A Context-Aware Framework to Evaluate Passing Decisions in Soccer

Robert Bajons | Tobias Harringer

Vienna University of Economics and Business

Sep 12, 2026

Spot the difference

Example 1
Example 2

Overview

On ball action is a chain of events:

\[ \text{choose action} \rightarrow \text{choose target} \rightarrow \text{execute action} \]

Choose action: a player selects action \(T = t\) from \(t \in \{\mathrm{pass},\mathrm{dribble},\mathrm{shot}\}\)

Choose target: depending on the choice of action \(T\), a player chooses a target. Target is a matter of definition:
- For a pass, target could be:
  - Intended end location of a pass
  - Intended receiver of a pass
- For a dribble, the target could be: intended end location of the dribble

Execute action: where did a player place the ball (similar for pass/dribble/shot)

There is a distinction between decision and execution!

Valuing an action I

The state-of-the-art approach:

\[\text{action value} = \text{realized value} - \text{current value} = V(s')-V(s)\]

\(V \dots\) some kind of value function, \(\qquad s \dots\) state before action, \(\qquad s' \dots\) state after action

Common examples:

xT difference (Singh 2019), OBV (Hudl StatsBomb 2021), VAEP difference (Decroos et al. 2019), EPV added (Fernández et al. 2021)
Difference primarily in what is contained in \(s\), which data is used, and the estimation procedure

Problems:

What happens after action \(V(s')\) often depends on other players (teammates /defenders) and only partly on the action taker
What about the decision? Was the pass/dribble/shot the best option? Was the target of the pass the best option?

Valuing an action II

Remember: On ball action is a chain of events:

\[ \text{choose action} \rightarrow \text{choose target} \rightarrow \text{execute action} \]

Suggestion: Instead of purely looking forward in time (VAEP, EPV added), progressively add action-related information:

Before the action the situation \(S\) has a certain value: \(\Vv(S) := \Ex^{\pi}[\Rs \mid S]^{\ast}\)
Value changes depending on choice \(t \in \{\text{set of actions}\}^{\ast\ast}\): \[\Vv(S,t) := \Ex^{\pi}[\Rs \mid S,\, T=t]\]
Value changes depending on choice of target \(\theta \in \{\text{set of targets}\}^{\ast\ast}\): \[\Vv(S,t,\theta) := \Ex^{\pi}[\Rs \mid S,\, T=t,\, \theta_T=\theta]\]
Value changes depending on execution \(\ell \in \{\text{set of execution outcomes}\}^{\ast\ast}\): \[\Vv(S,t,\theta,\ell):= \Ex^{\pi}[\Rs \mid S,\, T=t,\, \theta_T=\theta,\, L=\ell]\]

\(^{\ast}\Rs\) is the reward function and can (in principle) be defined as needed, \(\pi\) denotes the empirical behavior policy, i.e. the action-selection and outcome distributions observed in the data

\(^{\ast\ast}\)These sets need to be defined properly

Valuing an action III

Decision-Execution Score (DES):

\[\begin{equation} DES := \Vv(S,t,\theta,\ell) - \Vv(S) = \Dsel(S;t) + \Dtar(S,t; \theta) + \Xex(S,t,\theta;\ell), \label{eq:telescope} \end{equation}\]

\[ \begin{align} \Dsel(S;t) &:= \Vv(S,t) - \Vv(S) && \text{(type selection),} \label{eq:dsel}\\ \Dtar(S,t; \theta) &:= \Vv(S,t,\theta) - \Vv(S,t) && \text{(target selection, } t\in\{\mathrm{pass},\mathrm{dribble}\}\text{),} \label{eq:dtar}\\ \Xex(S,t,\theta;\ell) &:= \Vv(S,t,\theta,\ell) - \Vv(S,t,\theta) && \text{(execution).} \label{eq:xex} \end{align} \]

\(\Dsel \rightarrow\) value of selecting type \(t\) over other types
\(\Dtar \rightarrow\) value of selecting target \(\theta\) over other targets, conditional on selection \(t\)
\(\Xex \rightarrow\) value of execution, conditional on selection \(t\) and \(\theta\)

Theoretically cool things:

DES is mean zero \(\rightarrow\) all levels (\(\Dsel, \Dtar, \Xex\)) individually are also mean zero
Levels independent of each other: one can be negative while the others are not!\(^{\ast}\)

\(^{\ast}\)E.g., you can make a bad decision in choosing a pass (\(\Dsel\) negative), but given you choose pass, you can choose a better than average option (\(\Dtar\) positive)

Evaluating passing skill

We restrict ourselves to \(t = \mathrm{pass}\):
- Action selection decision is ignored, we assume pass was chosen
- Target for a pass: \(\theta_{\mathrm{pass}} = k \in \{1,\dots,10\} \quad(\text{intended receiver})\)

Relevant stages:
- \(\Vv(S,\mathrm{pass}) \rightarrow\) value of the passing situation
- \(\Vv(S,\mathrm{pass},k) \rightarrow\) value of choosing receiver \(k\) as target (the decision)
- \(\Vv(S,\mathrm{pass},k,\ell) \rightarrow\) the realized value (the execution)

The value:

\[\begin{equation} DES_{\mathrm{pass}} := \Vv(S,\mathrm{pass},k,\ell) - \Vv(S,\mathrm{pass}) = \Dtar_{\mathrm{pass}}(S,\mathrm{pass};k) + \Xex_{\mathrm{pass}}(S,\mathrm{pass},k;\ell). \label{eq:des_pass} \end{equation}\]

\(\Dtar_{\mathrm{pass}} = \Vv(S,\mathrm{pass},k)-\Vv(S,\mathrm{pass}) \rightarrow\) the decision value
\(\Xex_{\mathrm{pass}} = \Vv(S,\mathrm{pass},k,\ell)-\Vv(S,\mathrm{pass},k) \rightarrow\) the execution value

Back to example 1

Example 1

Probabilities show the likelihood of passing to any of the three available options.

Example 1

Probabilities show the likelihood of a successful pass to any of the three available options.

Value of the passing to a target player

Step 1: Obtain value for a successful pass to target

Value of the passing to a target player

Step 1: Obtain value for a successful pass to target
- Where does the ball end up if pass to the target is successful?

Value of the passing to a target player

Step 1: Obtain value for a successful pass to target
- Where does the ball end up if pass to the target is successful?
- What is the value of potential successful end locations?

Value of the passing to a target player

Step 1: Obtain value for a successful pass to target
- Where does the ball end up if pass to the target is successful?
- What is the value of potential successful end locations?
- Combine both to obtain expected value of a pass to target given success

Value given success = \(0.133^{\ast}\)

\(^{\ast}\)Given a pass to target player (#11) is successful, the value of the game state (likelihood of scoring - likelihood of conceding within 20s) is 0.133.

Value of the passing to a target player

Step 2: Obtain value for a unsuccessful pass to target

Value of the passing to a target player

Step 2: Obtain value for a unsuccessful pass to target
- Where does the ball end up if pass to the target is unsuccessful?

Value of the passing to a target player

Obtain value for a successful pass to target
- Where does the ball end up if pass to the target is unsuccessful?
- What is the value/risk of potential unsuccessful end locations?

Value of the passing to a target player

Step 2: Obtain value for a unsuccessful pass to target
- Where does the ball end up if pass to the target is unsuccessful?
- What is the value/risk of potential unsuccessful end locations?
- Combine both to obtain expected value of a pass to target given failure

Value given failure = \(0.004^{\ast}\)

\(^{\ast}\)Given a pass to target player (#11) is unsuccessful, the value of the game state (likelihood of scoring - likelihood of conceding within 20s) is 0.004. Even when the pass is unsuccessful in this situation the chance of regaining the ball and scoring is slightly higher than the chance of conceding a goal.

Value of the passing to a target player

Step 3: Obtain a pass success probability
- What is the probability that a pass is successful, given it is aimed to the target player?
- Pass targeted to #11 has 60.6% estimated success probability given the player positioning

Expected value of the pass:

\[ \begin{equation*} \begin{aligned} \text{Value of pass to target} &= \text{Pass success probability} \times \text{Value given success} \\ &+ (\text{1 - Pass success probability}) \times \text{Value given failure} \\ & = 0.606 \times 0.133 + (1-0.606)\times 0.004 = 0.082 \end{aligned} \end{equation*} \]

Value of the passing situation

Situational value depends on player selection likelihood and value for each player:

Value of the passing situation

Situational value depends on player selection likelihood and value for each player:
- How likely is a player the target of a pass?

Value of the passing situation

Situational value depends on player selection likelihood and value for each player:
- How likely is a player the target of a pass?
- What is the value of passing to a (relevant) player?

Value of the passing situation

Situational value depends on player selection likelihood and value for each player:
- How likely is a player the target of a pass?
- What is the value of passing to a (relevant) player?
- Combine those to obtain expected value of a the passing situation

Value of a situation = \(\sum_{k=1}^{10}\) Probability that player \(k\) is target \(\times\)

Value of pass to player \(k\) = \(0.03^{\ast}\)

\(^{\ast}\)The most likely target is player # 7 (61.5% likelihood) and has an expected value of 0.022. Player # 11 (the actual target) is the most the valuable target (0.082), but the run is difficult to spot and hence target probability is low (10.4%). The situational value is a weighted average of all values and their respective target probabilites.

Decision vs execution of a pass

The components allow us to measure value of the execution and the decision behind the pass:

Decision value:

\[ \begin{equation*} \begin{aligned} \text{Decision Score}^{1} &= \text{Value of pass to target} - \text{Value of situation} \\ &= 0.082-0.03 = 0.052 \end{aligned} \end{equation*} \]

\(^{1}\)The positive decision score reflects that a pass to the target player # 11 had more expected value than any other option. This value is completely independent of how the pass is executed.

Execution value:

\[ \begin{equation*} \begin{aligned} \text{Execution Score}^{2} &= \text{Value of realized end location}^{\ast} - \text{Value of pass to target} \\ &= 0.212-0.082 = 0.13 \end{aligned} \end{equation*} \]

\(^{2}\)The positive execution score reflects that the actual pass was successful and executed into an exceptionally valuable location.

\(^{\ast}\)Value of realized end location can be directly obtained from value surface given success.

Pass: Foden (#47) to target Haaland (#9)

Pass evaluation:
- Situational value: 0.141
- Decision: -0.095
- Execution: -0.041
- Total (DES): -0.136

Plot info:
- Boxes per candidate receiver: V = expected pass value; ps = pass success probability; \(\pi\) = likelihood of being the target
- Rings: yellow = passer; gold = intended receiver; cyan = true receiver

Results

Broadcast tracking data from
- Premier League seasons 2025/26\(^{\ast}\)
- Important: Gradient sports provides indication of intended target

\(^{\ast}\)For training of the models 6 other seasons of data were used.

Carefully filter passes:
- Checks for sensible passes (broadcast tracking is a difficult problem)
- Only passes played by foot (no headers)
- Only passes where intended target was annotated
- Total amount of passes used: 238778

Results: Raw Player rankings

Decision score
Execution score

\(^{\ast}\)Rankings based average per-match aggregate of the scores; only players with \(> 200\) passes

Testing player skill

Each score is mean-zero by construction: average behavior scores 0

\[ \mathbb{E}[\Dtar]^{\ast} = 0 \quad\Rightarrow\quad \text{a player's average is tested against a } \textbf{known zero} \]

Aggregate to per-match sums (passes within a match are dependent \(\rightarrow\) matches are the natural unit)
Test each player’s average match-sums vs 0 (one-sample \(t\)-test / robust alternative)
Correct for testing ~300 players \(\rightarrow\) FDR control (Benjamini and Hochberg 1995)

Interpretation:

Positive = better than average
Negative = worse than average
Significance = beyond chance

\(^{\ast}\)This holds for \(\Xex\) as well

Decision score vs execution score

Unadjusted
Adjusted

Summary

A pass is chain of events:

\[\text{choose target} \rightarrow \text{execute action} \Rightarrow \text{Decision vs. Execution}\]

Decision: who did the player target and how valuable was this compared to other options?
- Independent of other player \(\rightarrow\) stable player trait
- Requirement: Information on intended target!

Execution: where did a player place the ball?
- Dependent on teammates & opponents \(\rightarrow\) noisy; not a clean passer trait

You can consistently choose well; whether one pass lands perfectly depends additionally on other factors

Thank you for your attention!

(Scan QR code for slides, and more information!)

References

Benjamini, Yoav, and Yosef Hochberg. 1995. “Controlling the False Discovery Rate: A Practical and Powerful Approach to Multiple Testing.” Journal of the Royal Statistical Society Series B: Statistical Methodology 57 (1): 289–300. https://doi.org/10.1111/j.2517-6161.1995.tb02031.x.

Decroos, Tom, Lotte Bransen, Jan Van Haaren, and Jesse Davis. 2019. “Actions Speak Louder Than Goals: Valuing Player Actions in Soccer.” Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (New York, NY, USA), KDD ’19, 1851–61. https://doi.org/10.1145/3292500.3330758.

Fernández, Javier, Luke Bornn, and Daniel Cervone. 2021. “A Framework for the Fine-Grained Evaluation of the Instantaneous Expected Value of Soccer Possessions.” Machine Learning 110 (6): 1389–427. https://doi.org/10.1007/s10994-021-05989-6.

Hudl StatsBomb. 2021. “Introducing on-Ball Value (OBV).” September 16. https://blogarchive.statsbomb.com/news/introducing-on-ball-value-obv/.

Singh, Karun. 2019. “Introducing Expected Threat (xT).” https://karun.in/blog/expected-threat.html.

A Context-Aware Framework to Evaluate Passing Decisions in Soccer

Spot the difference

Overview

Valuing an action I

Valuing an action II

Valuing an action III

Evaluating passing skill

Back to example 1

Example 1

Example 1

Example 1

Value of the passing to a target player

Value of the passing to a target player

Value of the passing to a target player

Value of the passing to a target player

Value of the passing to a target player

Value of the passing to a target player

Value of the passing to a target player

Value of the passing to a target player

Value of the passing to a target player

Value of the passing situation

Value of the passing situation

Value of the passing situation

Value of the passing situation

Decision vs execution of a pass

Example 2

Results

Results: Raw Player rankings

Testing player skill

Decision score vs execution score

Summary

References

Appendix