EPV — Expected Possession Value

Definition

Expected Possession Value (EPV) is a continuous, real-time framework that assigns a value to every point on the pitch at every moment of a match, representing the probability that the current possession will end in a goal, given the positions and movements of all 22 players and the ball.

EPV is the most comprehensive spatial value model in football analytics, combining pitch control (see Pitch Control Models) with ball value estimation into a single unified surface.

History & Origins

EPV was developed by Javier Fernández and Luke Bornn, with key contributions from Dan Cervone and others, published in a series of papers from 2018–2019.

Javier Fernández was working at FC Barcelona's Innovation Hub (Barça Innovation Hub), one of the first football clubs to build a dedicated analytics research department. His background was in data science and machine learning.

Luke Bornn brought a cross-sport perspective: a statistics professor at Simon Fraser University (Canada), he had previously developed spatial models for the NBA at the Sacramento Kings and later consulted for AS Roma. His work on player tracking in basketball directly influenced the football EPV framework.

Their key papers include:

• "Wide Open Spaces: A statistical technique for measuring space creation in professional soccer" (MIT Sloan 2018)
• "Decomposing the Immeasurable Sport: A deep learning expected possession value framework for soccer" (MIT Sloan 2019)

EPV built on the foundation laid by William Spearman's pitch control model (2018) but went further by combining spatial control with a learned value function, answering not just "who controls this space?" but "what is this space worth right now?"

How It Works

EPV integrates three components:

1. Pitch Control

For every point on the pitch, estimate which team is more likely to gain control of the ball there (see Pitch Control Models). This uses player positions, velocities, and physical movement models.

2. Ball Value Model

Independently estimate the expected goal probability from possessing the ball at each location. This is similar to xT (see xT - Expected Threat) but continuous rather than grid-based, and conditioned on the full game state (not just location).

3. Action Models

Estimate the probability and outcome of possible actions from the current ball location:

• Pass model: probability of completing a pass to each location, and the EPV at the destination
• Carry model: probability and value of carrying the ball forward
• Shot model: xG if a shot is taken

Combined

EPV at any point = weighted sum of the expected value of all possible actions, accounting for the probability of success for each action given the current spatial configuration.

The result is a dynamic surface that changes 25 times per second (at tracking data frame rate), showing the evolving value landscape of the pitch.

What It Reveals

• Pass value: the EPV gain/loss of every possible pass, factoring in both the value of the destination zone AND the probability of the pass succeeding given the defensive setup
• Off-ball movement value: a player's run that opens space increases EPV at nearby locations even if they don't receive the ball — one of the few frameworks that quantifies this
• Space creation: directly measures how player movement changes the value surface
• Decision-making quality: compares the action a player chose against the optimal action available (the "EPV-maximizing" option)
• Defensive contribution: shows how pressing, covering, or positioning reduces the opponent's EPV

Key Properties

• Real-time and continuous (no grid discretization)
• Accounts for all 22 players' positions and movements
• Separates "what happened" from "what should have happened" — enabling player evaluation beyond outcomes
• The most theoretically complete value framework in football analytics

Limitations & Debates

• Requires tracking data: the full EPV framework cannot work with event data alone, severely limiting its accessibility
• Computational intensity: computing EPV surfaces for every frame of a match is resource-heavy
• Model complexity: the framework involves multiple interacting models (pitch control, pass probability, ball value), each with their own assumptions and error propagation
• Validation difficulty: there's no ground truth for "the value of possessing the ball at this exact point at this exact moment" — validation is indirect
• Not publicly available: unlike xG or xT, no public EPV dataset or simple lookup table exists. Reproducing it requires tracking data + significant modeling effort
• Barça-centric development: much of the foundational work was done at FC Barcelona, leading to questions about how well the models generalize to different playing styles and leagues

Relationship to Other Metrics

• xT (see xT - Expected Threat) → static, grid-based, location-only version of ball value. EPV is the continuous, context-aware evolution
• Pitch Control (see Pitch Control Models) → one component of EPV (the spatial control layer)
• VAEP (see VAEP - Valuing Actions) → action-level valuation using ML on event data. Complementary approach — VAEP works without tracking data, EPV needs it but is more comprehensive
• xG (see xG - Expected Goals) → EPV includes xG as the shot component of the overall value model

Key People

• Javier Fernández — primary architect of EPV, FC Barcelona Innovation Hub
• Luke Bornn — co-developer, cross-sport spatial statistics expertise
• Dan Cervone — contributed to the deep learning EPV framework
• William Spearman — pitch control model that EPV builds upon

Resources

• Fernández & Bornn, "Wide Open Spaces" (MIT Sloan 2018)
• Fernández et al., "Decomposing the Immeasurable Sport" (MIT Sloan 2019)
• Friends of Tracking YouTube series — EPV concepts and implementations

---

Tags: #football #analytics #EPV #metrics #possession-value #spatial-analysis