/w=48,quality=90,fit=scale-down){kind=small}
Atom ScottIntroduction
Hello everyone! Atom here from Playbox.
In Part 1, we gave you a behind-the-scenes look at how match data is extracted from broadcast footage. We hope you enjoyed it!
Now, in Part 2, we’re diving into practical analysis using that very data. The theme this time is “Analysing the Ball Carrier.” We’ll explore how performance can be evaluated using modern metrics like xG (Expected Goals) and VAEP, which quantify the quality of play in a more nuanced way.
This article is written by Yuki, who we also introduced in Part 1. He is currently working as an analyst for the University of Tsukuba football team, while also conducting research in computer vision. His perspective, bridging the gap between theory and pitch, is not to be missed!
This article is part of our three-part series, “Behind the Scenes of Football Analysis.” We hope you continue to enjoy the journey!
- Behind the Scenes of Football Analytics: Part 1 – Extracting Match Data from Broadcast Footage
- Behind the Scenes of Football Analytics: Part 2 – Analysing the Ball Carrier (xG, VAEP)
- Behind the Scenes of Football Analytics: Part 3 – Analysing Off-Ball Attacking Patterns (OBSO)
1. Quantifying Football Actions
As explained in Part 1 of this series, tactical analysis in football often draws on two key data sources: tracking data, which logs the positions of individual players, and event data, which captures what happens when a player is on the ball.
Yet, when viewed in isolation, a single action from the event data is merely a recorded occurrence — it doesn't tell us how valuable or difficult that action was. Only by introducing relevant metrics and examining surrounding actions can we begin to reflect on questions like: Was that play truly outstanding? How much did it contribute to the team's success?
For instance, consider the following clip:
Trossard dribbles down the wing and delivers a cross, which Martinelli meets with a header to score.
How would you evaluate this play?
Some might exclaim, "How did he even score that?!" Others might say, "Well, he was bound to score from there." Or perhaps, "As soon as the cross came in, it was practically a goal already."
But what was the actual difficulty and value of the play? These reactions are subjective, and naturally, opinions vary from person to person.
That’s where the method we’ll explore today comes in.
By extracting player positions and action types from broadcast footage, we can use metrics such as xG (Expected Goals) and VAEP (Valuing Actions by Estimating Probabilities) to quantify both the difficulty of a shot and the quality of the actions leading up to it.
These two metrics allow us not only to objectively assess shot events in terms of goal likelihood, but also to evaluate the contribution of players involved in earlier phases — through passes, dribbles, and more.
2. What Is xG?
Overview of xG
xG (Expected Goals) is a metric that quantifies the probability of a given shot resulting in a goal, expressed as a value between 0 and 1.
An xG of 0 (0%) suggests a shot with virtually no chance of scoring, while an xG of 1 (100%) indicates a shot that is almost certain to result in a goal.
For example, a shot with an xG of 0.10 is expected to result in one goal per ten attempts.
Football is a low-scoring sport, but xG allows us to objectively assess the quality of each shooting opportunity by assigning it a numerical value.
How is xG calculated?
So how exactly is xG determined?
In its early form, xG models relied on relatively simple variables such as:
- The location of the shot
- The type of shot (e.g. header, volley)
- The shooter’s body orientation or state
At the University of Tsukuba Football Club (the author's alma mater), for instance, xG was calculated manually by:
- Dividing the penalty area into a grid and assigning values to each zone
- Applying multipliers based on shot context (e.g. direct shot, weak foot, etc.)
For example, a central shot from inside the six-yard box might be given a base value of ×0.8. If taken with the player’s stronger foot on the first touch, another multiplier of ×0.8 might apply — resulting in an xG of 0.64 for that particular shot.
However, real-world football scenarios are rarely that simple.
Modern xG models now typically use AI-based approaches, incorporating a wide range of contextual features to estimate scoring probability.
Here are just a few examples of the types of data often used:
- Distance to goal
- Shooting angle
- Type of shot
- Type of preceding action (e.g. pass, dribble)
- Positions of defenders and the goalkeeper
- Player speed and momentum
Using AI models trained on large datasets of past shots, these modern systems can generate more detailed and accurate xG values.
3. Analysing Match Scenarios Using xG
Let’s now take a look at some examples of how xG can be applied in real match analysis.
Scene-by-Scene Evaluation
We'll begin by reviewing a few match situations using an AI-based xG model trained on historical shot data. By doing so, we can examine the xG value calculated for each scene.
1. Mitoma’s Shot – Japan vs Australia (AFC Asian Qualifiers Final Round)
The xG for Mitoma’s shot was calculated at 0.032.
In other words, the model estimated only a 3.2% chance of the shot resulting in a goal — indicating that it was a highly difficult attempt.
This low value can be attributed to factors such as the tight angle from which the shot was taken and the distance from goal, which was not particularly close.
2. Mitoma’s Goal – Japan vs China (AFC Asian Qualifiers Final Round)
In this scene, Mitoma’s headed goal was assigned an xG of 0.135, meaning there was a 13.5% chance of it resulting in a goal.
With an xG of just 0.135, this goal highlights Mitoma’s ability to convert difficult chances — a clear indicator of his high-level finishing.
Although the angle to the goal was narrow and the shot was taken with a header, the distance to goal was significantly shorter compared to the previous attempt — which likely contributed to the higher xG value.
In this case, the AI model used was trained on relatively simple features such as distance to goal, shooting angle, and body part used to take the shot.
As mentioned earlier, incorporating additional features — such as the goalkeeper’s positioning and preceding events — into the model would allow for even more precise xG calculations.
Match-Level Analysis Using xG
Next, let’s take a look at an example of match-level analysis using xG.
Below are the stats from the recent La Liga Matchday 35 clash between Barcelona and Real Madrid (data sourced from FOTMOB).
The match ended in a 4–3 victory for Barcelona, but let’s examine the xG figures.
- Barcelona xG: 4.26
- Real Madrid xG: 2.74
Barcelona’s xG slightly exceeded their actual goal tally, suggesting that they converted the key chances they were expected to score. At the same time, the small gap between xG and goals may indicate that, with slightly sharper finishing, they could have added one more.
In contrast, Real Madrid’s xG was lower than their actual goal count, which indicates a high level of finishing — they were able to convert relatively difficult chances into goals.
The figure below is a visualisation of Tsukuba University's xG data from a match during the 2024 season, displayed as a heat map. (Event data provided by Bepro)
It reveals that most goal-scoring opportunities occurred near the centre of the goal line, with additional effective chances created just in front of the left side of the goal area and around the edge of the penalty arc. By visualising xG in this way, teams are able to reflect on their attacking performance, while also gaining insight into possible weaknesses in the opponent’s defence.
Furthermore, by combining xG data with information such as the type and location of the final pass leading to the shot, it becomes possible to make data-driven suggestions for training — such as designing shooting drills based on actual match trends.
xG is a relatively simple metric, but it’s also intuitive and easy to understand.
Personally, I consider it a very useful indicator because of its versatility — it can be used in a variety of analytical contexts, such as:
- A single shot
- All shots in a match
- A team’s shooting performance over an entire season
- An individual player’s shooting record
And above all, xG directly relates to what matters most in football: goals.
4. What is VAEP?
So far, we’ve looked at xG, which evaluates shots individually.
Now let’s turn our attention to VAEP, a framework that allows us to evaluate not only shots but also other on-the-ball actions throughout a match.
Overview of VAEP
VAEP (Valuing Actions by Estimating Probabilities) is a framework developed by Tom Decroos and colleagues at the DTAI Sports Analytics Lab at KU Leuven (Catholic University of Leuven) in Belgium.
VAEP assigns value to every on-the-ball action — such as passes, dribbles, or tackles — based on how much it increases or decreases a team’s expected chance of scoring.
As mentioned earlier, xG allows us to evaluate how likely a shot was to result in a goal. However, the number of shots in a match is relatively low, and xG only considers the player who takes the shot. This limits the scope of analysis.
By contrast, VAEP enables us to quantify the value of all actions leading up to a shot — allowing for a broader and more detailed assessment of each player’s contribution to the team.
How is VAEP Calculated?
VAEP measures the change in a team’s “state value” before and after an action.
The calculation involves three main steps:
- Estimate the state value before the action
- Estimate the state value after the action
- Compute VAEP as:
VAEP = (state value after) – (state value before)
Here, state value is defined as:
Expected Goals (xG) – Expected Goals Against (xGA)
For example:
If a team’s state value before a cross is 0.10, and after the cross it increases to 0.25,
then the VAEP of that cross is +0.15 — meaning it added the equivalent of 0.15 goals in value to the team.
Conversely, if a pass results in the state value decreasing from 0.10 to 0.05,
then the VAEP is -0.05, suggesting that the action put the team in a worse position and benefited the opponent.
As you can see, calculating VAEP requires models that can estimate expected goals and expected goals against for each game situation.
To do this, two separate machine learning models — one for scoring and one for conceding — are trained.
These models consider a sequence of three consecutive actions (the target action plus the two that came before it) as a single input “state.” For each state, the following features are included:
- Type of action (pass, dribble, shot, etc.)
- Ball and player positions
- Time and distance between actions
- Distance and angle to goal
Each state is labelled as either a 1 (if a goal or concession occurred within the next k actions) or 0 (if not), and the models are trained on these labels.
Once trained, these models can be used to estimate how likely a team is to score or concede at any given moment — before and after each action — making it possible to compute VAEP values across a match.
5. Analysing Match Scenarios Using VAEP
Let’s walk through two examples to see how VAEP is calculated in practice.
In each case, we’ll focus only on the six actions leading up to the shot.
1. Mitoma’s Shot – Japan vs Australia (AFC Asian Qualifiers Final Round)
Now, let’s calculate the VAEP values for the same sequence we previously analysed using xG.
The figure below shows the results of the analysis, where offensive_value and defensive_value represent each action’s impact on scoring and conceding, respectively.
In this case, Mitoma’s shot had a low VAEP since it went off target.
However, the dribble just before the shot received a high VAEP score, indicating it was an effective move that advanced the ball into a dangerous area.
Let’s also highlight Morita’s long pass, which recorded the highest VAEP value in this sequence.
While it didn’t directly result in a goal, it significantly moved the ball closer to the opponent’s goal — a contribution that the model quantified as highly valuable.
2. Martinelli’s Goal – Liverpool vs Arsenal (Premier League Matchweek 36)
This is the scene we introduced at the beginning. How did you evaluate it?
Let’s take a look at the VAEP analysis below. In this sequence, Martinelli’s header resulted in a goal, which gives him a notably high VAEP score. Since VAEP is based on expected goals, actions that directly lead to scoring — such as the final shot — tend to receive higher values.
However, what deserves particular attention here is Trossard’s cross that provided the assist. Compared to the other actions in the sequence, his VAEP score is also quite high, reflecting the quality and decisiveness of his pass.
As these two examples illustrate, VAEP allows us to quantify not only the contribution of the player who takes the shot, but also the value of every action leading up to that moment.
By calculating VAEP for an entire match, we can compare which players were making the most effective contributions.
And because VAEP is calculated for each individual action, it also enables us to analyse when and where on the pitch the team was creating meaningful attacking moments.
Conclusion
The use of broadcast footage opens new doors for advanced tactical analysis, even in environments where official data is unavailable.
By quantifying the value of actions using metrics like xG and VAEP, we can begin to answer questions such as:
“Which players consistently make decisive contributions?” or
“This player may not have eye-catching stats, but is still performing at a high level.”
Such data-driven evaluations can support not only match analysis but also recruitment decisions.
Looking ahead, further advances in AI-powered auto-tracking and video-stream analysis could bring real-time xG and VAEP scores within reach.
Moreover, these tools have the potential to benefit not just professional clubs but also amateur teams, enabling them to enhance their tactical understanding through video.
At our company, we are continuing to refine our systems and expertise for analysing xG and VAEP from broadcast footage.
We are committed to supporting football clubs and media outlets with the adoption of these tools — so if you’re interested, please don’t hesitate to get in touch.
株式会社Playboxのホームページ 👉️ https://www.play-box.ai/
自動撮影・編集ができる手頃なAIスポーツカメラ「playbox」 👉️ https://www.play-box.ai/lp
お問い合わせ
Playboxへのご質問やご相談、ビジネスの提案など、お気軽に以下のメールアドレスまでご連絡ください。
