Passing to Win: Using Characteristics of Passing Information

for Match Winner Prediction

Taihu Li

1,2 a

, Jeewoo Yoon

1,2 b

, Daejin Choi

3,∗ c

and Jinyoung Han

1,2,∗ d

Department of Applied Artiﬁcial Intelligence, Sungkyunkwan University, Seoul, Korea

RaonData, Seoul, Korea

Department of Computer Science & Engineering, Incheon National University, Incheon, Korea

Keywords:

Machine Learning, Football, Pass Map, Match Winner Prediction.

Abstract:

Predictingthe football match results has received great attention both in sports industry and academic ﬁelds.

Many researchers have studied on predicting the match outcome using the simple features such as the number

of shots and passes. However, little attention has been paid to using pass interaction features, which can

represent how players in a match interact to each other. To this end, we propose a win-lose prediction model

that predicts a match result using the pass interaction and other features, achieving high accuracy of 79.5%. By

conducting an ablation study, we ﬁnd that the proposed interaction features play an important role in accurately

predicting match results. We believe our work can provide important insights both for industry and academic

researchers who want to understand the characteristics of winning teams.

1 INTRODUCTION

Recent advances in computing technology have

driven researchers to analyze diverse information of

football matches. Researchers can access not only

general match statistics such as match winner, num-

ber of shoots in a match, or total number of passes in

a match, but also in-game statistics of individual play-

ers (e.g., total running distances) and even interac-

tions among the players such as passing data between

two players (Linke et al., 2020; Pons et al., 2019).

Such comprehensive information can enable conduct-

ing an in-depth analysis on winning matches or in-

dividual playing performance (Johansen et al., 2013;

Bastida Castillo et al., 2018).

The abundant and comprehensive match data

has spurred researchers in industry and academia

to investigate match winners (Harrop and Nevill,

2014; Clemente et al., 2015) or winning strate-

gies (Georgievski et al., 2019), which has provided

valuable insights into understanding the key factors

to win a match. Harrop et. al revealed that increas-

https://orcid.org/0000-0002-2118-721X

https://orcid.org/0000-0002-9067-8653

https://orcid.org/0000-0001-5070-360X

https://orcid.org/0000-0002-8911-2791

∗

Corresponding Author

ing success rates of passes and shots and decreasing

the number of passes and dribbles are important to

win a match (Harrop and Nevill, 2014). Clemente et.

al built a players’ network and showed that the net-

works of winning teams tend to be dense (Clemente

et al., 2015). Georgievski et. al suggested that

the current rank of the team in a league need to

consider the degree of offensiveness/defensiveness of

teams (Georgievski et al., 2019).

In recent years, there have been much efforts on

predicting the match results or match winners, us-

ing simple statistical information of matches. For

example, Razali et. al proposed a machine learn-

ing model based on Bayesian Networks, which uses

match statistics like number of shots (Razali et al.,

2017). Pettersson et. al used a Long Short-term Mem-

ory (LSTM) model with the history of the match re-

sults of two teams to predict the match winner (Pet-

tersson and Nyquist, 2017). Hassan et. al proposed

an Artiﬁcial Neural Network (ANN) to predict the

match results using the data collected from TRACAB

that utilizes a beam-forming sensor & receiver equip-

ments (Hassan et al., 2020). However, although these

studies have provided valuable insights into under-

standing statistical features in predicting match win-

ners, little research has paid attention to model and

analyze how interactions among players through pass-

ing in a match can be used to predict match winners.

Li, T., Yoon, J., Choi, D. and Han, J.

Passing to Win: Using Characteristics of Passing Information for Match Winner Prediction.

DOI: 10.5220/0010659000003059

In Proceedings of the 9th International Conference on Sport Sciences Research and Technology Support (icSPORTS 2021), pages 54-60

ISBN: 978-989-758-539-5; ISSN: 2184-3201

In this paper, we propose a machine learning-

based prediction model for match winners, which

adopts both statistical match traits as well as interac-

tion patterns of two teams in a match. To this end, we

collected the match results and their associated data

including team statistics (e.g., possession rate), play-

ers’ individual statistics (e.g., number of ball steals),

and the pass matrices whose element is the pass-

ing counts between two players, from CHAMPION

DAtA

that logs all the football matches in Chinese

Super League and A League. Using the collected

dataset, we model the passing interaction as a directed

graph, called a pass map. Based on both the charac-

teristics of the pass maps (e.g., betweenness central-

ity in a graph) and the statistical information of two

teams in a match, the proposed model can predict the

match winner with 79.5% accuracy.

2 METHODS

We describe our methodology for developing a ma-

chine learning-based model to predict match win-

ners. In particular, we ﬁrst describe the data col-

lection method, e.g., crawling the match results with

comprehensive in-game statistics and passing infor-

mation. After modeling the passing patterns of each

team in a match as a pass map, we extract two feature

sets: (i) statistical features including in-game traits,

and (ii) interaction features with the characteristics

of the passing patterns. We then describe the pro-

posed machine learning-based model that can predict

the match result using the extracted features.

2.1 Data Collection

Figure 1: A screenshot of a pass matrix available on

CHAMPION DATA. Note that we anonymized the players’

names.

http://data.champdas.com/

We build the dataset to detect which team will win

the match by collecting the match information from

CHAMPION DATA, which provides the results of

the football matches in Chinese Super League and A

League with diverse match-relevant information in-

cluding statistical in-game traits both form the per-

spectives of teams and players, and the pass matrices

that represent how frequent players pass to each other.

An example of a pass matrix is illustrated in Figure 1.

That is, each row and column in a pass matrix M rep-

resents a player, and the element of the matrix at i-th

row and j-th column (i.e,, M

i j

) indicates the number

of passes from player i to player j. Note that two pass

matrices for each team are provided for a match. To

collect the match-related information from CHAM-

PION DATA, we developed a web crawler that fetches

the web pages including the match results with de-

tailed in-game information. Using the crawler, we

collected 2,682 match data from 75 teams. After ﬁl-

tering the tie match data, we ﬁnally gathered the 1,999

match results, 3,998 pass matrices, and statistical in-

formation of individual players and teams. Among

them, we use 80% as the training set of our model

ans use the remaining 20% for testing. Table 1 sum-

marizes in-game statistical information of individual

players and teams, respectively.

2.2 Pass Map Construction

Figure 2: An illustration of an example of a pass map of a

team in a match.

The passing information among team players has

been considered as an important factor to infer team’s

characteristics or even success in a match. Inspired by

this, we construct a pass map by using NetworkX

to characterize the passing patterns among the team

players in a match. That is, a pass map is deﬁned as a

directed graph G = (V, E, W ), where V and E are the

https://networkx.org/

Passing to Win: Using Characteristics of Passing Information for Match Winner Prediction

Table 1: The collected in-game statistical information of individual players and teams.

Team Stats. Player Stats.

# of shots, # of shots on target,

# of penalty kicks, # of free kicks,

# of front court free kicks, # of corner kicks

Possession rate, # of total passes,

Pass success rate, Dominance rate

# of key passes, # of cross passes,

# of break through, # of fouls obtained,

# of steals, # of intercepts, # of catches,

# of offside violations, # of clearance kicks,

# of pass blocks, # of shoot blocks,

# of yellow cards, # of red cards,

# of short passes, # of long passes,

Short pass rate (%), Long pass rate (%),

Direct pass rate (%), Cross pass rate (%),

Diagonal pass rate (%), Back pass rate (%)

sets of players (of a team) and passes among the play-

ers, respectively. Note that an edge e

i j

from node i and

node j exists when player i passes the ball to player j.

The weight of e

i j

is computed as the number of passes

from node i to node j. An example of a pass map of a

team in a match is illustrated in Figure 2.

2.3 Feature Extraction

From the collected dataset and the constructed pass

maps, we extract the features that are used to detect

which team wins the match. In particular, for a given

team, we compute two feature sets of the features, sta-

tistical and interaction features, described as follows:

• Statistical Features: We use 10 team and 21

player features, described in Table 1, as the statis-

tical features in a match. To compute the features

of a team from the players’ statistics, we calculate

the average values of the individual statistics of 11

players in the starting lineup.

• Interaction Features: We use the characteris-

tics of the pass maps of two teams in a match as

the interaction features. We use the NetworkX to

compute the following node features: in-degree,

out-degree, degree centrality (Tang et al., 2013),

closeness centrality (Bavelas, 1950), and betwee-

ness centrality (Freeman, 1977), each of which is

summarized in Table 2. Since the features are

computed from individual player’s perspective,

we simply compute average values and standard

deviations of each feature to generate a interac-

tion feature set of a team.

To compute the features of a match, we simply

concatenate two feature sets of each team (home and

away), which ﬁnally results in 92 statistical and 20 in-

teraction features for each match. See Table 3 in Ap-

pendix for all the listed features. The whole process

of feature extraction is illustrated in Figure 3.

Collected Dataset

Pass Matrices

Player Stats

Team Stats

Average,

Standard deviation

Statistical Features

Pass Map

Node Characterization

Interactional Features

Node Features

Average,

Standard deviation

Figure 3: An illustration of feature extraction process.

2.4 Match Winner Prediction Model

Statistical

Features

Interactional

Features

Statistical

Features

Interactional

Features

Home Team

Away Team

Win-Lose

Prediction Model

Win or Lose?

Figure 4: Overall architecture of the match winner predic-

tion model.

We deﬁne the prediction task that predicts which team

will win the match as a binary classiﬁcation prob-

lem. That is, we ﬁrst divide the given the constructed

dataset D = {(x

, y

)}

i=1

∈ R

, y

∈ {1...c}) with n

matches, m match features (i.e., statistical and inter-

action features), and c match result classes (i.e., pos-

itive when home team wins the match, negative oth-

erwise) into two datasets D

and D

test

, which repre-

sent the datasets for training and testing, respectively.

We then train our model using D

and ﬁnally predict

icSPORTS 2021 - 9th International Conference on Sport Sciences Research and Technology Support

Table 2: Five interaction features with their deﬁnitions and descriptions. g is the number of nodes and i is the index of the

nodes. x

i j

represents the total number of direct connections between N

and other g − 1 nodes, dis(i, j) indicates the distance

from node i to node j, and sd( j, i, k) means the shortest path from j to k passes through i.

Characteristics Deﬁnition Description

In-degree – How many players pass the ball to the given player

Out-degree – How many passes are concentrated to the player

Degree centrality C

) =

∑

j=1

i j

(i6= j)

g−1

How a given node plays a central role in connecting other nodes

Closeness centrality C

) =

g−1

∑

j=1

dis(i, j)

How many a given player receives direct passes from other players

Betweenness centrality C

) =

∑

j,k=1

sd( j,i,k)

∑

j,k=1

( j,k)

, ( j 6= k) How essential a given player is to connect small groups of the players

Precision Recall F1 Accuracy Precision Recall F1 Accuracy Precision Recall F1 Accuracy

0.625

0.728

0.673

0.558

0.765

0.8110.8120.812

0.795

0.838

0.848

0.828

INTER. ONLY STATS. ONLY STATS. + INTER.

Figure 5: Performance results of the match result prediction model. INTER. ONLY, STATS. ONLY, and STATS. + INTER.

denote interaction features only, statistical features only, and statistical + interaction features, respectively.

the match result class c based on the features (i.e.,

∈ R

) in D

test

To solve the classiﬁcation problem, we ﬁrst inves-

tigated popular machine-learning-based classiﬁers in-

cluding Support Vector Machines (Gunn et al., 1998),

Random Forest (Breiman, 2001), and eXtreme Gradi-

ent Boosting (XGBoost) (Chen and Guestrin, 2016).

After performance comparison, we select XGBoost as

the prediction model as it outperforms others. In the

experiment, we use the scikit-learn

and the XGBoost

python library

to conduct training and testing. The

formal deﬁnition of XGBoost is deﬁned as follows:

ˆy

= φ(x

) =

∑

k=1

) (1)

where ˆy

, f

, K, and f

) are the predicted class for

the i-th match, k-th independent tree, the number of

trees, and the prediction score given by the k-th inde-

pendent tree on the match features extracted from the

https://scikit-learn.org/

https://xgboost.readthedocs.io/

i-th match, respectively. The objective function of the

model, L(φ), can be calculated as follows:

L(φ) =

∑

l(y

, ˆy

) +

∑

Ω( f

) (2)

where l and

∑

Ω( f

) are the loss function between

the predicted class ˆy

and the target class y

and the

regularization term that penalizes the complexity of

the model, respectively. Here, we use squared er-

ror as the loss function. As a regularized term, we

use Ω( f ) = γT +

λ||w||

, where λ and γ controls the

penalty for the number of leaves T and magnitude of

the leaf weights w, respectively.

3 RESULTS

In this section, we report the performance results of

the prediction model for match winners. We then ex-

plore what features play important roles in the predic-

tion.

Passing to Win: Using Characteristics of Passing Information for Match Winner Prediction

3.1 Model Performance

Home win / Away lose Away win / Home lose

Home win / Away loseAway win / Home lose

Predict

Acutal

106

212

Figure 6: Confusion matrix of the model trained with both

team and interaction features.

# of shots on target (Home)

# of shots on target (Away)

# of total passes (Home)

# of total passes (Away)

Avg. of diagonal pass rate (Home)

Std. of # of cross pass (Away)

Avg. of direct pass rate (Away)

Dominance rate (Home)

Pass success rate (Home)

Std. of # of long pass (Home)

Std. of # of cross pass (Home)

Std. of # of clearance kick (Away)

# of team shots (Home)

Std. of # of long pass (Away)

Avg. of cross pass rate (Away)

Std. of # of clearance kick (Home)

Std. of closeness centrality (Away)

120

111

113

Std. of out-degree (Away)

Std. of degree centrality (Home)

Std. of in-degree (Home)

Statistical

Interactional

Feature Importance

Figure 7: Top 20 Important Features.

Figures 5 and 6 show the prediction result of the

proposed model. Here, we report four performance

metrics: (i) Precision (

T P

T P+FP

), (ii) Recall (

T P

T P+FN

(iii) F1 score (2 ×

precision×recall

precision+recall

), and (iv) Accuracy

(

T P+T N

T P+T N+FP+FN

) where TP, FP, FN, and TN represent

the true positive, false positive, false negative, and

true negative, respectively. In addition, we evaluate

the proposed model with three different feature sets:

(i) interaction, (ii) statistical, and (iii) both (i.e., in-

teraction + statistical). Overall, the model using both

feature sets outperforms all the other models. The ac-

curacy of the model is 79.5% while the ones of the

models solely using interaction or statistical features

are 55.8% and 76.5%, respectively. When we look

at the confusion matrix in Figure 6 that indicates the

number of the classiﬁed instances, the total number

of the instances correctly classiﬁed (upper-left and

lower-right) is 318 while the one at the other locations

is 82, meaning that the proposed model can predict

the match winners with high accuracy. Furthermore,

the performance of the model using both interaction

and statistical features is higher than other models us-

ing a single feature set (i.e., either interaction or statis-

tical), implying that interaction and statistical features

are complementary to each other.

3.2 Feature Importance

We further investigate what features play signiﬁcant

roles in predicting match winners by observing the

top 20 features in terms of the importance scores for

prediction calculated by the average gain across all

splits the feature is used in, as shown in Figure 7.

In general, the statistical features like the number of

shots, the number of total passes, and average diago-

nal pass rates or total passes are located at higher po-

sition, showing that statistical features are important

indicators in predicting match winners. Interestingly,

the standard deviations of closeness centrality, out-

degrees, degree centrality, and in-degrees are listed

in the top 20 features, which implies that whether all

the players in a match pass to each other with a similar

degree can be important predictors for match winners.

In other words, the passing interaction behavior in a

match is important for predicting match winners.

4 CONCLUDING DISCUSSION

In this paper, we proposed a machine learning model

that predicts the football match winners based on

the statistical and interaction features. We collected

the match results with their associated information

for 2,682 matches of Chinese Super League and A

League (2014-2020). By conducting an ablation

study, we revealed that the extracted interaction fea-

tures are complementary to statistical features.

There are a few limitations in our work. First, we

conducted experiments only on the football matches

in Chinese League, thus generalizing the methods and

the results in this paper to other leagues such as En-

icSPORTS 2021 - 9th International Conference on Sport Sciences Research and Technology Support

glish Premier League (EPL), Bundesliga, or LaLiga

should be cautiously considered. As a future work, we

plan to evaluate our proposed model to these leagues.

Second, we only considered eleven players in the

starting lineup, which has the rooms for improvement.

Despite the limitations, we believe our experimental

design and results can provide important insights for

both football industry and academic researchers who

want to lighten important characteristics of winning

teams.

ACKNOWLEDGEMENTS

This research was supported by the framework

of international cooperation program managed by

the National Research Foundation of Korea (NRF-

2020K2A9A2A11103842), and the MSIT(Ministry

of Science and ICT), Korea, under the ICT Creative

Consilience program (IITP-2021-2020-0-01821) su-

pervised by the IITP (Institute for Information &

communications Technology Planning & Evaluation).

REFERENCES

Bastida Castillo, A., G

omez Carmona, C. D., De la

Cruz S

anchez, E., and Pino Ortega, J. (2018). Ac-

curacy, intra-and inter-unit reliability, and comparison

between gps and uwb-based position-tracking systems

used for time–motion analyses in soccer. European

journal of sport science, 18(4):450–457.

Bavelas, A. (1950). Communication patterns in task-

oriented groups. The journal of the acoustical society

of America, 22(6):725–730.

Breiman, L. (2001). Random forests. Machine learning,

45(1):5–32.

Chen, T. and Guestrin, C. (2016). Xgboost: A scalable

tree boosting system. In Proceedings of the 22nd acm

sigkdd international conference on knowledge discov-

ery and data mining, pages 785–794.

Clemente, F. M., Martins, F. M. L., Kalamaras, D., Wong,

P. D., and Mendes, R. S. (2015). General network

analysis of national soccer teams in ﬁfa world cup

2014. International Journal of Performance Analysis

in Sport, 15(1):80–96.

Freeman, L. C. (1977). A set of measures of centrality based

on betweenness. Sociometry, pages 35–41.

Georgievski, B., Labadze, L., and Aboelsoud, M. E. (2019).

Comparative advantage as a success factor in football

clubs: Evidence from the english premier league (epl).

Journal of Human Sport and Exercise.

Gunn, S. R. et al. (1998). Support vector machines for

classiﬁcation and regression. ISIS technical report,

14(1):5–16.

Harrop, K. and Nevill, A. (2014). Performance indicators

that predict success in an english professional league

one soccer team. International Journal of Perfor-

mance Analysis in Sport, 14(3):907–920.

Hassan, A., Akl, A.-R., Hassan, I., and Sunderland, C.

(2020). Predicting wins, losses and attributes’ sen-

sitivities in the soccer world cup 2018 using neural

network analysis. Sensors, 20(11):3213.

Johansen, H. D., Pettersen, S. A., Halvorsen, P., and Jo-

hansen, D. (2013). Combining video and player

telemetry for evidence-based decisions in soccer. In

icSPORTS, pages 197–205.

Linke, D., Link, D., and Lames, M. (2020). Football-

speciﬁc validity of tracab’s optical video tracking sys-

tems. PloS one, 15(3):e0230179.

Pettersson, D. and Nyquist, R. (2017). Football match pre-

diction using deep learning. Psychology of Sport and

Exercise, 15(5):538–547.

Pons, E., Garc

ıa-Calvo, T., Resta, R., Blanco, H., L

opez del

Campo, R., D

ıaz Garc

ıa, J., and Pulido, J. J. (2019). A

comparison of a gps device and a multi-camera video

technology during ofﬁcial soccer matches: Agreement

between systems. PloS one, 14(8):e0220729.

Razali, N., Mustapha, A., Yatim, F. A., and Ab Aziz,

R. (2017). Predicting football matches results using

bayesian networks for english premier league (epl).

In Iop conference series: Materials science and en-

gineering, volume 226, page 012099. IOP Publishing.

Tang, X., Wang, J., Zhong, J., and Pan, Y. (2013). Pre-

dicting essential proteins based on weighted degree

centrality. IEEE/ACM Transactions on Computational

Biology and Bioinformatics, 11(2):407–418.

Passing to Win: Using Characteristics of Passing Information for Match Winner Prediction

APPENDIX

Table 3: A list of match features used for training the match winner prediction model.

index feature index feature

1 # of total penalty (Home) 57 Avg. of in-degree (Home)

2 # of total shots (Home) 58 Std. of in-degree (Home)

3 # of total shots on target (Home) 59 Avg. of out-degree (Home)

4 Possession rate (Home) 60 Std. of out-degree (Home)

5 # of total passes (Home) 61 Avg. of degree centrality (Home)

6 Pass success rate (Home) 62 Std. of degree centrality (Home)

7 Dominance rate (Home) 63 Avg. of closeness centrality (Home)

8 # of total free kick (Home) 64 Std. of closeness centrality (Home)

9 # of total frontcourt free kick (Home) 65 Avg. of betweenness centrality (Home)

10 # of total corner kick (Home) 66 Std. of betweenness centrality (Home)

11 # of total penalty (Away) 67 Avg. of # of catches (Away)

12 # of total shots (Away) 68 Avg. of # of key pass (Away)

13 # of total shots on target (Away) 69 Avg. of # of cross pass (Away)

14 Possession rate (Away) 70 Avg. of # of break through (Away)

15 # of total passes (Away) 71 Avg. of # of be fouled (Away)

16 Pass success rate (Away) 72 Avg. of # of offside (Away)

17 Dominance rate (Away) 73 Avg. of # of steal (Away)

18 # of total free kick (Away) 74 Avg. of # of intercept (Away)

19 # of total frontcourt free kick (Away) 75 Avg. of # of clearance kick (Away)

20 # of total corner kick (Away) 76 Avg. of # of block pass (Away)

21 Avg. of # of catches (Home) 77 Avg. of # of block shot (Away)

22 Avg. of # of key pass (Home) 78 Avg. of # of yellow card (Away)

23 Avg. of # of cross pass (Home) 79 Avg. of # of red card (Away)

24 Avg. of # of break through (Home) 80 Avg. of # of short pass (Away)

25 Avg. of # of be fouled (Home) 81 Avg. of # of long pass (Away)

26 Avg. of # of offside (Home) 82 Std. of # of catches (Away)

27 Avg. of # of steal (Home) 83 Std. of # of key pass (Away)

28 Avg. of # of intercept (Home) 84 Std. of # of cross pass (Away)

29 Avg. of # of clearance kick (Home) 85 Std. of # of break through (Away)

30 Avg. of # of block pass (Home) 86 Std. of # of be fouled (Away)

31 Avg. of # of block shot (Home) 87 Std. of # of offside (Away)

32 Avg. of # of yellow (Home) 88 Std. of # of steal (Away)

33 Avg. of # of red (Home) 89 Std. of # of intercept (Away)

34 Avg. of # of short pass (Home) 90 Std. of # of clearance kick (Away)

35 Avg. of # of long pass (Home) 91 Std. of # of block pass (Away)

36 Std. of # of catches (Home) 92 Std. of # of block shot (Away)

37 Std. of # of key pass (Home) 93 Std. of # of yellow (Away)

38 Std. of # of cross pass (Home) 94 Std. of # of red (Away)

39 Std. of # of break through (Home) 95 Std. of # of short pass (Away)

40 Std. of # of be fouled (Home) 96 Std. of # of long pass (Away)

41 Std. of # of offside (Home) 97 Avg. of short pass rate (Away)

42 Std. of # of steal (Home) 98 Avg. of long pass rate (Away)

43 Std. of # of intercept (Home) 99 Avg. of direct pass rate (Away)

44 Std. of # of clearance kick (Home) 100 Avg. of cross pass rate (Away)

45 Std. of # of block pass (Home) 101 Avg. of diagonal pass rate (Away)

46 Std. of # of block shot (Home) 102 Avg. of back pass rate (Away)

47 Std. of # of yellow (Home) 103 Avg. of in-degree (Away)

48 Std. of # of red (Home) 104 Std. of in-degree (Away)

49 Std. of # of short pass (Home) 105 Avg. of out-degree (Away)

50 Std. of # of long pass (Home) 106 Std. of out-degree (Away)

51 Avg. of short pass rate (Home) 107 Avg. of degree centrality (Away)

52 Avg. of long pass rate (Home) 108 Std. of degree centrality (Away)

53 Avg. of direct pass rate (Home) 109 Avg. of closeness centrality (Away)

54 Avg. of cross pass rate (Home) 110 Std. of closeness centrality (Away)

55 Avg. of diagonal pass rate (Home) 111 Avg. of betweenness centrality (Away)

56 Avg. of back pass rate (Home) 112 Std. of betweenness centrality (Away)

icSPORTS 2021 - 9th International Conference on Sport Sciences Research and Technology Support