The Future of Football Analytics — Beyond GPS to Spatial Intelligence

Introduction

The GPS-and-events era of football analytics — counting distances, speeds, and pass completions — is giving way to a more sophisticated paradigm: spatial intelligence, where every moment of a match is described not just by what actions occurred but by what the game state was, what actions were available, and what value was created or destroyed. This transition, from outcome measurement to state-action valuation, represents a fundamental shift in how performance is understood, and it is being driven by deep learning applied to spatiotemporal tracking data.

The Science

The limitations of current event-based analytics: Traditional football statistics measure outcomes: goals, assists, shots on target, pass completion percentage. These are noisy signals — heavily influenced by teammate quality, opponent quality, and chance. A central midfielder making 10 high-value passes into dangerous spaces might complete only 40% of them and appear statistically worse than a player making 40 short back-passes at 95% completion. Event-based statistics systematically undervalue risk-taking and space-creation.

The shift to state-action valuation: Models like VAEP (Valuing Actions by Estimating Probabilities) and Expected Possession Value (EPV) assign each game state a probability of eventual goal-scoring. Every action — whether successful or not — changes the game state. The value of an action is the change it creates in goal probability. This framework:

• Values every action including defensive clearances, interceptions, and ball recoveries
• Accounts for the risk-reward of attempted actions, not just completed ones
• Enables comparison of players in completely different tactical contexts

Deep learning in football analytics: Spatiotemporal tracking data — 22 players × 25 Hz position — cannot be adequately modelled by traditional statistical methods. Deep learning architectures specifically designed for spatiotemporal data include:

• Graph Neural Networks (GNNs): Model player interactions as dynamic graphs — each player is a node, movements create edges — enabling modelling of team structure and space creation
• Transformers applied to player sequences: Attention mechanisms that capture long-range temporal dependencies in player movement — identifying tactical patterns across multiple possessions
• Generative models (VAE, GAN): Generate synthetic match scenarios from historical data — enabling tactical simulation without needing to observe every possible scenario

Counterfactual analysis: One of the most powerful analytic frontiers — asking “what would have happened if the player had made a different decision?” Counterfactual models generate alternative scenarios from tracking data, allowing comparison of the actual action taken versus the theoretically optimal action. This enables coaching conversations grounded in data: “In this possession, you had three passing options — here’s the EPV of each option, and here’s why the ball was held rather than played.”

Player profile similarity and recruitment: Deep learning on tracking data generates player movement signatures that enable similarity search — identifying players from different leagues with similar spatial movement profiles to a transfer target. This is more predictive of adaptation to a new tactical system than traditional statistics-based similarity.

What Research Says

Fernández et al. (2019) presented the foundational EPV framework at the MIT Sloan Sports Analytics Conference, demonstrating that deep learning models applied to player tracking data could estimate continuous game state value — enabling the first practical implementation of comprehensive action valuation.

Bauer and Anzer (2021) developed a data-driven method for detecting counterpressing in Data Mining and Knowledge Discovery, demonstrating that tracking data can be used to identify counterpressing situations and measure their success — enabling real-time game state characterisation.

Pappalardo et al. (2019) published an open data set of spatio-temporal match events in soccer competitions in Scientific Data, providing the open resource on which spatiotemporal deep learning models are built, the frontier where the greatest near-term advances are expected.

Did You Know? Eleven Sports, StatsBomb, and several club analytics teams are working on implementations of “tactical language models” — trained on millions of possession sequences — that can describe tactical patterns in natural language. Rather than requiring an analyst to identify pressing triggers, these models can report “this team typically presses when their No. 10 drops into midfield and the opposition goalkeeper is on the ball” — extracting insights from pattern recognition that would take human analysts hundreds of hours to identify. The convergence of NLP and football tracking analytics is moving from research concept to operational tool at the leading edge of the industry.

Applied to Football

Applying next-generation analytics:

1. Transition from event-based to state-action metrics for player evaluation. VAEP and EPV are increasingly available through commercial analytics platforms — incorporate these into player evaluation alongside traditional metrics.
2. Counterfactual tactical analysis for coaching conversations. Data-supported coaching sessions using counterfactual scenarios (“here were your three options in this possession”) are more actionable and persuasive than descriptive event analysis alone.
3. Graph neural network models for team structure analysis. Player movement network analysis reveals how team structure creates or denies space — enabling systematic comparison of team organisation across opponents.
4. Player similarity search for recruitment. Deep learning similarity models on tracking data enable more accurate positional and tactical fit assessment for potential signings than event statistics alone.
5. Invest in tracking data infrastructure. The analytical gap between clubs with and without full-match tracking data will widen as deep learning approaches mature. Tracking data is the foundational resource that enables all advanced analytics.

Key Takeaways

• Event-based statistics are noisy outcomes measures; state-action valuation models provide a richer, more predictive framework
• VAEP and EPV quantify the value of every action — including defensive actions invisible in traditional statistics
• Deep learning architectures (GNNs, transformers) enable modelling of spatiotemporal football data at a scale and complexity traditional statistics cannot
• Counterfactual analysis enables evidence-based coaching conversations about decision quality, not just decision outcomes
• Tracking data infrastructure is the foundational investment enabling advanced football analytics

References

• Fernández, J., Bornn, L., & Cervone, D. (2019). Decomposing the immeasurable sport: a deep learning expected possession value framework for soccer. MIT Sloan Sports Analytics Conference 2019.
• Bauer, P., & Anzer, G. (2021). Data-driven detection of counterpressing in professional football. Data Mining and Knowledge Discovery, 35(5), 2009–2049.
• Pappalardo, L., Cintia, P., Rossi, A., Massucco, E., Ferragina, P., Pedreschi, D., & Giannotti, F. (2019). A public data set of spatio-temporal match events in soccer competitions. Scientific Data, 6(1), 236.

Written by Hüseyin Akbulut, MSc Sport Scientist · Founder of Sporeus

Hüseyin Akbulut is the founder of Sporeus and author of THRESHOLD (EŞİK), a 540-page Turkish-language book on endurance science.

• 540pp THRESHOLD Book
• MSc Sport Sciences
• Marmara University

Full profile →