Modeling Off-Ball Contributions in Team Sports

42 concepts40 papers1 proposal

Imagine analyzing a game of chess by only recording which piece captures another, completely ignoring where the remaining pieces are standing. This is the fundamental limitation of traditional sports analytics. For decades, data collection has relied on "event logs"—discrete records of actions involving the ball or puck, such as passes, shots, and fouls. As Luca Pappalardo et al. demonstrate in their release of a massive open-access soccer-log dataset, A public data set of spatio-temporal match events in soccer competitions, these systems focus exclusively on the player in possession.

This focus creates a massive information gap. In a 90-minute soccer match, an individual player holds the ball for an average of just under two minutes. The remaining 88 minutes are spent moving off the ball—positioning defensively, making decoy runs, or pressing opponents. By omitting these trajectories, event-based models remain blind to the vast majority of a player's tactical influence. Paolo Cintia et al. note in PlayeRank: Data-driven Performance Evaluation and Player Ranking in Soccer via a Machine Learning Approach that this omission severely limits the completeness of player assessments. Similarly, Yuesen Li et al. argue in Characterizing player’s playing styles based on player vectors for each playing position in the Chinese Football Super League that relying solely on on-the-ball events ignores critical contributions like spatial marking and decoy runs that never result in direct ball contact.

The Cost of the Blind Spot

When models lack full-pitch context, their downstream predictions and valuations become heavily biased. This blind spot manifests in two primary ways:

1. The Valuation of Defensive Inaction

In team sports, a defender's ultimate goal is often to ensure that nothing happens. If a defender positions themselves perfectly to cut off a passing lane, the opposing midfielder will choose a different route, and the defender's brilliant positioning goes completely unrecorded in an event log. Charbel Merhej et al. address this in What Happened Next? Using Deep Learning to Value Defensive Actions in Football Event-Data, explaining that event-based models cannot account for the defensive value of marking or positioning.

This creates a structural bias toward active, high-risk interventions (like tackles) while undervaluing players who control space. As Guiliang Liu et al. show in Deep Reinforcement Learning in Ice Hockey for Context-Aware Player Evaluation, event-based valuation models systematically undervalue players who preemptively disrupt plays through spatial control. However, as Oliver Schulte notes in Valuing Actions and Ranking Hockey Players With Machine Learning, integrating dense tracking data with reinforcement learning remains a significant challenge due to the limited public availability of such granular datasets.

2. Decontextualized Probability Models

Consider an Expected Goals (xG) model designed to calculate the probability of a shot scoring. An event log tells the model the coordinates of the shot, but it cannot tell the model if there is an open net or a wall of five defenders blocking the path. H. Eggels et al. point out in Explaining Soccer Match Outcomes with Goal Scoring Opportunities Predictive Analytics that this lack of defensive pressure data severely limits model accuracy. Mustafa Cavus et al. echo this in Explainable expected goal models for performance analysis in football analytics, noting that without integrating dynamic tracking data, models cannot account for the spatial distribution of opponents and the goalkeeper at the moment of the event.

This limitation is further compounded in defensive modeling; Lingfeng Ruan et al. demonstrate in Quantifying the Effectiveness of Defensive Playing Styles in the Chinese Football Super League that event-based data fails to capture defensive shapes and physical metrics like sprinting distances, preventing a comprehensive understanding of how off-ball movements influence game outcomes.

The scarcity of spatial tracking data also hinders player classification. Xing Wang et al. point out in The Differences in the Performance Profiles Between Native and Foreign Players in the Chinese Basketball Association that without spatial metrics like shot charts and shooting locations, it is difficult to distinguish between specific player archetypes and roles. Even when analyzing sequences of play, the lack of spatial context degrades tactical relevance. As Rory Bunker et al. argue in Supervised sequential pattern mining of event sequences in sport to identify important patterns of play: An application to rugby union, traditional sequential pattern mining treats events as equivalent regardless of where they occur on the field, ignoring the vastly different strategic implications of identical event sequences in different zones. Similarly, Zachery Born et al. note in Clustering Offensive Strategies in Australian-Rules Football Using Social Network Analysis that restricting tactical analysis to ball movement patterns fails to account for the complex interactions between offensive and defensive structures.

From Discrete Events to Continuous Trajectories

To resolve these limitations, researchers are shifting from discrete event streams to continuous spatiotemporal tracking data. Instead of treating a match as a sequence of isolated actions, modern frameworks model the game as a continuous, multi-agent interaction over a shared space.

This shift allows models to capture "micro-actions"—the subtle, off-ball movements that orchestrate scoring opportunities. In basketball, for example, Anthony Sicilia et al. developed DeepHoops: Evaluating Micro-Actions in Basketball Using Deep Feature Representations of Spatio-Temporal Data, leveraging raw tracking data to quantify the hidden value of off-ball screens and backdoor cuts. Similarly, in multi-agent systems, Motokazu Hojo et al. demonstrate in Automatically recognizing strategic cooperative behaviors in various situations of a team sport how machine learning can segment and classify cooperative behaviors that occur far from the ball, removing the object-centric bias that plagues traditional analysis.

By tracking every player simultaneously, models can construct dynamic representations of territory, often called "control areas." For instance, Ning Ding 0002 et al. utilize drone footage in Estimation of control area in badminton doubles with pose information from top and back view drone videos to generate probability maps of the court. This approach captures the continuous spatial interaction between teammates, moving beyond discrete shot counts to quantify how effectively a unit covers the playing area.

Quantifying the Unseen Strategic Layer

When full-field tracking data is integrated, the analytical focus expands from what happened to what could have happened. This enables counterfactual reasoning and deeper strategic modeling. Luc De Raedt et al. demonstrate this in A Markov Framework for Learning and Reasoning About Strategies in Professional Soccer, using a Markov framework to analyze passing lanes and the strategic exploitation of unoccupied space. By evaluating how movements alter the probability of future success, frameworks like the one proposed by Tom Decroos et al. in STARSS: A Spatio-Temporal Action Rating System for Soccer can account for space creation and positioning that does not involve direct ball contact.

Ultimately, this research direction aims to bridge the gap between what data models see and what human experts instinctively perceive. As Sho Takahashi et al. note in Visualization Methods for Encourage Experience of Sports in Soccer Videos, casual viewers focus on the ball, while experts track the underlying spatial structure. By incorporating advanced inputs—such as estimating where players are looking—Genki Suzuki et al. show in A Method for Player Importance Prediction from Player Network Using Gaze Position Estimated by LSTM that models can combine spatial physics with network analysis to identify key players whose contributions are entirely off-the-ball.

As surveyed by Artur Xarles et al. in Action Valuation in Sports: A Survey, capturing these simultaneous, off-ball actions remains the critical frontier for fully representing the strategic complexity of team sports.

Key Foundational Work

Know Thine Enemy: A Champion RoboCup Coach Agent (2006) — contributed 1 concept elements to this direction.
Automatically Discovering Offensive Patterns in Soccer Match Data (2015) — contributed 1 concept elements to this direction.
History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 35 papers
  2. Version 1 · 2026-10-04 · 54 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_event_based_data_limitations_1b8eca77,
  title = {event-based data limitations},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/event-based-data-limitations/rdr_1b8eca776036461fa3b05920d79f755f},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 2, frozen text):

@misc{lacuna_event_based_data_limitations_1b8eca77_v2,
  title = {event-based data limitations},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/event-based-data-limitations/rdr_1b8eca776036461fa3b05920d79f755f/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake