Thread by @JanVanHaaren

Last month I had the pleasure of giving a keynote talk at @BarcaInnoHub's Barça Sports Analytics Summit. I presented a few event-data based football metrics that we developed for player recruitment. My presentation slides are available on Google Drive.

docs.google.com/presentation/d…

🗣️ "Working with many football organizations across the globe, @SciSportsNL primarily focuses on helping football clubs to identify appropriate transfer targets. Football clubs' recruitment departments are typically faced with a multitude of questions which they need to answer."

🗣️ "To help answer the scouts' questions, we developed a whole range of performance and style metrics for football players. These metrics enable us to compose a data-driven profile for about any professional or semi-professional football player in the world."

🗣️ "From an analytics perspective, we distinguish between three types of data: matchsheet data, ball event data and tracking data. Typically, the availability of the data decreases as the granularity of the data increases."

🗣️ "We leverage ball event data to develop our player recruitment metrics. Ball event data is widely available, becomes increasingly information rich, and is easier to process and analyze than spatio-temporal player and ball tracking data."

🗣️ "To value the majority of the actions that football players perform with the ball such as passes, traditional statistics fall short as they fail to account for the context in which the players perform these actions."

🗣️ "To overcome this limitation of traditional statistics, we developed an approach to value on-the-ball player actions that takes the context into account. To value a given action, we compute the difference between the values of the game states before and after the action."

🗣️ "Hence, the key challenge is to determine the value of a given game state. We address this challenge by answering two underlying questions: How likely is it that the team in possession will score a goal from the game state? How likely is it that they will concede a goal?"

🗣️ "Intuitively, we determine the likeliness of scoring and conceding a goal by observing the outcomes of highly similar game states in the past. For example, if 3 out of each 100 similar game states in the past led to a goal, the corresponding likeliness of scoring is 0.03."

🗣️ "We describe a football game as a long sequence of events. Similarly, we describe a game state as the subsequence of the last three events that happened in the game. The visualization shows the game states before and after the Busquets pass in our earlier example."

🗣️ "To feed game states into a machine learning algorithm, we describe each game state as a feature vector and assign a label. Our feature vector captures characteristics of the actions and important contextual information. The label is 1 if a goal is scored within 10 actions."

🗣️ "We repreat the previous procedure for tens of millions of game states and train a probabilistic classification model using @CatBoostML. Using the trained models, we predict the likelihood of scoring or conceding a goal from the game state under consideration."

🗣️ "By valuing the pre-action and post-action game states using the trained machine learning models, we obtain a Contribution Rating of 0.03 for the Busquets pass in our earlier example."

🗣️ "Although many action valuation frameworks have been proposed, our framework is likely the most powerful such framework out there. Our framework is fully data-driven, considers both offensive and defensive contribution, and captures game states beyond just pitch location."

🗣️ "Our framework enables a number of player recruitment use cases such as identifying players who stand out in a region or in a given competition. For instance, Frenkie de Jong obtained the highest Contribution Rating for passes of all midfielders in the 2018-2019 Eredivisie."

🗣️ "In addition to the Contribution Rating, our framework enables a few derived metrics. The Decision Rating evaluates whether a player selected the best possible action in a game state. The Execution Rating evaluates how well a player performed the chosen action."

🗣️ "Football players often need to perform in high-mental-pressure situations, especially when playing for clubs that compete for the championship or battle against relegation. Hence, we developed a metric that evaluates how mental pressure affects the performances of players."

🗣️ "Intuitively, our framework compares player performances between low-mental-pressure and high-mental-pressure situations. To this end, we developed a data-driven mental pressure model that estimates both the pre-game and in-game mental pressure experienced by the players."

🗣️ "This visualization shows the pre-game mental pressure levels for the Everton players throughout the 2017-2018 Premier League season. Clearly, our pre-game mental pressure model responds to the events that happened over the course of the season."

🗣️ "This visualization shows the in-game and combined mental pressure levels for the Everton and Watford players in their game on November 5th, 2017. Clearly, both our in-game and combined mental pressure model respond to the events that happened over the course of the game."

🗣️ "Our results show that Neymar is one of the world's best players when pressure is low. However, Neymar is just an average player when pressure is high. In particular, he makes much poorer decisions when the mental pressure mounts."

🗣️ "The take-away messages are the following: First, event data are well suited for player recruitment metrics. Second, our action valuation framework is likely the most powerful framework out there. Third, our framework enables a multitude of applications and other metrics."

🗣️ "If you want to learn more about this joint work with @TomDecroos, @LotteBransen, @p_robberechts and @jessejdavis1, then check out our papers! An open-source implementation of our action valuation framework is available on GitHub: github.com/ML-KULeuven/so…."