects. For example, a long pass from one flank to the other may not immediately lead to a goal but can open up space to set up a goal chance several actions down the line.
To help fill the gap in objectively quantifying the performance of football players in matches, the Contribution Ratings module values each on-the-ball action that happens in a match. Our approach goes beyond traditional football statistics such as the percentage of successfully completed passes or the total number of dribbles by considering all types of actions (e.g., passes, crosses, dribbles, take-ons, and shots). In addition, our approach also accounts for the circumstances under which players perform their actions as well as the possible longer-term effects of their actions.
Intuitively, an action’s contribution rating reflects its expected impact on the scoreline. That is, an action valued at +0.03 is expected to contribute 0.03 goals in favor of the team performing the action, whereas an action valued at -0.03 is expected to yield 0.03 goals for their opponent. For example, while a pass valued at +0.03 might not have resulted in a goal in one particular situation, it is expected to yield three goals if the action were repeated 100 times in highly similar match situations.
Broadly speaking, football players perform their actions with the intention of increasing the chance of scoring a goal or decreasing the chance of conceding a goal. Hence, a natural way to measure the effect of an action is by computing the action’s impact on the scoreline. To do so, we compare the chance of scoring and conceding a goal from the match situation before the action with the chance of scoring and conceding a goal from the match situation after the action. We used play-by-play match event data for over 50,000 matches to train two machine learning models: one model that estimates the chances of scoring a goal from a given match situation in the near future, and one model that estimates the chances of conceding a goal. To this end, we need a mathematical description of a match situation and a corresponding label for each of both models.
Describing match situations
We describe each match situation by the characteristics of the last three actions that happened as well as statistics that are computed about this sequence of actions. We distinguish between three types of information:
- Characteristics of the individual actions such as the start and end location of the action, the type of action (e.g., pass, cross, dribble, take-on, shot, and interception), the body part that the player used to perform the action, the outcome of the action (i.e., successful or unsuccessful), and the time elapsed since the start of the match.
- Statistics about the sequence of actions such as the time elapsed and distance covered between any two consecutive actions as well as for the whole sequence of actions. This information helps to capture the speed of play.
- Characteristics of the match context such as the number of goals scored by each of both teams as well as the goal difference. This information helps to address teams adapting their playing style to the scoreline. For example, a team that is 1-0 ahead might play more defensively than a team that is 0-1 behind.
Labeling match situations
We assign two labels to each match situation: a label for the model that estimates the chance of scoring a goal in the near future and a label for the model that estimates the chance of conceding a goal in the near future. For the former model, we assign a match situation a label of 1 if the team in possession of the ball scored a goal within the next ten actions and a label of 0 otherwise. In contrast, for the latter model, we assign a match situation a label of 1 if the team in possession of the ball conceded a goal within the next ten actions and a label of 1 otherwise. We empirically found that looking ten actions ahead works well in practice for training the models.
The Insight platform allows to scout interesting players based on data collected during matches. The available filters for, among others, SciSkill, Potential, market value, age and position enable to reduce the list of players to a shortlist in just a blink of an eye. The newly released Contribution Ratings module provides additional insights into how a player contributes to the performances of his team by performing actions with the ball. In particular, the module compares each player to all other players in the same position in the same competition. To do so, the module computes the player’s total contribution per 90 minutes of play given that the player played at least 700 minutes.
The following three tables present the top-five-ranked players in terms of pass contribution in the 2018/2019 Premier League season, the top-five-ranked players in terms of shot contribution in the 2018/2019 Bundesliga season, and the-five-ranked players in terms of dribble contribution in the 2018/2019 Eredivisie season.
Figure 1. Top-five-ranked players in terms of pass contribution in the 2018/2019 Premier League season.
Figure 2. Top-five-ranked players in terms of shot contribution in the 2018/2019 Bundesliga season.
Figure 3. Top-five-ranked players in terms of dribble contribution in the 2018/2019 Eredivisie season.
The Contribution Ratings module provides additional insights into how a player contributes to the performances of his team by performing actions with the ball. Furthermore, the module shows how each player rates on different aspects compared to all other players in the same position in the same competition.
A scientific paper titled “Actions Speak Louder Than Goals: Valuing Player Actions in Soccer” that describes the technicalities of the contribution ratings model will be presented at the 25th ACM SIGKDD Conference on Knowledge Discovery and Data Mining later this year. A preprint of the paper is already available to read on arXiv.
Start your 14-day trial straightaway
We always strive to attract the brightest talents