UNIQUENESS PASSING MODEL 01.

INTRODUCTION

01. INTRODUCTION
You are watching a football match, possession is being circulated
around the middle in unspectacular fashion. Then all of a sudden,
your central midfielder (known for his vision) gets a moment of
time and hits a great pass. You know it's a great pass, it's
something different, yeah it's not normal, it's..... unique

I have attempted to capture this ‘uniqueness’ in a metric, in the

hope that it can become a reliable inclusion in player recruitment
shortlisting and analysis
UNIQUENESS PASSING MODEL 02. THE DATA

02. THE DATA

I will be using the recently released World Cup 2018 dataset from
Statsbomb. The dataset consists of over 63 games which totals to
174,418 events, of which 56,943 are passes.

Further Reading
Statsbomb Resource Centre
Sign-up to their public data user agreement, get access to the data &
get more info on the dataset
Accounts to Follow
StatsBombR Package
Ted Knutson Mike Goodman
The R package that will help you get your hands on that data super fast Founder / CEO Managing Editor
and super easy @mixedknuts @TheM_L_G

The Statsbomb Blog Thom Lawrence James Yorke

The epicentre of public analytics, great writing and includes examples Tummyball Chief Analyst and More
of how the free data could be utilised @deepxg @jair1970
UNIQUENESS PASSING MODEL
03. CLUSTER ANALYSIS
Clustering allows observations (rows) of data to be
compared across a dataset and to be grouped or clustered
together by their similarity.
The technique has been widely used due to ease of use, it’s
ability to improve comparative analysis and it’s intuitive
nature. It has it’s strengths, as well as it’s weaknesses.
For example, in the representation Player A
of multi-dimensional space, Player
A and Player B are close to each
other but it in different clusters.
Whilst Player B and Player C are Player B
more dissimilar but are clustered
Player C
together.
03. CLUSTER ANALYSIS
Further Reading
Introduction to Various Clustering Techniques
Introduction to K-means Clustering
Introduction to Principal Component Analysis
Identifying & Assessing Team Strategies
Will Gürpınar-Morgan
Use of k-means clustering to group
possession chains, in order to identify
team strategies
Square Pegs for Square Holes
Will Gürpınar-Morgan
Use of k-means clustering to identifying
player types to support an underpinning
recruitment strategy
The Evolution of the Full-Back
Mark Carey & Mladen Sormaz
Use of PCA and clustering to examine the
development of the full-back and group
into types
UNIQUENESS PASSING MODEL
04. CALCULATING DIFFERENCE
Leaning on the work of clustering, I will use the Nearest
Neighbour Search (NNS) function to calculate the ‘uniqueness’
of each pass
The NNS will calculate the dissimilarity of pass n with other
passes as the Euclidean distance within the multi-dimensional
space of the data
Instead of creating clusters, the ‘uniqueness’ metric will be the
sum of the distance of a certain number of passes. To make
this analysis scalable for more data, 1% of the total passes is
used as the comparison scope (569 passes in this case)
The higher the distance the more ‘unique’ the pass is deemed
04. CALCULATING DIFFERENCE
Each pass will be compared for it’s uniqueness
across the following characteristics:
• duration: the time the pass took from release to receive
• pass.length: distance from origin to destination
• pass.angle: the angle of the pass measured in radians
• pass.height.id: 3 grade variable for pass height (1) ground
pass (2) mid-height (3) high-ball
• x: the vertical location of the pass origin
• y: the horizontal location of the pass origin
• x.destination: the vertical location of the pass destination
• y.destination: the horizontal location of the pass
destination
UNIQUENESS PASSING MODEL 05. DOES IT MATCH?

05. DOES IT MATCH?

I grab a random selection of 10 passes to check the nearest Visualisation Explainer

neighbour search is doing it’s job and finding the 50 most similar Pass Origin - Circle denoting the origin of the pass

passes. Does it match? Yes, it’s doing a very good job at matching Pass Path - Line denoting the path of the pass

similar passes, so let’s delve a little deeper into the results Similar Passes - Lower Opacity for matched passes
UNIQUENESS PASSING MODEL 06. HISTOGRAM OF UNIQUENESS

06. HISTOGRAM OF UNIQUENESS

The histogram shows a Right-Skewed /
Positive Skewness distribution of uniqueness

Average Uniqueness 140.2

Quantiles
0 % 70.4
25 % 100.2

Frequency
50 % 119.4
75 % 168.8
100 % 608.9

The positive skew extends to the extreme of

608.9, this could mean that there are some
really unique passes that could offer a team
real value
Uniqueness
 UNIQUENESS PASSING MODEL 07. INITIAL DATA EXPLORATION

07. INITIAL DATA EXPLORATION

x Location of Pass Origin y Location of Pass Origin Pass Angle
Forwards
80
600
There is no Once again is no There is no
specific and clear specific and clear specific pattern
60
pattern other pattern other between pass
400
than a mild than a angle and
Uniqueness

y 40
curved dip. Uniqueness Left Side Right Side uniqueness aside
x location of the reaching high from a high
200 pass origin does 20
values in at the values being seen
not seem to be wings in forward passes
0 25 50 75 100 125
overly descriptive 0
200 400 600
x Uniqueness
Backwards

x Location of Pass Destination y Location of Pass Destination Pass Length

80
600
Once again is no Once again is no There is a distinct
specific and clear 60 specific and clear 600
pattern in pass
pattern other pattern other length with the

Uniqueness
400
than a mild y 40 than a longer the pass
Uniqueness

400
curved dip and a Uniqueness the higher the
distinct uplift in 20
reaching high uniqueness rating
200 very advanced values in at the 200

receives. 0
wings
200 400 600
0 25 100 125 0 30 60 90
50 75
xx Uniqueness Pass Length (m)
UNIQUENESS PASSING MODEL 08. LOOKING FOR PATTERNS

08. LOOKING FOR PATTERNS

x,y Location of Pass Origin x,y Location of Pass Destination
opacity of circle by Uniqueness Attacking Direction opacity of circle by Uniqueness Attacking Direction
1

2
2

1. An increase at the edges of the pitch 1. An increase at the edges of the pitch
2. An increase in the defensive box 2. An increase in the offensive box
3. An increase in wide areas around the box 3. An interest deep lying play-maker spot?
UNIQUENESS PASSING MODEL 09. WATCH THE VIDEO!

09. WATCH THE VIDEO!

Sergio Ramos Isco Simon Kjær Paulinho
v Russia v Russia v Croatia v Costa Rica
A simple slow pass out to Jordi Alba, A simple slow pass and short pass out A surprisingly unique pass, these A great pass in theory but as it overrun
this is the least unique pass in the to the wing. A continuation pass, situations are rarer than I imagined it became highly unique, in fact the
whole dataset which is surprisingly common and have a low conversion rate most unique pass in the dataset
Uniqueness = 70.5 Uniqueness = 74.0 Uniqueness = 330.4 Uniqueness = 580.9
UNIQUENESS PASSING MODEL
10. DON’T WATCH THE VIDEO!
In my hypothesis, the more
unique a pass is the better is it.
Paulinho’s pass vrs. Costa Rica
made me revaluate the passes
with the highest uniqueness
rating.
I am starting to have a feeling
that the higher values are in fact
so random that they score highly
but are in fact not of much use
and a bad decision to attribute to
a player as a positive.
Let’s have a look at some other
passes with high uniqueness
rating
10. DON’T WATCH THE VIDEO!
Carvajal
v Morocco
A cross field ball, awkward to receive
by Pique which resulted in the ball
running on further. Certainly not an
interesting pass!
Uniqueness = 529.6
Di Maria
v Nigeria
A very strange cross field volley under
slight pressure. Possible intent to play
to central defenders but overhit. Again,
not an interesting pass!
Uniqueness = 533.00
UNIQUENESS PASSING MODEL
11. A METRIC CRITIQUE
1. The passes with the highest scores
of uniqueness are very random. By
leaving these passes in for further
analysis players and teams will be
rewarded.
Having studied a vast number of these
passes it seems that the top 0.25% of
passes are totally random and can be
screened out.
The top 0.25 - 1.00% does still include
some random passes but also includes
those with real quality. These will be
left in the dataset, yet I must retain the
understanding that this is a weakness
of the model.
11. A METRIC CRITIQUE
2. There’s high prevalence 3. All passes have a unique score. If we
of passes in the defensive sum the scores in our further analysis
half that rank highly on the player making 10 very boring and
uniqueness scale. A pass simple passes would score higher than
to a full-back who dropped a player making 1 truly amazing pass.
deep & wide is rare and
high in uniqueness Having studied the video of a vast
number of these passes it seems that
Although these passes are indeed unique the top 25% of the unique attacking
they didn’t inspire the thinking behind the passes seem to be those that the
metric! When I run further analysis on the metric was intended to identify.
results defensive players dominate all of
the shortlists. I will therefore filter the Therefore I keep just these passes in
passes, to create a subset of unique the classification of unique attacking
attacking passes. passes — leaving us with 5,410 passes
x > 60 & x.destination > x or x > 80 & x.destination > 60
 UNIQUENESS PASSING MODEL 12. SHORTLISTING ‘UNIQUENESS’

12. SHORTLISTING ‘UNIQUENESS’

Shortlisting provides us with some insights via an eye-ball test of who PLAYER TEAM CLUB UPA PER 90
pops up on the lists. Defenders

Jérôme Boateng Germany Bayern Munich 13.7

These shortlists would become much more accurate after 20+ games.
Joshua Kimmich Germany Bayern Munich 11.5
I have applied a filter of >10 UPA & >180 total minutes played.
Sergio Ramos Spain Real Madrid 7.7

The shortlists largely pass the eye-test. Toni Kroos tops the pile, high Yasir Al Shahrani Saudi Arabia Al-Hilal Riyadh 7.6
possession teams like Germany and Spain show up and and Red Bull Marcelo Brazil Real Madrid 6.6
Leipzig features twice. Midfielders

Toni Kroos Germany Real Madrid 12.1

Yasir Al Shahrani and Tarek
Hamed are names that I don’t have Isco Spain Real Madrid 9.2

good knowledge of. Yet when Mesut Özil Germany Arsenal FC 8.9
reviewing their footage I can Tarek Hamed Egypt Zamalek SC 8.6
understand why they are scoring João Mário Portugal Inter Milan 8.2
highly. For example ——> Forwards

Artem Dzyuba Russia Zenit St. Petersburg 11.4

UPA PER 90: Overall Distribution Timo Werner Germany Red Bull Leipzig 9.0
Yussuf Poulsen Denmark Red Bull Leipzig 8.3
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 Piotr Zieliński Polamd Napoli 8.1
Thomas Müller Germany Bayern Munich 7.9
Messi De Bruyne Henderson Kroos Boateng
UNIQUENESS PASSING MODEL 13. BEST XI

13. BEST XI
06. BEST XI

NK

3
P
RA

TO
OP

OM
-T

FR
XI

K
IC
T
ES

-P
-B

XI
S

T
SE
Another way of eye-balling the

ES
S

-B
PA
ISCO ISCO

S
NG

SE
findings to is form a Best XI from DZYUBA WERNER

KI

S
11 11

PA
AC
T

NG
AT
9 9

the players scoring highest per

KI
UE

AC
IQ

T
UN

AT
position or in the top 3 per

UE
ÖZIL ÖZIL

IQ
UN
position (to account for
10 10

tight calls) SHANRANI

3
T. HAMED
6
POULSEN
7
MARCELO
3
KROOS
6
ZIELIŃSKI
7

BORGES MODRIĆ
8 8

S. RAMOS S. RAMOS
Although the Top 4 4

RankXI has some BOATENG BOATENG

KIMMICH KIMMICH
players that you would PICKFORD 5
2
PICKFORD 5
2

change, picking from the 1 1

top 3 players in each position

gives you a truly world class team
full of players deemed as the best
in their position by many
UNIQUENESS PASSING MODEL 14. EXPECTED COMPLETIONS

14. EXPECTED COMPLETIONS

The natural extension of identifying the Pass Completion Probability
by Uniqueness - All Passes
uniqueness is to see if the metric or the
process can help build a more precise pass There is a clear dip in

completion analysis.
completion probability
within the mid 200s to
a low of 0.47

Pass completion stats are often see as

largely irrelevant but they are intuitive so if
the stat can be improved it’s a big bonus. Pass Completion Probability
by Uniqueness - Unique Attacking Passes

Let’s quickly investigate how the A slow rise in

completion probability

uniqueness of a pass impacts on pass towards the upper

end of the uniqueness
range
completion rates
UNIQUENESS PASSING MODEL 15. REGRESSION TO THE MEAN

15. REGRESSION TO THE MEAN

There isn’t enough detail to the pass completion rates in
relation to the uniqueness score.
Adapting the NNS process may give us better success
1. Create a new variable of success, completed pass = 1 &
incomplete = 0
2. For each pass find the 50 most similar passes
This could provides a more precise way of assessing if a player
passes above expectation. Before shortlisting by xCompletion
Difference, it’s important to explain why I will filter out all players
that have not played above a 20 unique attacking passes. Players
with fewer passes can appear extremely low or high, with a higher
number of passes most players regress closer to the mean.
20
Key

3. Calculate the pass completion rate for those passes as

0 - 10 Unique Attacking Passes

10 - 20 Unique Attacking Passes

expected completion 15
20 - 30 Unique Attacking Passes

4. Calculate the xCompletion Difference = Success - expected 10

30 - 40 Unique Attacking Passes

completion 40 - 50 Unique Attacking Passes

5. Average the xCompletion Difference per player for all UAPs 5

0
-40% 0% -40% -80%
xCompletion.Rate
UNIQUENESS PASSING MODEL 16. xCOMPLETION RATES - PER POSITION

16. xCOMPLETION RATES - PER POSITION

DEFENDERS MIDFIELDERS FORWARDS
PLAYER TEAM xCOMP.RATE PLAYER TEAM xCOMP.RATE PLAYER TEAM xCOMP.RATE
Toni Kroos Germany 14.9 % Luis Suárez
Thomas Meunier Belgium 12.2 % Uruguay 5.7 %
Luka Modrić Croatia 12.4 %
Olivier Giroud France 4.5 %
Nahitan Nández Uruguay 10.2 %
Harry Maguire England 10.9 %
Lucas Torreira Uruguay 9% Neymar Brazil 3.9 %
Jérôme Boateng Germany 10 % William Carvalho Portugal 8.7 % Piotr Zieliński Poland 0.7 %
Yury Gazinskiy Russia 7.8 %
Eden Hazard Belgium 0.1 %
Raphaël Guerreiro Portugal 8.6 % Thomas Delaney Denmark 5.8 %
Sergej Milinković-Savić Serbia 5.3 % Artem Dzyuba Russia -1.1 %
Mário Fernandes Russia 6.4 % Dries Mertens Belgium 5% Dusan Tadic Netherlands -1.4 %
Jesse Lingard England 4.3 %
Sergio Ramos Spain 6.2 % Marcus Berg Sweden -4.3 %
Javier Mascherano Argentina 3.8 %
Philippe Coutinho Brazil 3.6 % Bryan Ruiz Costa Rica -5 %
Jordi Alba Spain 1.2 %
Granit Xhaka Switzerland 3.5 % Yussuf Poulsen Denmark -6.4 %
Martín Cáceres Uruguay 0% Christian Eriksen Denmark 3%
Aleksandr Samedov Russia -7.1 %
Andrés Iniesta Spain 3%
Joshua Kimmich Germany -1.1 % Ivan Rakitić Croatia 2.7 % Kylian Mbappé France -8 %
João Mário Portugal 1.7 % Ola Toivonen Sweden -8.5 %
Marcelo Brazil -6 % Aleksandr Golovin Russia 1.4 %
Aleksandar Mitrovic Serbia -8.9 %
Paul Pogba France 1.3 %
Benjamin Pavard France -6.6 % Jordan Henderson Timo Werner Germany -9.2 %
England 0.9 %
Isco Spain 0.1 %
Lucas Hernández France -7.3 % Lionel Messi Argentina -9.2 %
Viktor Claesson Netherlands -0.3 %
Thomas Müller Germany -9.2 %
Kieran Trippier England -8.6 % Celso Borges Mora Costa Rica -1.1 %
Héctor Herrera Mexico -1.1 % Ivan Perišić Croatia -10.3 %
Yasir Al Shahrani Saudi Arabia -9 % Salman Al Faraj Saudi Arabia -1.2 % Mario Mandžukić Croatia -11.7 %
Sergio Busquets Spain -2.2 %
Hörður Magnússon Iceland -10.6 % Harry Kane England -14.1 %
Kevin De Bruyne Belgium -5.3 %
N'Golo Kanté France -5.8 % Xherdan Shaqiri Switzerland -15 %
Ludwig Augustinsson Sweden -11.7 % Roman Zobnin Russa -6.4 % M'Baye Babacar Niang Senegal -18.4 %
Marcelo Brozović Croatia -7.5 %
Šime Vrsaljko Croatia -12.2 % Ante Rebić Croatia -22.2 %
Andrej Kramarić Croatia -7.9 %

Henrik Dalsgaard Denmark -12.5 % Dele Alli England -16.6 % Antoine Griezmann France -22.3 %
UNIQUENESS PASSING MODEL 17. PLOTTING THE RESULTS - DEFENDERS

17. PLOTTING THE RESULTS - DEFENDERS

PLAYER TEAM xCOMP.RATE

Thomas Meunier Belgium 12.2 %

Harry Maguire England 10.9 % 15

Total Number of Unique Attacking Passes

AVG. XCOMP DIFF

Jérôme Boateng Germany 10 % 20-30
30-40 JÉRÔME BOATENG
Raphaël Guerreiro Portugal 8.6 %
40-50

Mário Fernandes Russia 6.4 % 12.5

Sergio Ramos Spain 6.2 % KIMMICH

Jordi Alba Spain 1.2 %

10
Martín Cáceres Uruguay 0%

Joshua Kimmich Germany -1.1 %

Marcelo Brazil -6 % 7.5 AL SHAHRANI SERGIO

RAMOS
Benjamin Pavard France -6.6 %
MAGNÛSSON MARCELO GUERREIRO AVERAGE UPA.90
CÁCERES
Lucas Hernández France -7.3 % VRSALIJKO TRIPPIER MÁRIO
FERNANDES
Kieran Trippier England -8.6 % 5
AUGUSTINSSON MEUNIER
JORDI ALBA
Yasir Al Shahrani Saudi Arabia -9 % DALSGAARD PAVARD
MAGUIRE

Hörður Magnússon Iceland -10.6 % LUCAS HERNÁNDEZ

2.5
Ludwig Augustinsson Sweden -11.7 % -25% -10% 0% 5% 20%

Šime Vrsaljko Croatia -12.2 %

Henrik Dalsgaard Denmark -12.5 %

UNIQUENESS PASSING MODEL 18. PLOTTING THE RESULTS - MIDFIELDERS

18. PLOTTING THE RESULTS - MIDFIELDERS

PLAYER TEAM xCOMP.RATE
Toni Kroos Germany 14.9 %
Luka Modrić Croatia 12.4 %
Nahitan Nández Uruguay 10.2 %
15
Lucas Torreira Uruguay 9% Total Number of Unique Attacking Passes

AVG. XCOMP DIFF

William Carvalho Portugal 8.7 % 20-30
Yury Gazinskiy Russia 7.8 %
30-40
Thomas Delaney Denmark 5.8 %
40-50
Sergej Milinković-Savić Serbia 5.3 %
12.5
Dries Mertens Belgium 5%
KROOS
Jesse Lingard England 4.3 %
Javier Mascherano Argentina 3.8 %
Philippe Coutinho Brazil 3.6 %
Granit Xhaka Switzerland 3.5 % 10
Christian Eriksen Denmark 3%
Andrés Iniesta Spain 3% INIESTA MASCHERANO
ISCO
Ivan Rakitić
AL FARAJ
Croatia 2.7 %
C.BORGES
João Mário Portugal 1.7 % J.MÁRIO COUTINHO
Aleksandr Golovin Russia 1.4 % 7.5 N.NÁNDEZ
RAKITIĆ MILINKOVIĆ-
Paul Pogba France 1.3 % KOVIĆ
HENDERSON
Jordan Henderson England 0.9 % AVERAGE UPA.90
POGBA GAZINSKIY
DELE ALLI XHAKA
Isco Spain 0.1 % DE BRUYNE MODRIĆ
GOLOVIN
Viktor Claesson Netherlands -0.3 % TORREIRA
KRAMARIĆ ZOBNIN
Celso Borges Mora Costa Rica -1.1 % 5 BUSQUETS
W.CARVALHO
BROZOVIČ H.HERRERA ERIKSEN
Héctor Herrera Mexico -1.1 % DELANEY
LINGAARD
Salman Al Faraj Saudi Arabia -1.2 % CLAESSON
MERTENS
Sergio Busquets Spain -2.2 % KANTÉ

Kevin De Bruyne Belgium -5.3 %

2.5
N'Golo Kanté France -5.8 %
-25% -10% 0% 5% 20%
Roman Zobnin Russa -6.4 %
Marcelo Brozović Croatia -7.5 %
Andrej Kramarić Croatia -7.9 %
Dele Alli England -16.6 %
UNIQUENESS PASSING MODEL 19. PLOTTING THE RESULTS - MIDFIELDERS

19. PLOTTING THE RESULTS - MIDFIELDERS

PLAYER TEAM xCOMP.RATE
Luis Suárez Uruguay 5.7 %
Olivier Giroud France 4.5 %
Neymar Brazil 3.9 % 15
Total Number of Unique Attacking Passes

AVG. XCOMP DIFF

Piotr Zieliński Poland 0.7 % 20-30

Eden Hazard Belgium 0.1 % 30-40

40-50
Artem Dzyuba Russia -1.1 %
12.5
Dusan Tadic Netherlands -1.4 %
Marcus Berg Sweden -4.3 %
DZYUBA
Bryan Ruiz Costa Rica -5 %
Yussuf Poulsen Denmark -6.4 % 10
Aleksandr Samedov Russia -7.1 %
Kylian Mbappé France -8 % WERNER
T.MÜLLER
Ola Toivonen Sweden -8.5 % GRIEZMANN Y.POULSEN
ZIELIŃSKI
7.5 LUIS SUÁREZ
TOIVONEN TADIC
Aleksandar Mitrovic Serbia -8.9 % NIANG MITROVIC
KANE AVERAGE UPA.90
Timo Werner Germany -9.2 % MANDŽUKIČ GIROUD
PERIŠIČ
RUIZ
Lionel Messi Argentina -9.2 % SAMEDOV
REBIC HAZARD
SHAQIRI
Thomas Müller NEYMAR
Germany -9.2 % 5
MESSI
Ivan Perišić Croatia -10.3 % MARCUS
BERG
Mario Mandžukić Croatia -11.7 %
MBAPPÉ
Harry Kane England -14.1 %
Xherdan Shaqiri Switzerland -15 % 2.5
-25% -10% 0% 5% 20%
M'Baye Babacar Niang Senegal -18.4 %
Ante Rebić Croatia -22.2 %
Antoine Griezmann France -22.3 %
UNIQUENESS PASSING MODEL 20. xG & Uniqueness

20. xG & Uniqueness

The idea in my head of a
‘unique’ pass at the start of this
investigation was a creative, There does seem to be a
defence splitting and shot mild positive correlation.
creating pass. However, the relationship
was not statistically
Let’s analyse the relationship significant. With more data
between uniqueness and these conclusions could be
expected goal value of resulting tested with greater clarity.
shots. Find all shots within 5
seconds of a unique attacking
pass and sum the expected
goals value.
UNIQUENESS PASSING MODEL
21. FINAL THOUGHTS
1. The Uniqueness Passing Model
seems to help with shortlisting
and player identification for
further assessment. However, the
same could be said about passes
in the final third per 90 (and there
is minimal computational
requirements!). It ‘works’ but it’s
pretty clunky, & sellotape together
2. I don’t have enough data to
thoroughly test if UAPs is
replicable and predictive to shots,
expected goals or goals
themselves. I would be interested
to investigate this further if
(when) Statsbomb release more
data to play with.
21. FINAL THOUGHTS
3. I hope the working example of the
NNS will prompt further
investigation into how it can be
used more extensively within
public football analytics.
4. The Statsbomb dataset is
amazing, well designed and a real
gift to the community. We are so
lucky to have access to it! So
utilise it to death!