VAEP — Valuing Actions by Estimating Probabilities

Definition

VAEP is a framework for valuing every on-ball action in football by estimating how much that action changes the probability of scoring and conceding within the next few actions. It assigns a numeric value to passes, dribbles, tackles, interceptions, shots, and every other on-ball event.

History & Origins

VAEP was developed by Tom Decroos, Lotte Bransen, Jan Van Haaren, and Jesse Davis at KU Leuven (Belgium), published in their 2019 paper "Actions Speak Louder than Goals: Valuing Player Actions in Soccer" at the KDD conference.

It emerged from a broader research effort at KU Leuven around the SPADL (Soccer Player Action Description Language) framework, which standardizes event data from different providers into a common format. VAEP was designed as the valuation layer on top of SPADL.

The work built on earlier possession value concepts but took a fundamentally different approach from grid-based methods like xT - Expected Threat|xT: instead of discretizing the pitch, VAEP uses machine learning to directly model the relationship between action context and outcome probabilities.

KU Leuven's football analytics group, including Jan Van Haaren (now at Club Brugge) and Jesse Davis, has been one of the most prolific academic contributors to football analytics.

How It Works

1. Represent each game state using features describing the last 3 actions: action type, coordinates, body part, time, result, and distance/angle to goal
2. Train two models:
   • P(scoring within next 10 actions | current game state)
   • P(conceding within next 10 actions | current game state)
3. For each action, compute the change in both probabilities:
   • ΔP_scores = P_scores(after action) − P_scores(before action)
   • ΔP_concedes = P_concedes(after action) − P_concedes(before action)
4. VAEP value = ΔP_scores − ΔP_concedes

Positive VAEP means the action improved the team's situation. Negative VAEP means it worsened it.

Key Properties

• Values all on-ball actions, not just passes/carries (like xT) or shots (like xG - Expected Goals|xG)
• Captures both offensive and defensive contributions in a single framework
• A successful tackle that wins the ball in a dangerous area gets positive VAEP
• A misplaced pass that gives the opponent a chance gets negative VAEP
• Uses ML models (typically gradient-boosted trees like XGBoost) rather than Markov chains
• Context-aware: considers the sequence of recent actions, not just location

vs xT

| Aspect | xT - Expected Threat|xT | VAEP |
|--------|-----|------|
| Method | Markov chain on pitch grid | ML model on action sequences |
| Actions valued | Passes, carries | All on-ball actions |
| Defensive actions | Not captured | Captured |
| Interpretability | Very high (lookup table) | Moderate (ML model) |
| Data requirements | Moderate | Higher |
| Granularity | Zone-level | Action-level |

Current Discussion & Debates

• Atomic VAEP (A-VAEP): an extension that decomposes VAEP into finer sub-components, separating the contributions of different action features
• VAEP vs. xT for player valuation: VAEP is generally considered more comprehensive, but xT is simpler to explain to coaches and scouts. Some practitioners use both
• 10-action window: the choice of "next 10 actions" is somewhat arbitrary. Different windows produce different valuations
• Still limited to on-ball actions: like xT, VAEP doesn't capture off-ball movement, pressing runs, or space creation — this requires tracking data and frameworks like EPV - Expected Possession Value|EPV
• Computational cost: unlike xT (a static grid), VAEP requires model inference for every action, making it heavier to compute at scale
• Adoption: increasingly used in academic research and by data-driven clubs, but less widespread in media/public analytics compared to xG - Expected Goals|xG

Accessing VAEP

• socceraction Python library — full implementation of SPADL + VAEP (GitHub)
• Requires event data with coordinates (StatsBomb, Opta, Wyscout via kloppy for normalization)
• Original paper: Decroos et al., "Actions Speak Louder than Goals" (KDD 2019)

Related Metrics

• xG - Expected Goals — values shots only; VAEP values all actions
• xT - Expected Threat — simpler grid-based alternative for ball progression
• EPV - Expected Possession Value — tracking-data alternative with continuous spatial modeling
• Packing Rate — geometric approach to valuing line-breaking actions

---

Tags: #football #analytics #VAEP #metrics #possession-value