What can AI do for football analytics and what can football do for AI?

Football as a two-way testbed for AI

The paper's answer runs in both directions. For football, AI offers predictive and prescriptive tools that can support decisions by managers, coaches, and players; for AI, football supplies a demanding, data-rich environment where statistical learning, computer vision, and game theory must work together rather than in isolation. The authors describe this combined domain as a microcosm for AI research.

That framing is a research agenda, not a claim that an integrated system already exists. The paper reviews the state of the art, identifies where the three fields intersect, and illustrates one intersection in depth. The authors' long-term target is an Automated Video Assistant Coach (AVAC): a system that would process raw broadcast footage and advise coaching staff before, during, and after matches. The AVAC is presented as a vision of future human-centric AI research, not as a deployed product.

Why football arrived late to data science

The paper contrasts football with baseball, basketball, and tennis, where predictive analytics has been investigated for decades and baseball data has been collected systematically since the 19th century. Football's data-science transformation is described as still in its infancy, with decisions about transfers, scouting, and pre- and post-match analysis only recently being informed by data.

The authors attribute the delay to the game's structure and to its culture. Football is played outdoors on a larger pitch, with many players, few substitutions, and longer uninterrupted sequences than basketball, making it harder to control and analyze. Credibility in football decision-making also long rested with managers, retired players, and scouts rather than data specialists. The paper quotes Kuper and Szymanski's remark that "until very recently, soccer had escaped the Enlightenment" to capture this cultural hesitation.

Three frontiers where the AI fields meet

The paper organizes football analytics around three foundational AI areas: statistical learning (SL), computer vision (CV), and game theory (GT). Each has already contributed on its own. Statistical learning tools such as principal component analysis identify player types, models predict trajectories, and actions such as passes or tackles are valued by their effect on the probability of scoring. Computer vision supports player tracking, pose estimation, and injury prediction from gait and fatigue analysis. Game theory illuminates strategic interactions in penalty kicks and compares player decisions with mathematically principled baselines.

The authors' contribution is to name the intersections. Frontier 1 combines game theory with statistical learning; Frontier 2 combines statistical learning with computer vision; Frontier 3 combines game theory with computer vision. The paper states that most state-of-the-art football analytics falls under a single area, with some initial activity at Frontier 2, and that it identified no notable research activity at Frontier 1 or Frontier 3 for football or sports generally. The microcosm is the integrated domain requiring all three fields at once.

Penalty kicks as a game with measurable payoffs

The paper's detailed case study sits at Frontier 1. It revisits Palacios-Huerta's work on penalty kicks and adds statistical learning of player attributes. The game-theoretic setup is a bimatrix game: two players, a kicker and a goalkeeper, each choosing from a set of strategies, with payoffs recorded in two matrices. A mixed strategy is a probability distribution over those choices, and a Nash equilibrium is a strategy profile where neither player can gain by unilaterally deviating.

The paper also defines an epsilon-Nash equilibrium, in which a player could gain by deviating but the gain is bounded by epsilon. These definitions matter because they set the benchmark against which real penalty-kick behavior is compared. The authors describe empirical game theory as a way to build such games from real interaction data or simulations, and they note that empirical game-theoretic analysis can be extended to meta-game models of set pieces, automatically clustering and identifying useful meta-strategies.

What the 15,000-kick dataset adds

The case study uses a dataset of over 15,000 penalty kicks, drawn from the main European leagues, and visualizes the shot distribution within it. The paper also illustrates the distinction between natural and non-natural sides, a concept used in the analysis of kick placement.

The key move is to combine the game-theoretic analysis with statistical learning of player attributes. The paper describes this as blending game-theoretic analysis with learned predictive models to provide more granular analysis tools, and it presents the combination as providing deeper insights into penalty-kick taking than the classical analysis alone. The authors also connect this to counterfactual analysis: predictive models can be used to consider outcomes of alternative scenarios that never actually took place, and trajectory prediction models enable ghosting, which predicts what a different team or player would have done in a given play.

Player vectors turn actions into comparable attributes

The statistical-learning component rests on player vectors, which follow Decroos and Davis's definition of playing style: a player's preferred areas on the field and the actions they tend to perform in each location. The construction is concrete. The authors collect event-stream data from all Premier League matches Liverpool Football Club participated in from 2017 to 2019 and filter for passes, dribbles, shots, and crosses. For each player and action type, they overlay a 60 by 40 grid on the pitch and count how often the player performed that action in each cell.

Each grid matrix is then reshaped into a vector and compressed with non-negative matrix factorization, producing a smaller vector whose dimensions quantify the player's preference for performing that action type in a given area. Each player ends up with four vectors, one per action type, concatenated into a final 18-dimensional vector. This is the learned representation that the paper combines with the game-theoretic penalty-kick analysis, and it is also the basis for related applications such as assessing a player's contribution to a team or measuring the mutual chemistry of a player pair.

Vision supplies the raw input the other fields need

Computer vision is the field that turns video into data the other areas can use. The paper describes multi-person pose estimation as a central part of vision-based football analysis and groups methods into bottom-up approaches, which first detect human joints and then group them into pose instances, and top-down approaches, which first detect body instances and then run single-person pose estimation on each. The computation cost of top-down methods grows linearly with the number of people in an image, while bottom-up cost stays constant; top-down approaches are often more accurate when instances overlap significantly.

The authors experimented with G-RMI, a top-down approach. In its first stage, Faster-RNN detects person instances; in the second, a fully convolutional network densely classifies whether each spatial position is near a given keypoint class and refines predictions with offsets, with a specialized form of Hough voting aggregating the results. The paper presents example predictions and says it plans to build on this approach for the challenges it identifies. This is a report of an experiment and a plan, not a validated end-to-end football system.

The agenda's value lies in the unclaimed intersections

The paper's most decision-relevant claim is a priority ordering rather than a performance number. It states that the most pertinent research problems lie in the underexplored intersection of statistical learning, computer vision, and game theory, and that no notable research activity was identified at Frontier 1 or Frontier 3 for football or sports generally. The penalty-kick case study is the paper's evidence that Frontier 1 can produce new insight: combining equilibrium analysis with learned player attributes goes beyond the classical game-theoretic treatment.

The authors also argue that football is a uniquely challenging AI domain because it is reasonably well controlled compared with other physical domains, involves heterogeneous human factors such as skill sets and injury risk, and has a long cause-and-effect feedback loop because scoring is relatively infrequent even in professional play. They see the rapidly emerging availability of multi-modal sensory data as making football an ideal platform for developing and evaluating AI algorithms. The paper closes by pointing to downstream extensions to other sports, both real and virtual, which remain envisioned impacts rather than demonstrated results.

Sources used in this answer