Wide Open Spaces: A statistical technique for measuring

space creation in professional soccer

Javier Fernandez Luke Bornn
F.C. Barcelona Simon Fraser University, Sacramento Kings
[email protected] [email protected]

1 Introduction
Soccer analytics has long focused on the outcomes of discrete, on-ball events; however, much of the sport’s
complexity resides in off-ball events. In the words of Johan Cruyff: “it is statistically proven that players
actually have the ball 3 minutes on average. So, the most important thing is: what do you do during those
87 minutes when you do not have the ball? That is what determines whether you are a good player or
not.” The creation and closure of spaces is a recurrent subject in observation-based tactical analysis, yet it
remains highly unexplored from a quantitative perspective.

We present a method for quantifying spatial value occupation and generation during open play. Here
direct space occupation refers to space created for oneself, while space generation refers to opening up
space for teammates by attracting opponents out of position. We irst build a novel parametric pitch con-
trol model that incorporates motion information, relative distance to the ball, and player position in order
to provide a smooth surface of potential ball control. Through the mixture of all players’ control surfaces
we obtain a fuzzy degree of potential ball control at the team level in any given moment. We also con-
struct a model for the relative value of any pitch position, based on the position of the ball and using feed
forward neural networks. From all this (a player’s invested pitch zones, a team’s pitch control, and the
relative value of each zone), we employ the full spatio-temporal dynamics of each player to construct two
novel spatial value creation metrics, accounting for both occupation and generation of spaces.

Through the analysis of a irst division Spanish league match, we show a handful of approaches to bet-
ter understand a missing key factor for performance analysis in soccer: off-ball attacking dynamics. The
quanti ication of space occupation gain and space generation allows us to observe Sergio Busquets’ high
relevance during positional attacks through his pivoting skills, the dragging power of Luis Suarez to gen-
erate spaces for his teammates, and the capacity of Lionel Messi to occupy spaces of value with smooth
movements along the ield, among many other characteristics.

The level of detail we can reach with automated quantitative analysis of space dynamics is beyond
what can be reached through observational analysis. The capacity of evaluating space occupation and
generation opens the door for new research on off-ball dynamics that can be applied in speci ic matches
and situations, and directly integrated into coaches’ analysis. This information can be used not only to
better evaluate players’ contributions to their teams, but also to improve their positioning and movement
through coaching, providing a key competitive advantage in a complex and dynamic sport.

2018 Research Papers Competition

1 Presented by:
2 Occupation and Generation of Spaces
During the last two decades, successful elite soccer teams have been increasingly adapting possession-
centered playing styles, especially within Spanish, English and German leagues. Beyond the basic idea
of simply controlling the ball as much as possible, these strategies comprise of a large set of on-ball and
off-ball actions to generate better scoring chances. Some of these are: creating superiorities (numeri-
cal, positional or qualitative), creating disorder on the opponents defense through movement and team-
collaboration, building up plays from the goalkeeper, and executing passes with offensive intention, among
many others. Above all these actions there is a main underlying concept: the generation and occupation
of spaces. Pep Guardiola once said:

We have to pass the ball, yes, but with clear intention. Pass it to drag players to one side and
creating space in the opposite side. Then, move the ball there. That’s our game.

Occupying space on the ield is fundamentally about a player’s act of continually positioning himself
in an area of high value. The value of space can be de ined in terms of the relative position of the ball,
the closeness to the opponent’s goal, and more speci ically the level of ownership of space, regarding the
density of opponents within the given area. Furthermore, we can cluster the types of occupation of space
depending on the speed of the player. Speci ically we identify two types: active occupation, when the
player moves at running speed to earn the space, and passive occupation, when the player is below run-
ning speed (jogging or walking). For instance, if a player is closely marked and then runs towards a free
space faster than the opponent, he will obtain a gain on owned space through active space occupation. As
another example, if the player is walking towards a given area and nearby opponents move away from that
area, the player will be gaining space through passive occupation.

A more complex concept is that of space generation. We de ine the generation of space as the action of
dragging opponents out of certain areas to create new available space in previously covered areas. Specif-
ically, we identify situations where a player drags an opponent away from another teammate whom the
opponent was close to originally. The dragging concept is, at its simplest, creating space for a teammate
by pulling their defender towards oneself. Notice that unoccupied space could also be generated when
dragged players leave a clear area; however, we are not considering this case for this study. Similarly to
the Space Occupation Gain (SOG) concept above, later we also explore the concept of Space Generation
Gain (SGG). In this way, we separate out space created for oneself from space created for teammates, in
both a passive and active manner.

Figure 1 presents an example of both space occupation and space generation during an of icial Span-
ish irst division match. The three images show a process where Andrés Iniesta moves to clear up space
away from the ball and then attacks a high value space inside the box. When he moves to this space he
drags three defenders towards himeslf while also receiving a pass. The attraction of the three defenders
leaves open space for Lionel Messi, who in this newfound space receives a pass free of a mark, and subse-
quently sends a lob pass onwards for Suarez, who meanwhile was running towards the goal line in search
of space of value to score. A more detailed video example of occupation and generation of spaces can be
seen at the following link, where players are highlighted when adding space for themselves or teammates:
http://www.lukebornn.com/sloan/space_occupation_1.mp4

Before providing explicit details on how to calculate space occupation and generation, we irst need a
better notion of space ownership and value, as creating space for your own goalkeeper who is 80 meters
2018 Research Papers Competition
2 Presented by:
 Figure 1: A game situation presenting both space occupation and generation. From left to right: in the
irst frame Iniesta moves back to occupy a space of value with higher control. In the second frame, Iniesta
observes an open space to attack. He moves towards the space, dragging three defenders. In the third
frame, the three dragged defenders leave an open space for Messi that can now receive the ball free of
mark, while Suarez runs towards the goal line enabling him to receive a pass.

behind the location of play is worth much less than creating space in high-threat areas nearer the ball and
goal. The next two sections present a novel pitch control model for evaluating space ownership, and a
dynamical model for space value according to ball and player position.

3 Modelling Pitch Control: A Parametric Approach

Pitch control is a recurrent concept in the analysis of space dominance in team sports. It can be de ined as
the degree or probability of control that a given player (or team) has on any speci ic point in the available
playing area. The emergence of player tracking data has given rise to different pitch control (or dominant
region) models. A widely applied model is the Voronoi tesselation, which takes into account the position
of all players on the ield and calculates the closest player to each given spatial point, inding dominant
cells for each player. This model has been used for quantifying the dominant area of attacking and defend-
ing teams in constrained playing areas [1], to evaluate the space dominance based on passing behaviour
[2], to improve models for pass probabilities [3], and to evaluate positional value of players on rebounds
in basketball [4], among many other applications. From the original model for team sports presented by
Taki and Hasegawa [5] there have been several extensions for faster computation, as well as extensions to
incorporate motion and weighted valuation of dominant space [3, 6]. Beyond its bene its, all the different
Voronoi tesselation-based approaches start from the idea of inding regions that are exclusively dominated
by a given player. This concept disregards the concept that ownership of space is continuous, not discrete,
with uncertainty in who controls areas between players. Also, the distance between players and the ball
is also believed to in luence the relative positioning and degree of space control, especially for sports with
wider playing spaces such as soccer; however, this is not taken into account by the mentioned approaches.

2018 Research Papers Competition

3 Presented by:
 We propose a novel pitch control model that takes into account the location, velocity and distance to
the ball of all the players, providing a smooth surface of control for each team. For any given location, the
in luence that every player has in that place is computed and summarized, resulting in a probability of
control. An additional objective of this approach was to provide a model that could be applied in a given
data frame, without requiring signi icant data for learning its parameters. This is particularly important
for clubs in competitions such as the Spanish League where tracking data is not available for direct usage.
Also, such a model would allow easier reproducibility.

3.1 Player In luence Area

Depending on their location in time, players might have different levels of in luence on nearby zones. When
a player is far away from the ball his level of in luence can be understood as a wider area, based on the rea-
soning that if the ball moves towards the player he would have a more time to reach the ball within a
larger space. On the opposite, when closer to the ball, the player has less possibilities of reaching the ball
if it moves from its current location. Also, the player’s velocity plays an important role in de ining the area
of in luence. A player at running speed might have more in luence in the areas in the direction of speed
compared to if they were walking or jogging. Further, the player may have higher levels of in luence in
close spaces than in farther spaces.

Based on this reasoning we propose de ining the player in luence area through a bivariate normal dis-
tribution, whose shape can be adjusted to account for the player’s location, velocity, and relative distance
to the ball. At any given location a degree of in luence or control can be queried through the distribution’s
probability density function.

Speci ically, the player’s in luence I at a given location p for a given player i at time t is de ined by a
bivariate normal distribution with mean µi (t) and covariance matrix Σi (t), given the player’s velocity ⃗s
and angle θ. For a given location in space p at time t, the probability density function of player i in luence
area is de ined by a standard multivariate normal distribution. The player’s in luence likelihood is then
de ined as the normalization of f at the given location p by the value of f at player’s current location pi (t),
as shown in Equation 1.

fi (p, t)
Ii (p, t) = (1)
fi (pi (t), t)
This formulation provides an initial model for obtaining a degree of in luence within a [0, 1] range for
any given location on the ield. The mean and covariance matrix can be dynamically adjusted to provide a
player dominance distribution that accounts for location and velocity. In Appendix A.1 we provide speci ic
details for this equation.

Figure 2 presents the player in luence area in two different situations regarding the player’s distance to
the ball and velocity. Here we can observe how depending on the distance to the ball the range of in luence
of the player varies. Also, the distribution of player in luence is reshaped to be oriented according to the
direction of movement and stretched in relation to the speed. If the player is in motion, the distribution
is translated so the higher level of in luence is near points where the player can reach faster, according to
his speed. This model can easily be expanded to handle player-speci ic movement characteristics, such as
acceleration and maximum speed.

2018 Research Papers Competition

4 Presented by:
(a) Player in luence area for player in possession of the(b) Player in luence area for player 15 meters away from
ball and no speed (lower than walking speed) the ball, running at 6.36 m/s in a 45 degrees angle

Figure 2: Two situations representing the player in luence area

3.2 Modelling Team Pitch Control

When de ining a team’s degree of control at any given location on the ield, it is desirable to take into
account the level of in luence each individual player of both teams is having on that point. Since many
players can have in luence at a given location at a certain time, the model should be able to account for the
aggregated in luence of each player and provide a value of control within a continuous range, instead of
strict areas such as the case of Voronoi Tesselation.
Based on this, we present a pitch control model that summarizes the level of in luence of every player,
and outputs a degree of control for any part of the pitch. Equation 2 presents the pitch control level at
a location p at time t, where i and j refers to the index of the player in each opposing team. Here the
logistic function transforms the substraction of the accumulated individual in luence area of each team
into a degree of control within the [0, 1] range. Also, since we are de ining a team-oriented pitch con-
trol model, a single player without any in luence of any other player at its current location only controls
logistic(1) = 0.73 of the space. This provides the need of higher density of players near a given area to pro-
vide higher level of control in that area. For the statistically-inclined, note that this formulation represents
the probability of control of a given team, where each team’s latent surface is captured by a kernel-based
non-parametric point process,
∑ ∑
P C(p, t) = σ( I(p, t) − I(p, t)) (2)
i j

where σ is the logistic function. Since the pitch control model follows the de inition of player in luence
area in Figure 2, the model is taking into account the location of the ball, the players’ velocities and the
location of all the players on the ield. Equation 2 is a simpli ied version of pitch control calculation based

2018 Research Papers Competition

5 Presented by:
 Figure 3: Pitch control surface indicating the degree of control for team in red. Arrows show players
velocities, and contour lines allow to visualize the surface geometry. Numbers in white indicate the pitch
control value at their drawing location. Axis dimensions are in meters.

on players’ in luence areas. Note we can include a constant within σ to add more lexibility, if desired.
Figure 3 presents the pitch control surface in a given situation of the match. At location (82, 8), near the
ball, it can be observed clearly how the yellow team’s high density provides lower level of control for the red
team near the ball. Also the velocity of the player in possession of the ball (red team) provides the red team
an advantage in the running direction. At location (80, 25), the red player is creating a positional advantage.
Meanwhile, at location (50, 30) the yellow player has minimal control of space because of the high density
provided by the three surrounding opponent players. For a single time frame, this pitch control model
provides a synthesis of player locations, player velocities and ball-relative positioning in one variable. Also,
by exploiting the dynamics of pitch control time, it becomes a versatile tool for evaluating multiple types
of spatio-temporal characteristics of the game such as the creation of positional advantages, the in luence
of density and pressure speed in defending situations, and the creation and generation of spaces.

2018 Research Papers Competition

6 Presented by:
4 Quantifying Pitch and Space Value
Although the control of space is a fundamental element to identify occupation and generation of spaces,
we still need an additional part of the equation: the value of space. The sole fact of moving for inding
better passing options is an advantage itself. However, it can be easily argued that not every space has the
same value. A trivial method for determinating the value of space is its distance to the opponent goal. Its
well known that spaces near the goal have an increased value, given the advantage that would provide to
dominate them. But exploring more deeply into the dynamics of soccer, and based on the opinion from F.C.
Barcelona expert analysts, it can be also argued that the value of space changes dynamically depending on
multiple positional factors, such as the position of the ball and the players. In order to quantify in a de-
tailed way the value of the space generated or occupied we provide a novel model for inding the relative
pitch value on every position of the ield, depending on the location of the ball. The following link presents
a video where the dynamic evaluation of pitch value depending on the ball position, as detailed below, can
be observed: http://www.lukebornn.com/sloan/field_value.mp4.

Instead of de ining a priori a model for space valuation we would like to extract a sense of space value
from the spatio-temporal behaviour of players during multiple matches. For this we set the following
hypothesis: considering a suf iciently high number of situations, the defending team distributes itself
throughout the ield in a manner which covers high value spaces. Although it is clear that at any given
point defenders will deviate based on overloads, speci ic offensive player positioning, and other scenar-
ios, in general, most players will remain close to high value areas. An extreme example of this will be the
case where the attacking team places all players in the middle of ield. It is arguable that, although this
would impact the position of the defending team, they will most probably still keep players near the box
and their own goal. Note that similar ideas are used when identifying defensive matchups based on de-
fender locations in basketball [7].

Based on this, we propose learning the sum in luence that a defensive team would have in a given
location on the ield, given the location of the ball. Let Vk,l (t) be the value of location pk,l of the pitch at
time t, and let pb (t) be the location of the ball at time t, we want to learn a function f n with parameters θ
that values space as a function of of the ball,

Vk,l (t) = f n(pb (t), pk,l (t); θ) (3)

At a irst glance it seems the relation between the position of the ball and the ield location for predict-
ing location value is complex and non-linear. While we have tested several linear models, we have observed
signi icantly improved performance with non-linear alternatives, so stick to those here. In order to solve
the proposed problem, we use a feed forward neural network with one hidden layer, that aims to learn the
parameters θ of the mapping de ined in Equation 3. For the learning process we build a dataset where the
target value Vk,l (t) is calculated by minimizing the loss 4, which corresponds to the sum of player in luence
for every defending player d in a given situation.
∑
Dk,l (t) = Id (pb (t), pk,l (t)) (4)
d
{
1 Dk,l (t) > 1
Vˆk,l (t) =
Dk,l (t) otherwise

2018 Research Papers Competition

7 Presented by:
(a) Pitch value for ball vertically centered at (b) Pitch value for ball at the center of the ield
the irst quarter of the ield

(c) Pitch value for ball at the third quarter of (d) Pitch value for ball vertically centered in
the ield on top of the left lane the fourth quarter of the ield

Figure 4: Predicted pitch value in a [0,1] range for given ball location (white circle)

Defensive situations are found by selecting game situations where the opponent has possession of the
ball. Then, the sum of player in luence is found for every location (k, l) within a 21 by 15 grid, for every
defending player i. Situations are selected so they are separated in time by at least three seconds. Here
we employ Metrica Sports tracking-data of 20 matches of irst and second B Spanish division, consisting of
2.4 million examples. For learning the parameters we use a feed forward neural network with one hidden
layer using the adam optimization algorithm [8]. Speci ically, we aim to ind the optimal parameters θ∗ that
minimize the loss function L as presented in Equation 5. We selected mean square error as loss function
L and sigmoid function as the activation function f .

1∑
n
L(θ) = arg min L(ye , f (xe , θ)) (5)
θ n
e=1

We found the best model through a 10-fold cross-validation process. In order to obtain a valuation of a
ield location for a given ball location, we now query the learned model. Figure 4 shows three different
ball position scenarios and the obtained ield valuation. This model has learned that nearby locations to
the ball have increasing value for a certain range, while understanding effectively how to translate this
value depending on ball position. The model still lacks from the natural intuition that space generated
at the higher valued locations of the irst quarter of the ield should not have an identical valuation than
those of higher valued locations at the last quarter. In other words, the cumulative value of space is higher
when further up the ield, closer to the opponent’s goal. In order to adapt to this intuitive thinking we
2018 Research Papers Competition
8 Presented by:
(a) Distance to goal pitch value normalization (b) Normalized pitch value for ball vertically
surface centered at the irst quarter of the ield

(c) Normalized pitch value for ball at the cen- (d) Normalized pitch value for ball vertically
ter of ield centered the fourt quarter of the ield

Figure 5: Predicted pitch value in a [0,1] range for given ball location (white circle) normalized by a dis-
tance to goal model

normalize the obtained pitch value by the distance to the goal of every location normalized on a [0, 1]
range. Figure 5 presents the normalization surface and three different pitch value situations, where the
results still adapt to ball location but show a more consistent valuation of the pitch which adjusts for the
threat of the ball location, according to expert analysts. We see that when one’s own goalkeeper has the
ball, the overall value of space is limited, but when in the opponent’s box, space is much more valuable
alongside the looming threat of a shot on goal.

5 Occupation and Generation of Spaces

Previously, we approached the occupation and generation of spaces as actions focused on the improvement
of the quality of team positioning, with the purpose of reaching better goal scoring chances. The quality of
positioning of a given player is then related with having the best possible control of the space, and doing
so for spaces with higher value. We could then express the quality of owned space Q as a function of the
level of ownership (control) P C and the value of space V , as presented in Equation 6.

Qi (t) = P Ci (t)V (t) (6)

Based on the de initions of Sections 3 and 4 we can model P Ci (t) through our team pitch control model,
and V (t) using the ball-relative ield value model. Figure 6 presents the team pitch control, the pitch value
2018 Research Papers Competition
9 Presented by:
and the obtained quality of owned space, in a given match situation. We can now de ine with detail our
two main proposed concepts: Space Occupation Gain and Space Generation Gain.

5.1 Space Occupation Gain

Now that we have the necessary tools to represent the value of space ownership in a given time, we can
de ine a model for identifying gain in space occupation in time. As mentioned in section 2 we propose
the Space Occupation Gain (SOG) concept as the relative amount of quality of owned space earnt during a
time window. An opposite concept is that of Space Occupation Loss (SOL), which relates to a negative gain
during the time window. We irst de ine the concept of gain in time G as the mean difference of quality
of space occupation Q during a time window [t + 1, t + w + 1], for a given player i. This is expressed in
Equation 7.
∑t+w+1
t′ =t+1 Qi (t′ )
Gi (t) = (7)
w
Given the dynamic nature of football, players are involved in a continuous process of winning and losing
space. A small gain of space can happen when the nearby defenders follow the ball when it moves away
from the player, leaving the player a better control of space. However the same can happen in a high speed
running situation between the attack and the defender, where the attacker is moving slightly faster. In
another case, a medium or high gain of space can happen when the player moves towards a free space.
Given this, it is necessary to de ine a level of space gain from which the earned space can be considered
an actual occupational advantage and not a consequence of slower-moving contextual factors in a given
situation. We set a constant ϵ as a threshold to account for space occupation gain only in the cases the gain
is above that threshold. We can do the equivalent for space occupation loss. Both expressions are de ined
in Equations 8 and 9.
{
Gi (t) if Gi (t) ≥ ϵ
SOGi (t) = (8)
0 otherwise
{
−Gi (t) if Gi (t) ≤ −ϵ
SOLi (t) = (9)
0 otherwise
An additional concept for re ining the idea of gaining space quality is the way that space is gained,
speci ically regarding a player’s speed. We present two de initions: active and passive space occupation

(a) Pitch control surface (b) Pitch value based on ball position (c) Value of the owned space as prod-
uct of pitch control and ield value

Figure 6: Pitch control, ield value and value of owned space for attacking team in red, for attacking di-
rection left to right

2018 Research Papers Competition

10 Presented by:
gain. A space is occupied actively when the player moves towards that space with a greater speed than a
jogging pace (>1.5m/s [9]). Otherwise we consider that space to be occupied passively.

5.2 Space Generation Gain

The generation of space for teammates is a concept that involves two or more teammates during certain
attacking situation. Two main types of actors are present: one generator and one or more receivers. The
generator is a player that moves toward a certain space while dragging opponents during the process. This
dragging behaviour causes the freeing up of space previously occupied by the dragged opponent. When
that opponent was previously close to one or more other teammates we say those players are receiving
space generated by the attracting player. In order to express this concept mathematically, we need to de ine
a value for closeness. We will say that a player is close to another if the distance between them in a given
time t is below a constant δ. Also, it is desirable to de ine another constant α for constraining the minimum
attracting distance, which refers to the difference in distance between the starting and end position of the
generator and the attracted opponent. This allows us to avoid innaccurate attractions when players are
very close to each other initially. Given this, let di,j (t) be the distance between players i and j, Equation 10
expresses the concept of space generation SG between any pair of teammates (i, i′ ) and any opponent j,
for a time window [t, t + w].

SGi,i′ (t) = ∃j (di′ ,j (t) ≤ δ) ∧ (di,j (t + w) ≤ δ) ∧ (di′ ,j (t + w) > δ) ∧ (di,j (t + w) − di,j (t) > α) (10)

Once we can identify when a space generation behaviour is occurring, we would like to focus on the
cases in which we actually have a gain in space due to the dragging effect. Analogously to the SOG de ini-
tion, we express the Space Generation Gain (SGG) as space generation situations where the gain is above
a threshold ϵ, as presented in Equation 11.
{
Gj (t) if SGi,j (t) ∧ Gj (t) ≥ ϵ
SGGij (t) = (11)
0 otherwise
Essentially, we are attributing space gain to a player when a defender leaves his mark and moves to-
wards a teammate, subject to the conditions that the defender was close to the player and ended close to
the teammate during a time window. It is important to clarify that while SOG and SGG represent two
frequent and relevant cases of space gain within soccer, other types of situations and movements might
contribute as well to the total space created by a player during a match. An additional possible concept is
that of potential space, referring to a space that the player is more likely to reach, within his positioning,
but not in his immediate in luence area. We will now focus on analyzing SOG and SGG within a match
context.

6 Match Analysis
The ability to create and occupy spaces are two commonly trained concepts in modern soccer. During
training, coaches interrupt and reshape individual drills to teach players how to orient and move toward
spaces and away from low value local zones on the ield. When analyzing off-ball performance, coaches
appeal to video analysis. Although elite soccer analysis staff typically have a great capacity to understand
complex concepts through match visualization, the dynamics of space creation are so frequent and hap-
pen in such short time windows, that it becomes impractical for video analysts to grasp them all, even for
2018 Research Papers Competition
11 Presented by:
a single match. However, is important to note that there is no existance of ground truth data regarding
the quanti ication of spaces in soccer. Hence we have performed an extensive validation of the devel-
oped concepts through video and studying individual situations within games, with the help of two expert
soccer video analysts from F.C. Barcelona, in order to ine-tune our quantitative approach. The follow-
ing videos are examples of the video-based validation tool we have used: http://www.lukebornn.com/
sloan/space_occupation_1.mp4, http://www.lukebornn.com/sloan/space_occupation_2.mp4

Based on this, we provide a complete summary of off-ball movement statistics for a speci ic Spanish
irst division of icial match between F.C. Barcelona and Villareal F.C. in January 2017. Speci ically, we pro-
vide an analysis focused on the concepts of space occupation and space generation, using Metrica Sports
optical tracking data. This match ended with a 1-1 result, where the irst goal was scored by Villareal
F.C. at the 49th minute (second half), and the F.C. Barcelona equalizer came at the 90th minute by Lionel
Messi. Situationally, this presents a game where F.C. Barcelona was in need of scoring during the inal
minutes, and were required to occupy and generate the most spaces possible to reach scoring chances. In
order to identify space occupation and generation actions we calculate for the attacking situations of F.C.
Barcelona all the instances where a player had controlled possession of the ball with his feet. From each
of those situations, and alongside expert football analysts from F.C. Barcelona, we de ine a window w of
three seconds after each of these cases, reaching a total of 845 different situations. The closeness factor δ
is set to 5 meters, based on the minimum distance an opponent is on average to a player in possession of
the ball. We also set the minimum attraction distance for space generation α to 3 meters.

Table 1 presents the space occupation statistics for F.C. Barcelona, sorted in descending order by the
total amount of Space Occupation Gain (SOG). At irst glance it can be seen that over 41% of gain of space
occupation was performed by Iniesta, Sergio Busquets and Lionel Messi. Notably, these three players oc-
cupy different positions and have different roles within the team. Busquets is a pivot and has a speci ic
role of helping to drive the ball with controlled possession during build-ups, and to accompany the game
creation during positional attacks. Iniesta is an attacking mid ielder with great control of the ball, and
special skills in moving and inding spaces between lines. Messi is an attacker but not attached to a spe-
ci ic position, and is allowed to cover wide areas of the pitch to ind space and request the ball. The three
players share, however, a long-time tradition of possession-centered and off-ball movements quality dur-
ing their career. Suarez and Neymar, two highly mobile players, appear with a lower count of situations
where space was gained. This can be associated with the high level of strictly closed marking these players
suffered during the match.

It is interesting to observe that for most players the active occupation of spaces is considerably more
frequent than passive occupation. This is particularly noticeable on left and right backs Digne and Sergi
Roberto, who need to cover wider spaces and show a high mean distance to ball for SOG, a characteris-
tic shared by central defenders Pique and Mascherano. A remarkable case is that of Lionel Messi, whose
passive SOG is considerably higher than the active one. The passive characteristic of SOG does not mean
the player is not occupying the space intentionally, but rather that he is not moving at running speed, but
slower. Much has been argued in recent years about several moments during matches where Messi walks
through zones of the ield. However, that walking behaviour is not a detachment from the match but a
conscious action to move through empty spaces of value and claim the control of valuable space, and ul-
timately the ball. Messi does this very effectively, placing him near the top of players in terms of space
gained during the whole match, despite the lack of active gain. A relevant characteristic of this is that 71%
of the time the gain in space is done in front of the ball rather than behind. The in front and behind the ball

2018 Research Papers Competition

12 Presented by:
 ∑
# SOG SOG µ SOG Active (%) Passive (%) FRT BEH MBD Mins
Name
Iniesta 96 (14.8%) 15.77 0.16 56.25 43.75 49 47 15.19 94.86
S. Busquets 90 (13.9%) 14.85 0.16 47.78 52.22 44 46 16.65 94.86
Messi 81 (12.5 %) 14.72 0.18 33.33 66.67 58 23 17.50 94.86
A. Gomes 74 (11.4%) 12.58 0.17 68.92 31.08 40 34 15.93 68.61
Suarez 70 (10.8%) 12.27 0.18 57.14 42.86 57 13 13.46 94.86
Neymar 61 (9.4%) 9.46 0.16 59.02 40.98 48 13 18.31 94.86
S. Roberto 51 (7.8%) 7.34 0.14 78.43 21.57 25 26 25.10 94.86
Pique 29 (4.4%) 4.92 0.17 48.28 51.72 6 23 21.05 94.86
Mascherano 29 (4.4 %) 4.54 0.16 41.38 58.62 2 27 22.03 94.86
D. Suarez 22 (3.4%) 4.07 0.18 77.27 22.73 13 9 17.47 26.25
A. Turan 17 (2.6%) 3.51 0.21 52.94 47.06 12 5 12.71 23.32
Digne 26 (3.2%) 3.48 0.13 80.77 19.23 13 13 16.23 71.54

Table 1: Statistics of ∑
space occupation for F.C. Barcelona in an of icial Spanish League match against Vil-
lareal F.C. Symbols #, and µ represent the total, sum and mean of their associated variable. SOG refers
to Space Occupation Gain, while FRT and BEH indicate the amount of times SOG occurs in front or behind
the ball. MBD represents the mean ball distance, and Active (%) and Passive (%) the player percentage of
times the space was occupied through active or passive occupation.

statistics show a clear tendency for central defenders to gain space behind the ball, while attackers show
a higher rate of space gain in front of the ball. Noticeably Busquests, Iniesta and the right and left backs
(Digne and S. Roberto) have a balanced ratio of space gain behind and in front of the ball.

Table 2 presents the statistics for Space Occupation Loss (SOL) and Space Generation Gain (SGG). The
SOL statistics show a clear tendency of higher space loss for players that are more often in possession of
the ball such as Iniesta, Messi, Neymar and Suarez. The space loss can be directly associated with pressure
by the opponent, who tends to increase density near to attacking players to reduce their range of action,
especially for highly skilled players. Regarding the generation of space, we obtain a different picture from
the space occupation skills. Here, Neymar and Suarez appear to be, alongside Messi, the players that most
often drag opponents to create space. With a 4-3-3 system and high-quality players, a speci ic attacking
strategy is that of spreading out attacking players to drag defenders out of position and provide wider
spaces for attacking action. Busquets, a pivoting specialist, appears also at the top of the table showing his
value in supporting space creation. Notably the left and right back, Digne and S. Roberto do not generate
much space. Given that they move towards the border lines of the ield, it is less likely that opponents are
dragged by back defenders.

A more detailed perspective of space generators and receivers is presented in Figure 7. Here we can
observe the amount of times generators are producing space for receivers, and discover some collaborative
playing behaviour. First to observe is that Busquets receives space from most of the players at least once,
possibly showing his ability to stay at the center of play. A renowned skill of F.C. Barcelona is the third-
man pass, which consists of the following: if a player A wants to pass to player C, but is marked, he passes
to player B, dragging the opponents toward him, enabling C to receive the ball in more space. This plot
might show a third-man behaviour through Busquets. Notably, Suarez, Neymar and Messi generate space
commonly for each other, especially Suarez who provides considerable space to both. A special connection

2018 Research Papers Competition

13 Presented by:
 ∑ ∑
# Generated # Received SGG µ SGG # SOL SOL µ SOL Mins
Name
Neymar 28 (18.9%) 6 (4.1%) 5.97 0.21 51 -8.53 -0.17 94.86
Suarez 25 (16.9%) 18 (12.3%) 5.60 0.22 52 -9.12 -0.18 94.86
Messi 22 (14.9%) 24 (16.4%) 4.32 0.20 68 -11.61 -0.17 94.86
S. Busquets 15 (10.1%) 24 (16.4%) 3.83 0.26 38 -6.16 -0.16 94.86
Pique 14 (9.5%) 9 (6.2%) 3.66 0.26 19 -2.77 -0.15 94.86
Iniesta 13 (8.9%) 21 (14.4%) 2.62 0.20 75 -11.79 -0.16 94.86
A. Turan 8 (5.4%) 7 (4.8%) 2.26 0.28 8 -1.29 -0.16 23.32
S. Roberto 7 (4.7%) 2 (1.4%) 1.55 0.22 31 -4.62 -0.15 94.86
A. Gomes 9 (6%) 18 (1.2%) 1.49 0.17 44 -6.25 -0.14 68.61
Mascherano 5 (3.4%) 9 (6.2%) 0.80 0.16 23 -3.39 -0.15 94.86
D. Suarez 2 (1.4%) 8 (5.5%) 0.46 0.23 16 -3.14 -0.20 26.25

Table 2: Statistics of space generation, and space occupation

∑ loss for F.C. Barcelona in an of icial Span-
ish League Match against Villareal F.C. Symbols #, and µ represent the total, sum and mean of their
associated variable. # Generated and # Received indicate the total times a player generated or received
generated space, accompained by the team-relative percentage. SGG refers to Space Generation Gain and
SOL refers to Space Occupation Loss.

between Suarez and Messi is also shown for this game, where both were able to generate a high amount
of space for each other.
A further vision of space gain and generation can be grasped from Figure 9. Here we present the spa-
tial heatmap for SOG and SGG situations. At irst glance we observe the amount of space gained through
occupation is considerably higher than through generation, a more complex process. Iniesta presents an
interesting case where he can generate more space next to the left border line of the ield, while he is bet-
ter at gaining spaces for himself at the interior of the ield. Also, he produces a notable amount of space
near the box. Busquets shows an incredible collaborative behaviour by generating space almost anywhere
around the ield. He also presents wide areas of SOG, but more intensively near the mid ield, his natural
habitat. Suarez presents a notable ability to generate space within the box, where he concentrates most
of his generating contribution. Here he arises as a specialist in dragging defenders either while making
spaces for himself or while generating spaces for others. Messi also shows a great ability in generating
spaces around the attacking zones of the ield, while Neymar concentrates on the left wing, focused on
high speed diagonal runs towards the box. Defenders, as expected, show very little generation of space.

7 Discussion
In a sport where the average possession of the ball by a player is 3 minutes in a 90 minute game, the analy-
sis of team-collective dynamics through off-ball movements becomes a critical element for understanding
performance. We have shown how through spatio-temporal data it is possible to extract meaningful in-
formation relating the occupation of spaces of value and the generation of spaces for teammates. Beyond
the bigger picture that overall performance statistics of multiple matches can provide, the understanding
of off-ball movements demands the need for a more specialized per-match or even per-situation analysis.
Through the understanding of the frequencies, quality, position and effectiveness of space occupation and
generation, a coach can provide speci ic guidance to players to help the team playing dynamics beyond
what he can do with the ball.
2018 Research Papers Competition
14 Presented by:
 Figure 7: A heatmap showing the total times space was generated by generators (y-axis) for receivers
(x-axis)

In order to provide a deeper understanding of space, we have presented two novel approaches for
pitch control and pitch value modelling. Our pitch control model takes into account critical factors when
understanding the dominance of space such as the velocity and position of the player. It also provides a
key element that was missing in previous dominant region models: the idea of a soft surface of control
where for a given location on the ield, nearby players have a certain level of in luence, instead of de ining
strict dominance margins such as in Voronoi-based models. On the other hand, the proposed pitch value
model presents a way of quantifying the value of every location on the ield in a dynamic way, relative to
the location of the ball. This way, we can account for both the control of space a team has and the value of
that space, to obtain a measure of spatial value controlled.

For future studies, the proposed pitch control and ield models can be directly applied for reaching
more comprehensive pass probability and reward models, and in general to incorporate a new perspective
on dominant regions based approaches for understanding team sports. But more generally, this study sets
a base for new research on off-ball behaviour in soccer. New perspectives are still to be studied, such as
the effect of different pressure strategies, the concept of potential space and how it could be exploited, the
overall dynamic balance of space control between the two teams and its association to performance, as
well as many other research lines that address a critical question when training to succeed in soccer: what
should I do when my teammate is in possession of the ball.

2018 Research Papers Competition

15 Presented by:
Figure 8: Space Occupation Gain and Space Generation heatmap for every ield player playing over 60
minutes. The scaling factor is based on the maximum Space Occupation and maximum Space Generation
among all the team, respectively.

2018 Research Papers Competition

16 Presented by:
References
[1] Pedro Silva, Paulo Aguiar, Ricardo Duarte, Keith Davids, Duarte Araújo, and Júlio Garganta. Effects of
pitch size and skill level on tactical behaviours of association football players during small-sided and
conditioned games. International Journal of Sports Science & Coaching, 9(5):993–1006, 2014.

[2] Robert Rein, Dominik Raabe, Jürgen Perl, and Daniel Memmert. Evaluation of changes in space con-
trol due to passing behavior in elite soccer using voronoi-cells. In Proceedings of the 10th Interna-
tional Symposium on Computer Science in Sports (ISCSS), pages 179–183. Springer, 2016.

[3] Michael Horton, Joachim Gudmundsson, Sanjay Chawla, and Joël Estephan. Classi ication of passes
in football matches using spatiotemporal data. arXiv preprint arXiv:1407.5093, 2014.

[4] R Masheswaran, Y Chang, Jeff Su, Sheldon Kwok, Tal Levy, Adam Wexler, and Noel Hollingsworth. The
three dimensions of rebounding. MIT SSAC, 2014.

[5] Tsuyoshi Taki, Jun-ichi Hasegawa, and Teruo Fukumura. Development of motion analysis system
for quantitative evaluation of teamwork in soccer games. In Image Processing, 1996. Proceedings.,
International Conference on, volume 3, pages 815–818. IEEE, 1996.

[6] Dan Cervone, Luke Bornn, and Kirk Goldsberry. NBA court realty. MIT Sloan Sports Analytics Con-
ference, Boston, MA, USA, 2016.

[7] Alexander Franks, Andrew Miller, Luke Bornn, and Kirk Goldsberry. Counterpoints: Advanced de-
fensive metrics for NBA basketball. In 9th Annual MIT Sloan Sports Analytics Conference, Boston, MA,
2015.

[8] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint
arXiv:1412.6980, 2014.

[9] Angel Ric, Carlota Torrents, Bruno Gonçalves, Lorena Torres-Ronda, Jaime Sampaio, and Robert Hris-
tovski. Dynamics of tactical behaviour in association football when manipulating players’ space of
interaction. PloS one, 12(7):e0180773, 2017.

[10] A geometric interpretation of the covariance matrix. http://www.visiondummy.com/2014/04/

geometric-interpretation-covariance-matrix. Accessed: 2017-11-10.

2018 Research Papers Competition

17 Presented by:
A Appendix
A.1 Player In luence Area Details
Section 3.1 presents a player in luence model that accounts for the position, velocity and distance to ball of
a given player. The in luence degree Ii at Equation 13 is expressed in terms of the probability density func-
tion of a bivariate Gaussian distribution de ined by Equation 12. In this section we detail the calculation
of each of the elements involved in the equation.
1 1
fi (p, t) = √ exp(− (p − µi (⃗si (t)))T COVi (t)−1 (p − µi (t))) (12)
(2π)2 detCOV i (t) 2
fi (p, t)
Ii (p, t) = (13)
fi (pi (t), t)
The covariance matrix can be dynamically adjusted to provide a player dominance distribution that
accounts for location and velocity. Using the singular value decomposition algorithm we can express the
covariance matrix as a function of its eigenvectors and eigenvalues as expressed in Equation 14, where
V is the matrix whose columns are the eigenvectors of Σ, and L is the√diagonal matrix whose non-zero
elements are the corresponding eigenvalues [10]. Let R = V and S = L, we can de ine R as a rotation
matrix and S as a scaling matrix, allowing to express the covariance as in Equation 15. Based on this, the
rotation matrix and scaling matrix can be de ined as Equations 16 and 17, where θ is the rotation angle of
the speed vector and, sx and sy are the scaling factors in the x and y direction.

Σ = V LV −1 (14)

Σ = RSSR−1 (15)

[ ]
cos(θ) −sin(θ)
R= (16)
sin(θ) cos(θ)
[ ]
sx 0
S= (17)
0 sy
In order to ind the scaling factors, we take into account both the player’s magnitude of speed Si (t) (as
meters per second), and the distance to the ball Di (t). Based on the opinion of expert soccer analysts we
have de ined the range [4, 10] as the minimum and maximum distance in meters of player’s pitch control
surface radius Ri (t), based on the distance to the ball, following the transformation function shown at
Figure 9. Setting 13m/s as the maximum possible speed reachable, we calculate the ratio between players
and the maximum speed, as shown in Equation 18. Then, the scaling matrix is expanded in x direction and
contracted in y direction by this factor, as expressed in Equation 19. Given this, we can express a function
COV for obtaining the covariance matrix as shown in Equation 21. Finally, the distribution mean value
µi (t) is found by translating the players location at time t by half the magnitude of speed vector ⃗s, following
Equation 21.

s2
Srati (s) = (18)
132

2018 Research Papers Competition

18 Presented by:
 [ ]
Ri (t)+(Ri (t)Srati (⃗si (t)))
2 0
Si (t) = Ri (t)−(Ri (t)Srati (⃗si (t))) (19)
0 2

COVi (t) = R(θ, t)Si (t)Si (t)R(θi (t), t)−1 (20)

µi (t) = pi (t) + ⃗ŝi (t) · 0.5 (21)

Figure 9: Player in luence radius relation with distance to the ball

2018 Research Papers Competition

19 Presented by: