Pitch Control Models

Definition

Pitch control is a continuous spatial model that estimates, for every point on the pitch at a given moment, the probability that a given team could control the ball there. Unlike Voronoi diagrams (see Voronoi Diagrams) which use only static positions, pitch control accounts for player velocities, acceleration capabilities, and physical movement models to produce a probabilistic surface.

History & Origins

William Spearman and the Physics-Based Approach

The concept was formalized for football by William Spearman in his 2018 paper "Beyond Expected Goals", presented at the MIT Sloan Sports Analytics Conference. Spearman, then working at Hudl Statsbomb (later Hudl / Second Spectrum), proposed a physics-based model where each player's influence on a pitch location is determined by their ability to reach that location, modeled as a function of:

• Current position
• Current velocity
• Maximum acceleration and top speed
• Time to react and change direction

The result is a smooth probability surface (0 to 1) for each team at each frame, where 0.5 represents a contested zone and values closer to 0 or 1 represent clear control by one team.

Expected Possession Value (EPV)

In 2019, Javier Fernández and Luke Bornn published "Wide Open Spaces: A statistical technique for measuring space creation in professional soccer" and the related Expected Possession Value (EPV) framework. EPV combines pitch control with a ball value model to produce a real-time value surface of the entire pitch — essentially answering "what is the expected outcome of possessing the ball at this exact location, given the current positions and movements of all players?"

Fernández developed this work while at FC Barcelona's Innovation Hub, and Bornn brought experience from the NBA (he had applied similar spatial models at the Sacramento Kings and in his academic work at Simon Fraser University). Their collaboration represented a convergence of basketball and football spatial analytics.

Friends of Tracking

The Friends of Tracking YouTube series, initiated by David Sumpter (Uppsala University) and featuring Laurie Shaw, William Spearman, and others, made pitch control models accessible to a broader audience starting in 2020. Laurie Shaw published an open-source Python implementation of Spearman's model, which became a widely-used reference implementation for researchers and hobbyists.

This was a pivotal moment: pitch control went from being a concept discussed in proprietary settings to something anyone with Python skills and tracking data could experiment with.

How It Works

Player Influence Model

For each player at each frame:

1. Model the time to reach any point on the pitch, given current position, velocity, and physical parameters (max speed ~13 m/s, reaction time ~0.7s, max acceleration varies)
2. Convert time-to-reach into an influence probability: the faster a player can reach a point, the higher their influence there
3. A player running toward a zone has much more influence on that zone than a stationary player at the same distance

Team Control Surface

1. For each point on the pitch, sum the influence of all 11 players from each team
2. The ratio of Team A's total influence to the combined influence of both teams gives the pitch control probability for Team A at that point
3. Repeat for a grid of points across the pitch (e.g., every 0.5m) to produce the full surface

Integration with Ball Value

In the EPV framework:

1. Compute pitch control at every point
2. Overlay a ball value model (what's the expected goal probability from possessing the ball at each location — similar to xT but continuous and context-aware)
3. The product gives EPV: the value-weighted control surface
4. This can be used to evaluate passes ("what's the EPV gain of passing to that open zone?") or off-ball movement ("which player's run created the most EPV?")

What It Reveals

• Space creation: quantifies exactly how much usable space a player's movement creates, even if they never receive the ball
• Pressing effectiveness: shows how a team's press reduces the opponent's controlled space
• Pass difficulty: the pitch control at the destination of a pass indicates how contested the receiving zone is
• Defensive vulnerabilities: areas where pitch control dips below 0.5 for the defending team reveal exposed zones
• Off-ball value: one of the few frameworks that can quantify the contribution of players who don't touch the ball

Key Properties

• Probabilistic (0–1) rather than binary (Voronoi's "closest player wins")
• Dynamic: captures the effect of movement and velocity, not just position
• Continuous: no grid discretization needed (though computed on a grid for practicality)
• Requires tracking data — cannot be derived from event data

Limitations & Debates

Data Availability

Pitch control requires high-frequency tracking data (22 players + ball, typically 25fps). This data is expensive and proprietary for most competitions. Public experimentation is limited to the few available open datasets (Metrica Sports, synthetic data).

Physical Model Assumptions

The time-to-reach model relies on assumptions about player acceleration, top speed, and reaction time. These vary significantly between players (a center-back and a winger have very different movement profiles), and most implementations use uniform parameters. More sophisticated versions could incorporate player-specific physical profiles, but this adds complexity and data requirements.

Computational Cost

Computing pitch control for every frame of a 90-minute match at 25fps across a fine spatial grid is computationally intensive. Real-time applications require optimization (spatial subsampling, approximations, GPU computation).

Ball Height

Most pitch control models operate in 2D (the pitch surface). They don't account for aerial balls, where a tall player has an advantage regardless of horizontal distance. Some extensions address this, but it remains an active area.

Validation

It's difficult to validate pitch control models against ground truth, because "control" is a continuous concept without clear binary outcomes in most situations. Researchers typically validate indirectly (e.g., does pitch control at pass destinations predict pass success?).

Relationship to Other Concepts

• Voronoi diagrams → the simplest possible pitch control model (binary, position-only)
• xT (see xT - Expected Threat) → static zone-based value, no player context
• EPV → pitch control × ball value model = the most complete spatial value framework
• VAEP (see VAEP - Valuing Actions) → action-level valuation using ML, complementary but different approach

Key People

• William Spearman — formalized pitch control for football (2018), Hudl/Second Spectrum
• Javier Fernández — EPV framework, FC Barcelona Innovation Hub
• Luke Bornn — EPV co-author, spatial statistics, Simon Fraser University / Sacramento Kings / AS Roma
• Laurie Shaw — open-source pitch control implementation, Friends of Tracking
• David Sumpter — Friends of Tracking series, Soccermatics

Notable Implementations & Resources

• Spearman, "Beyond Expected Goals" (MIT Sloan 2018)
• Fernández & Bornn, "Wide Open Spaces" (MIT Sloan 2018)
• Fernández et al., EPV framework papers (2019)
• Laurie Shaw's open-source implementation — Python, widely referenced
• Friends of Tracking YouTube series — tutorial episodes on pitch control
• Metrica Sports open tracking dataset — free data for experimentation

---

Tags: #football #analytics #visualization #pitch-control #spatial-analysis #EPV