Home/Soccer Analytics/Expected Possession Value (EPV) Model

Expected Possession Value (EPV) Model

Expected Value ModelsLevel 4 — Expert

Prerequisites

xG Model With Defender & GK Positioning

Unlocks

Anticipatory Pass Modeling (Future Position, Not Current)Backward Pass Value ParadoxCreative Decision RatingExpected Ball Progression ModelOn-Ball Value (OBV)Pass Risk-Gain DecompositionPixel-Level Field Value CalculationPlayer Decision-Making Risk Preference ProfilingPosition-Driven EPV SkewPossession Value Decomposition (Action × Success × Outcome)SARSA/RL Event Valuation

What It Is

Expected Possession Value (EPV) estimates, at any moment during a possession, the probability of scoring a goal minus the probability of conceding a goal before the possession ends — given all available spatial and contextual information. Adapted from basketball (where it was called Expected Possession Value by the Cervone et al. group). The key innovation over xG: EPV is continuous across the entire match, not just at shot moments. Every pass, carry, and dribble has an EPV before and after, enabling you to value every action in the game, not just shots.

Correct Execution

EPV is computed by: (1) encoding the current game state (player positions, ball position, phase, block type, game dynamic); (2) integrating over all possible future trajectories of that possession from historical data; (3) outputting a probability estimate of eventual goal (for) minus goal (against). The EPV curve for a possession rises and falls with each action — a good pass increases EPV, a backward pass may decrease or maintain it, a turnover drops it sharply. The change in EPV from one moment to the next (ΔEPV) is the action value.

Progression Levels

Diagnostic Tree

Coaching Cues

  • "Every action has a value — not just shots. EPV makes that measurable." — Javier Fernandez, 2019
  • "The EPV curve for a possession tells you when the team had it right and when they didn't."

Common Errors

  1. Confusing EPV with xG: xG applies only at shot moments. EPV applies to every moment of a possession. They are not interchangeable.
  2. Using EPV without contextual conditioning: EPV models trained on all possessions may perform poorly in specific phase/block/dynamic combinations — condition or sub-model as needed.
  3. Building EPV from event data alone: EPV requires player position data (tracking) to accurately model spatial control. Event data alone is insufficient.
  4. Treating EPV as a complete model of football value: EPV captures attacking possession value but misses: value of not conceding, value that persists through brief turnovers, and deliberately negative-EPV tactical plays.

Edges

🔑 Hidden Causal Lever

EPV Horizon Blindness to Slow-Developing Plays

xg-modelsexpected-possession-value →

EPV with fixed horizons (10 events, 10 seconds) assigns zero credit to initiating actions of slow-developing plays. A midfield pass triggering a goal 30 events later gets no credit. Corners leading to second-phase goals are invisible.

What most people do
Accept the fixed horizon and treat EPV as a complete valuation.
What the best do
Track "beyond-horizon goal contribution" alongside EPV. Build extended attribution models that trace full causal chains regardless of possession boundaries.
Why it's an edge: Teams excelling at slow buildup or second-phase set pieces are systematically undervalued by any standard EPV implementation.
How to exploit: Build complementary "extended attribution" tracing back from goals to the full causal chain. Use as a second lens for players with mediocre EPV but whose teams consistently score from slow sequences.
StatsBomb CTO, 2019-10-25
⚡ Conventional Wisdom Is Wrong

Intentional Negative-EPV Plays Can Be Strategically Positive

xg-modelsexpected-possession-value →

Kicking the ball out for an opponent throw-in near their corner to set up a press is EPV-negative but strategically correct. EPV drops to zero at dead-ball moments but the team expects to win back from the restart in favorable shape.

What most people do
Treat all negative-EPV actions as mistakes. Penalize players who register them.
What the best do
Catalog "strategically negative-EPV" play types. Exclude from player penalization. Separately measure post-dead-ball value.
Why it's an edge: Analysts who naively use EPV undervalue tactically sophisticated teams. Coaching staff will distrust analytics that penalizes actions they know are correct.
How to exploit: Build a "post-dead-ball value" model estimating expected EPV from the restart. A conceded throw-in deep in the opponent's half may have net-positive value once the restart is modeled.
StatsBomb CTO, 2019-10-25
🔑 Hidden Causal Lever

EPV Surfaces Value in Actions That Generate Zero xG

xg-modelsexpected-possession-value →

EPV values every action by its impact on goal probability within the full possession, not just at the shot. This means a backward pass that opens a channel, a decoy run that pulls a defender, or a ball receipt that draws pressure all have measurable EPV deltas — even though they generate zero xG and zero xT. The majority of valuable actions in football produce no shots and no zone progression, making them invisible to xG and xT frameworks.

What most people do
Evaluate midfielders and deep-lying players using xG contribution or xT delta, which only captures value at the extremes (shots or zone changes).
What the best do
Use EPV to measure the value of maintaining or creating options — the "option value" of holding the ball in a threatening position even without advancing it. A player who maintains 0.15 EPV for 8 seconds while teammates reposition is providing value that neither xG nor xT records.
Why it's an edge: The transfer market systematically underprices players whose primary contribution is creating and maintaining possession value without direct goal involvement. These are the players who make everyone else better but lack impressive xG or xA numbers.
How to exploit: Compute per-player EPV maintenance (average EPV during possessions they're involved in) separately from EPV progression (delta per action). Players with high maintenance but low delta are system enablers — cheap to acquire, expensive to replace.
Javier Fernandez, FC Barcelona, StatsBomb Innovation in Football Conference, 2019-10-22. "Even backward passes can have positive EPV if they create better forward options."

Sources

  • Javier Fernandez, FC Barcelona, StatsBomb Innovation in Football Conference 2019, YouTube, 2019-10-22 — presented Barcelona's EPV-inspired framework; described the basketball EPV origins (Cervone et al.); showed EPV curve for a live possession; introduced ΔEPV as action value; noted football requires modeling passes to any pitch location rather than just to player positions
  • StatsBomb CTO, StatsBomb Innovation in Football Conference, YouTube, 2019-10-25 — detailed EPV limitations: possession definition ambiguity, temporal horizon problem, risk of conceding not explicit, intentional negative-EPV tactical plays; motivated the SARSA alternative