Why Some Clubs Turn Data Into Wins — and Others Don’t
What research shows about tactical modelling, automation, and the gaps between science and practice.
The following summary critically reviews the research paper titled “Perspectives on data analytics for gaining a competitive advantage in football: computational approaches to tactic” by Sigrid Olthof and Jesse Davis. All data, figures, and analysis presented here are drawn from their original work; I do not claim any authorship or ownership of the content. This summary has been written to provide a concise and technically informed synthesis of the paper’s findings, methodologies, and implications, while maintaining fidelity to the authors’ intellectual contributions.
Introduction
The authors situate football within the broader “Moneyball revolution”, describing how data-driven methods now support “team’s performance evaluation, roster construction, and tactical decisions to enhance performance” across multiple sports. They argue that football has followed a similar trajectory: video analysis is now standard, but clubs increasingly seek extra competitive advantage through computational tools such as expected goals models, tactical metrics, and recruitment analytics.
Expected goals (xG) is presented as the canonical example: a “well-known” metric that has “added value to operational decision-making in the transfers and recruitment of players” while also permeating post-match media discourse as a performance descriptor. At the same time, the commentary stresses that xG is only one part of a much broader computational ecosystem: there are models to “quantify how players contribute to a team’s performance”, to “automate analysis of individual player tendencies”, and to “model the movement patterns of teams as well as players,” many of which have already been implemented by clubs, federations, and data providers.
The central problem the article addresses is a gap between communities. Computational work on football tactics largely lives in AI, data mining and operations research venues, while sport scientists work closer to practice with “evidence-based” and often experimental designs. As the authors put it, “knowledge and skills on computational research mainly reside within the computational science community,” and collaboration with sport science “is relatively limited”. They contend that because these communities employ different tools and perspectives, tighter integration would “yield a strong multi-disciplinary research approach to advance football analytics”.
The paper sets three aims.
To “overview some of the historically prominent lines of computational research” in football, with an emphasis on machine-learning based solutions to everyday football questions.
To highlight “emerging trends” that either strongly benefit from or outright require domain expertise, such as automating parts of match analysis and evaluating decision-making.
To provide guidance and concrete initiatives that can “strengthen the ties” between sport scientists and computational researchers and help their work translate into competitive advantage for teams.
Computational methods in football
This section characterises modern football analytics as driven by large-scale datasets: many teams, often multiple leagues, and in some cases longitudinal data spanning several seasons. Two primary data sources underpin this work. Event data record on-ball actions (passes, shots, cards, substitutions) with attributes such as location, timestamps, and involved players. These are usually manually annotated and only describe “one of the 22 players on the pitch” per event. Crucially, there is now a sizeable public corpus of around “4,000 games” of open event data that can be used for non-commercial research.
Tracking data are derived from optical tracking or wearable systems and provide the spatiotemporal coordinates of all players and the ball several times per second. While optical tracking can be extracted from in-stadium systems or broadcast footage, there is “little publicly available tracking data”, meaning most work in this space depends on proprietary datasets.
Given this data landscape, computational football research tends to cluster around three themes:
Identifying and evaluating tactics
Designing predictive performance indicators (such as xG or possession-value models)
Modelling player movement.
The authors emphasise that much of this research is published in computer science venues and framed as machine learning problems, which can make it hard for sport scientists to discover. To address this, they focus on “the problems that have been tackled” rather than on detailed algorithmic choices, while still grounding everything in ML-style formulations such as prediction, action valuation, and pattern discovery.
Evaluating tactics
The first line of work covers how teams use data to analyse and evaluate tactical behaviour across in-possession, out-of-possession, transitions, and set-pieces. Clubs invest in analytics departments, “data science skills and computational techniques” with the explicit aim of gaining competitive advantage within their game model.
A key tactical area is pressing and counterpressing. The authors describe rule-based and machine-learning approaches that detect counterpressing situations by combining event and tracking data, then evaluate them via outcomes such as speed and success of ball regain or whether the sequence leads to a shot. One example uses a risk–reward framing: pressing is risky because leaving the team’s shape “opens up space if the press is broken”, but it carries the reward of quickly regaining possession in advanced positions. Such models help teams that “have not been effective in pressing” to diagnose and refine their approach.
Set pieces, especially corners, are another rich area. Studies combine event and tracking data to compare “in-swinging vs. out-swinging corner kicks and zonal vs. man marking schemes” and to characterise movement patterns and marking roles. Computational models can ask very concrete questions: which routines yield more shots; how different structures perform against different defensive setups; and how adjusting marking schemes might reduce the chance of conceding.
Beyond pressing and set pieces, the authors highlight computational approaches to more specific tactical questions: “optimal timing to substitute players”, the causal effect of throw-in locations, or the “benefits of crossing”. The common pattern is that these methods formalise a tactical decision (when to substitute, where to restart, whether to cross) and then estimate expected outcomes under different choices, providing a systematic complement to coaching intuition.
Measuring player and team success
This section focuses on predictive, model-based metrics like xG and expected possession value that quantify performance beyond raw goals or simple counts. xG models take features of a shot (shooter location, assist context, goalkeeper position) and return “an estimated probability of the shot being converted into a goal”. Empirically, xG is “more predictive of future success than looking at goals scored”, and has been adopted to support recruitment, transfers, and team evaluation.
The authors note that while it is often claimed that xG is more stable than goals for individual players, “to our knowledge, this has also been explored in blogs and not in scientific publications”. They stress that the lack of peer-reviewed proof of superior stability “does not diminish its value as a useful indicator” but that this is an area where more rigorous evidence would be welcome.
Expected possession value (EPV) models generalise beyond shots to value all on-ball actions. The central idea is that good actions increase your team’s chance of scoring and/or decrease the chance of conceding in a near-term time window, such as “the next 10 actions or 10 seconds”. Early frameworks like expected threat (xT) valued locations solely based on their contribution to scoring probability; later work incorporates both offensive reward and turnover risk, so an action is assessed by how it changes both scoring and conceding probabilities.
A major technical axis is how the “game situation” is represented. Event-data models may use only the current ball location, or richer sequences including previous actions, ball progression speed, and contextual features such as time remaining and scoreline. Tracking-based models use spatiotemporal features of all players, the ball, and the goals, and can be visualised as continuous value maps over the pitch. Commercial providers now maintain their own proprietary EPV-style models.
The authors also discuss work on passing behaviour as a partial step away from EPV. Metrics quantify risk-taking (attempting low-probability passes), decision-making (what might have happened under alternative pass choices), creativity (how unusual a player’s pass selection is relative to typical options), and off-ball availability (how well players position themselves to become viable passing options). These studies share the same core idea: treat each pass, or possible pass, as a decision under uncertainty and evaluate its contribution to team success.
Modeling player movement
The third historical line concerns models that exploit full tracking data to describe and predict player movement, broken into pitch control and ghosting.
Pitch control models estimate, for any location on the pitch, the probability that a given team will reach that point before the opposition. Building on the foundational work of Taki and co-authors, modern versions compute control surfaces based on players’ positions, velocities, and accelerations over recent frames. From the perspective of the team in possession, areas they control are “considered safe options to pass the ball into”. Some models then weight these surfaces by field value, reflecting that certain zones are inherently more dangerous. The authors point out that current pitch-control models typically do not incorporate “individual characteristics” like differences in maximum speed, but their advantage lies in being “extremely easy to visualize and interpret”.
Ghosting refers to models that learn to generate future trajectories given recent movements. Here, sequences of positions over the last few seconds are used to train models that predict where players would move under a particular tactical scheme. This allows analysts to compare observed movements to a model of “the team’s planned strategy as designed by the coaching staff”, for example when defending a counterattack. Although ghosting is more mature in basketball, early football work shows how it can highlight discrepancies between intended and actual execution, and how these deviations affect chances of conceding.
Emerging topics
The article then turns from historical strands to newer directions that are driven more explicitly by football-specific questions and practical needs. These emerging topics aim not only to improve predictive performance but also to “support practitioners” by saving time, improving interaction with data, and providing richer context around model outputs.
Automating analysis
Match and video analysts still spend much of their time “watching and annotating video footage” with coding software, a process that is “tedious and time-consuming” and prone to human perceptual error. The authors argue that computational methods can automate parts of this workflow, especially when tracking data is available, thereby enabling quicker and potentially more accurate analysis.
Formations are one example: early work automatically identified team formations from tracking data, while more recent studies examine how formations evolve over time and across phases of play (build-up, attacking, etc.), with phases themselves being “automatically detected by a learned model”.
Set-piece analysis has seen progress in automated pattern detection. For corners, one line of work clusters recurrent attacking movement patterns by representing players’ trajectories via their start and end positions. Another line automatically classifies the defensive role of each player (distinguishing man-markers, zonal defenders, or specific back-post roles) based on their movements, with the taxonomy of roles derived from expert coaches. These tools help coaches to understand typical routines, choose effective variations, and adjust defensive setups to minimise the probability of conceding.
The same logic extends to open play. Instead of manually searching for overlapping runs or specific rotations, models can be trained to detect patterns from a small set of labelled examples provided by a domain expert. The process is iterative and interactive: analysts supply clips of interest; a model is trained; and then “new labels” are acquired in a targeted way as the system presents borderline or uncertain cases for expert feedback.
From a competitive standpoint, the authors argue that such tools allow teams to cover more matches and more situations within fixed time constraints. However, they insist that “training such models requires input from sport scientists”, both in defining the behaviours of interest and in providing feedback to refine the models. Automation should free analysts to spend more time on quality control and on translating findings into coach- and player-facing recommendations.
Evaluating decision-making
Here the focus shifts from valuing actions in isolation to assessing the quality of players’ decisions given the options available. Existing EPV-style models can say whether an action increased or decreased the team’s expected outcomes, but not whether it was the best choice in a constrained, difficult situation. As the authors put it, “the chosen action may have been the best solution out of a range of bad options”, so simply penalising a drop in EPV can be misleading.
Emerging approaches therefore model the full “affordance set” of feasible actions in a game state, allowing analysts to compare the selected action to plausible alternatives. This leads to questions about how to reward or penalise outcomes: should a player be credited for a good choice poorly executed, or criticised for a bad choice that happens to work out? Reinforcement-learning based work extends this by simulating what would happen if players “systematically performed different actions in certain situations”, such as shooting more or less from particular regions, and estimating the effect on expected goals over time.
The researchers also note early work on off-the-ball decision-making. One example model evaluates how off-ball runs contribute to receiving passes, entering dangerous regions, and increasing pitch control. In a case study with a first-division club, coaches used such models to present players with game situations where both pass choices and off-ball positioning were evaluated. The authors emphasise that while this was an “uncontrolled intervention”, it illustrates how computational models can frame “discussion with coaches and players about on-pitch decision-making and tactical adjustments”; provided the outputs are communicated in a way that is understandable and credible.
Generative AI
Generative AI (GenAI) is introduced as a newer but fast-growing area in football analytics. One prominent example is TacticAI, a collaboration between Google DeepMind and Liverpool FC that trained a model to generate plausible corner-kick routines from tracking data. Strikingly, the generated patterns were so realistic that “human analysts could not distinguish” them from real corners, raising the prospect of using GenAI as a creative assistant for set-piece design.
Another use of GenAI is to impute tracking data that are missing from broadcast footage. Because standard broadcast only captures a subset of players at any one time, generative models can be trained to predict off-screen positions, effectively reconstructing full-pitch trajectories from partial observations. This could significantly expand the availability of tracking-grade data from TV feeds.
The authors also mention, more speculatively, commercial products that use large language models to synthesise scouting reports, combining qualitative-style text with quantitative metrics. They note that, so far, “we are not aware of any peer-reviewed research on this topic,” signalling that GenAI for textual football analysis is ahead in practice compared to academic scrutiny.
Trustworthy AI
The final emerging topic is the broader notion of trustworthy AI, which the authors frame as essential once models leave the lab and impact real-world decisions. They argue that “just optimizing predictive performance is insufficient to justify deployment,” especially in settings governed by legal frameworks such as GDPR and in environments where athletes’ rights and welfare are at stake.
They stress that models only provide competitive advantage if their outputs are “meaningful, actionable, and interpretable” to coaching staff. Collecting more data and training more complex models raises both ethical concerns over data use and practical concerns over whether staff will trust, understand, and act on the results. The remainder of the section is organised into three dimensions of trustworthiness: interpretability, fairness, and transparency.
Interpretability
Interpretability is defined as “the ability to understand why models make certain predictions,” and the authors highlight both design-time and post-hoc strategies. A priori, models can be made more interpretable by using features that align with football domain knowledge (for example, established tactical metrics or intuitive pitch zones) and by choosing model classes that are inherently more transparent, such as simple trees or linear models. This helps analysts and coaches see how inputs relate to outputs and builds trust in the modelling process.
Post-hoc techniques explain or approximate black-box models after training. One popular approach is SHAP, which uses game-theoretic principles to assign each feature a contribution to a specific prediction. This has been used for xG, allowing experts to see which factors drove a particular shooting probability. Other methods distil complex models into simpler surrogates, either globally (to “extract knowledge” about the model’s general behaviour) or locally (to explain individual predictions). In possession-value models, such distillation has been used to understand which aspects of a situation most strongly influence action values. The authors frame these tools as essential for checking whether models “agree with domain knowledge” and for making them acceptable in practice.
Fairness
Fairness focuses on systematic bias or discrimination across groups defined by sensitive attributes such as gender or ethnicity. The authors note evidence that player evaluations can be influenced by race and gender, meaning that if models are trained on these human annotations they may “perpetuate this bias”.
They give the specific example of using xG-based “goals above expectation” metrics to measure finishing skill: there is recent evidence that this measure is biased against good finishers by underestimating their true ability. As they put it, “such biases can undermine the validity of the metrics and hence call into question their suitable for use in practice”. This underscores the need to interrogate models not only for accuracy but also for whether they encode and amplify pre-existing biases in football data and evaluation processes.
Transparency
Transparency is treated as an open-ended but crucial concept, involving both data and model documentation. For data, the authors reference ideas like “datasheets for datasets” that describe how and why data were collected, what populations they cover, and what limitations they have. For models, “model cards” and similar documentation can clarify design choices, intended use cases, and known failure modes.
In football, this matters because design decisions (which leagues to include, what features to derive, how to label events) can significantly affect performance and generalisability. The authors raise the question of whether models trained on men’s top leagues (Premier League, La Liga) are valid when applied to women’s football or lower divisions, noting empirical work that suggests xG models may not transfer straightforwardly across these contexts. Transparency about training data and scope is therefore critical for avoiding misapplication.
Advancing collaboration between sport and computational scientists
Having laid out the technical landscape, the authors turn explicitly to collaboration. They argue that there is “tremendous potential” in bringing sport science and computational science closer together, especially as models are moving “closer to practice”. Sport scientists have rich, real-world questions grounded in coaching, medicine, and performance support, while computational researchers can provide methodological advances and tools, but also face the risk that powerful ML frameworks will be misused without sufficient domain understanding.
Collaboration is therefore framed as a way to ensure that models are both technically sound and contextually relevant, and that their outputs are interpreted correctly. The authors then distil lessons learned from past joint projects and outline structural initiatives that can facilitate this integration.
Anatomy of a collaboration
The authors identify several practical factors that shape successful collaborations.
First, work should start from “a good question that addresses a relevant problem” for staff in coaching, medical, or support roles. They warn that vague requests such as “Can AI solve my problem?” often lead to results that are “not interesting, practical, or useful” in context. A better starting point is a well-specified question that can be naturally mapped to computational tasks like prediction, decision-making, pattern discovery or scheduling.
Second, access to suitable data is indispensable, and both quantity and quality matter. Computational researchers are generally accustomed to working with “existing and relatively clean data,” and may be less familiar with collection protocols and practical issues such as missingness or measurement noise. Sport scientists can contribute by clearly describing data provenance, quality issues, and domain-specific preprocessing needs, as well as constraints on data sharing.
Third, mutual understanding is essential. The commentary stresses that computational researchers should invest time in learning enough football and sport science to grasp what matters tactically and physically, while sport scientists benefit from learning key ML terminology and concepts. The same idea might be called a “key performance indicator” in one community and a “feature” in another; explicit bridging of vocabularies helps avoid miscommunication.
Fourth, publication strategies need to be discussed upfront, because the two communities differ in culture. Computer science places heavy weight on peer-reviewed conference papers emphasising methodological novelty, while sport science values journal articles that combine theory with applied evidence and practical implications. Applying existing algorithms to a sports question is often “out-of-scope” for CS venues, whereas highly technical methodological contributions may not be relevant for most sports journals. The authors propose that “a collaborative effort could logically result in two publications”: one aimed at computational audiences focusing on the technical advances, and another aimed at sports science, contextualising the work, detailing applied analyses, and highlighting implications for practice.
Moving forward
To structurally foster collaboration, the authors highlight two kinds of initiatives: joint events and open science. Joint events (seminars, workshops, and multidisciplinary conference sessions) “literally” put sport and computational scientists in the same room, creating opportunities to meet, share problems, and spark joint projects. They cite Dagstuhl seminars on sports strategy and tactics and special sessions at the World Congress on Science and Football as examples. These gatherings can also provide platforms to disseminate work that might otherwise struggle to fit neatly into a single disciplinary venue, although they often remain small and infrequent.
Open science, especially shared data and open-source software, is presented as a key accelerator. The authors point to existing public datasets and high-level packages for xG, possession-value models, and general tracking-data analytics (such as Floodlight and Kloppy), as well as initiatives like PySport that curate sport-science tools. Such resources “reduce the barrier to entry” and make it easier to replicate and extend analyses, both in academia and in practice.
At the same time, they caution that open tools can be misused if applied outside their valid scope. They argue that researchers releasing software should provide “clear instructions and guidelines about how to use them and when they are (not) applicable,” ideally co-written with sports experts so that documentation is accessible and sport-specific edge cases are considered. Because best practices and methods in AI evolve quickly, continuous collaboration is needed “to promote the appropriate and up-to-date use of computational methods in sport science research”.
Conclusion
The paper closes by reiterating its three aims and its central thesis. First, by collecting “historically important topics” in one place, the authors hope to create an accessible entry point for sport scientists into the computational football literature. Second, they have highlighted emerging areas (automation, decision-making evaluation, generative and trustworthy AI) where domain expertise is especially critical and where closer integration of communities can yield substantial competitive advantage. Third, they have spelled out practical guidance and structural initiatives for collaboration, emphasising “continuous communication, willingness to understand each other’s domain knowledge, and joint events and open science” as key ingredients.
Overall, the article frames football analytics not just as a collection of models and metrics, but as a multidisciplinary endeavour in which computational and sport scientists jointly shape tools that clubs can use to improve tactics, recruitment, and performance in a way that is rigorous, interpretable, and practically useful.
Learn More
Here you have a direct link to get the free xG Mini-Guide, suited for those who want to fully understand the model in 5 minutes.
References
Olthof, S., & Davis, J. (2025). Perspectives on data analytics for gaining a competitive advantage in football: computational approaches to tactics. Science and Medicine in Football, 1-13. https://doi.org/10.1080/24733938.2025.2533784
Exactly. Data accumulation is no longer the bottleneck — prioritisation and sense-making are. This is the focus of our work at FAIR Research Organization: structuring complex datasets into interpretable, decision-oriented indicators rather than multiplying raw metrics.
Absolutely brilliant synthesis of where computational football analytics is heading. The point about affordance sets for evaluating decisions is fascinating because most EPV framworks penalize players for drops in expected value without considering the qualty of alternatives available. This feels especially relevant when assessing defenders under high pressure who might make the least-bad pass rather than an optimal one. The gap between CS venues and sport science journals probably explains why so many clubs still dunno how to properly deploy these models in practice.