Detection of Possible Match-fixing in Tennis Games

Kim Yong-wook

, Jinyoung Han

and Choi Seoung-rak

Department of Sports Science, Hanyang University, Ansan, South Korea

Department of HCI, Hanyang University, Ansan, South Korea

Keywords: Match-fixing, Tennis, Benford’s Law.

Abstract: This study seeks ways that can immediately detect match-fixing in a tennis match. We first explore the number

of rallies observed in tennis matches of the ATP and WTA leagues to determine whether they follow the

Benford’s law. We also artificially manipulate practice games to investigate whether the number of ralleys

observed in manipulated matches also follows the Benford’s law. Experimental results demonstrate that the

numbers collected from fixed games and the expected frequencies predicted by Benford’s Law are different.

Based on the lessons learned, we develop a machine-learning-based model for detecting whether a given

match is fixed or not. Our model shows a high accuracy in detecting fixed tennis matches, which has a great

utility for fair tennis play.

1 INTRODUCTION

Sports must always conform to the ethics of fair play.

Particularly when contending for victory, there is an

underlying principle that athletes must follow fair

rules in a righteous way and strive to show their best

performance. However, in a world of professional

sports where there are conflicts of vast interests,

dramas continue to unfold, sliding back and forth

between legal and illegal paths. There is a clear limit

to the fundamental roles that professional athletes are

basically paid to play. Furthermore, athletes fall into

temptation of illegal match fixing each time their

desire for fame and wealth combines with the greed

of sports clubs (European Commission, 2018; Lastraa

et al., 2018; Asser Institute, 2014).

In a fair competition, under the fair play system,

outcomes cannot be predicted with an absolute

certainty. However, an act of artificial manipulation

of the game, such as a player's slowdown, the bribery

of referees, or match-fixing by referees, can certainly

change the outcome of a game. Match-fixing is a

process of pre-determining results that might provide

a certain opportunity for avid gamblers to win bets in

sports gambling (Moriconi, 2018). It can be a serious

act that weakens the foundation of sports. Information

Commissioner’s Office (ICO) and The Federation

Internationale de Football Association (FIFA) have

acknowledged “prevention of match-fixing” as the

most critical sports issue to be addressed in the 21st

century and have taken various preventive measures

(IOC, 2016; FIFA, 2017; Aquilina and Chetcuti,

2014).

Match-fixing occurs extensively in a wide variety

of professional sports leagues, ICO-sponsored

Olympics, FIFA-sponsored World Cups, as well as

the world championship games organized by various

sports federations. The European Union (EU)

Commission has reported more than 5,200 match-

fixing cases worldwide since the year 2000 (European

Commission, 2014; European Commission, 2018).

Since 2000, match-fixing cases have been most

commonly found in the European league football

matches. However, the cases are expected to rise if

the leagues that the EU does not monitor, e.g., Asian

leagues and the American leagues, are also included

(European Commission, 2012; Katsarova, 2016). In

particular, leagues operating in the Asian region have

been under relatively less scrutiny than the European

leagues, making them a softer target for match fixers

(Hill, 2010). A match-fixing scandal that surfaced in

Korea’s professional football league in 2011 caused

two footballers to commit suicide, 10 to be

permanently suspended, and 59 to receive criminal

penalties (THE KOREA TIMES, 2012). This incident

was perpetrated by an overseas criminal organization

that managed to bribe footballers with as little as USD

5,000. It delivered a devastating wound to the Korean

football league that ended the lives and professional

careers of many of its players forever. Even to this

124

Yong-wook, K., Han, J. and Seoung-rak, C.

Detection of Possible Match-ﬁxing in Tennis Games.

DOI: 10.5220/0006924201240131

In Proceedings of the 6th International Congress on Sport Sciences Research and Technology Support (icSPORTS 2018), pages 124-131

ISBN: 978-989-758-325-4

date since the conclusion of the investigations, the

true identity and nature of the overseas crime group

that was involved in the incident have not been

revealed. There is a very high possibility that the

criminal group is attempting to fix another game at

this moment.

Match-fixing is very common in tennis leagues. It

has been reported that match-fixing was attempted at

the 2016 Wimbledon Championship, one of the most

prestigious tennis tournaments. Novak Djokovic,

who ranked first in the world in 2016, revealed in an

interview that he was once offered a USD 200,000

monetary reward. At the same time, he also received

physical threats from a group that was trying to fix a

match (Huffington post, 2016). This happens owing

to a structural problem with regard to the financial

reality of professional tennis, in which players

heavily rely on the prize money and monetary support

from their sponsors to make a living.

According to Michael Russell’s interview with

Forbes, the 92nd ranked player in the world in 2013,

he earned USD 270,000 in revenue in 2012, but his

actual income was USD 85,000 after a deduction of

personal and tour expenses that amounted to USD

75,000 and 35,000 respectively, while taxes resulted

in the deduction of another significant portion of his

effective earnings. He further disclosed that although

he made a total of USD 2,100,000 in a 15-year-long

career, there is not much left after his tour expenses

and fee for coaches and trainers (Forbes, 2013). This

means that the lower ranked players with less income

in the current world of professional tennis are more

prone to the temptation of match-fixing.

As match-fixing has become a serious social issue,

a series of related researches have been carried out on

match-fixing, such as studies that examine the match-

fixing cases and explain their cause (Hill, 2013;

Carpenter, 2012) and those that analyze the relevance

of match-fixing with the sports betting industry (Bag

and Saha, 2011; Boeri and Severgnini, 2011). In

addition, there are studies that offer legal

countermeasures to eradicate match-fixing

(Rodenberg and Feustel, 2014) as well as attempts to

detect match-fixing via a mathematical approach

(Hill, 2011).

To completely stamp out match-fixing from the

world of sports, the introduction of preventive

measures along with ethics education and training for

the athletes is essential. It is also critical to detect one

in time, on occurrence. However, there are numerous

variables to be considered in a sport like tennis,

unexpected outcomes often arise depending on the

condition of the athletes playing. Therefore, when a

player or coach, who is artificially defeated in a fixed

match, and claims later that “the player was in a

tattered physical condition and hence unable to

perform his best”, on-site investigation no longer

remains a feasible option.

Currently, the best option is to conduct a post-

incident investigation by anti-corruption

organizations such as the Tennis Integrity Unit over

match-fixing suspicions in tennis tournaments.

However, such inquiries require conclusive evidence

to determine culpability, such as evidence of

monetary transactions surrounding suspected athletes;

moreover, only the authorities with jurisdiction to

conduct criminal investigation, such as the

prosecution or the police, have access to them. As a

result, organizations such as the Tennis Integrity Unit,

which lack the authority, begin with probing into

suspicious events. Once sufficient evidence is

gathered to support the allegations, they request

international investigative authorities to open official

investigations into the alleged match-fixing incidents

(TIU, 2018).

This approach investigates players belonging to

different nationalities under the jurisdiction of the

country where the match-fixing is suspected to have

occurred, and it inevitably accompanies highly

complicated and cumbersome administrative

procedures. Therefore, if clear evidence such as

testimonies of fraudulent deals or any circumstantial

evidence regarding illegal monetary transactions is

not secured, ongoing investigations often get

suspended. As discussed earlier, a post-match

investigation into a match-fixing incident requires

tremendous effort to gather sufficient evidence to

support a finding of actual unlawful transactions,

making it highly difficult to identify match-fixing in

reality (HM Government, 2017). Consequently, the

current method of revealing fixed matches and

imposing severe penalties on the conspirators is not

enough to eradicate the problem of match-fixing.

Then how do we prevent match-fixing? The best

approach for now would be to hinder the match-fixing

attempts in advance. The following two types of

preventive measures are currently being implemented

at the sports scenes. First, we can encourage the

athletes to be aware of the seriousness of match-

fixing by continuously providing them with ethics

education and training (Department for Culture,

Media, and Sport: UK, 2010). Second, if pre-game

signs of possible match-fixing appear, such as a

sudden change in the dividend rates offered by sports

betting sites, it would be prudent to prevent attempts

of match-fixing by informing the players that the

impending match is most likely, a fixed match. These

methods use trainings and warnings to discourage the

Detection of Possible Match-ﬁxing in Tennis Games

125

players from attempting to fix a match and they are

already being used by IOC, professional leagues, and

various sports federations (IOC, 2018; FIFA, 2018),

However, if athletes continue with their attempts

to fix matches despite the measures in force, there are

almost no additional preventive measures available

that can be applied at the scene. Even if there is a

method to immediately detect match-fixing at a game,

how would that prevent such incidents from occurring?

In that scenario, the athletes will be alerted that a

match-fixing attempt will most likely be exposed and

they will inevitably reduce such attempts. Therefore,

this study introduces a method to detect match-fixing

immediately in the field of tennis.

2 RESEARCH METHOD

The research method applied to this study for the

detection of match-fixing is briefly explained as

follows. First, we observe the number of rallies

overserved in matches of the ATP and WTA leagues

to determine whether they follow the Benford’s law.

Second, we artificially manipulate practice games to

collect the data on the number of rallies in each match

and verify whether the distribution follows the

Benford’s law. Lastly, we develop a machine-

learning-based model to detect whether a given match

is fixed or not. For leaning and testing the model, A-

Set (training set), which is the number of rallies

recorded from ATP and WTA, and B1-Set (validation

set) and B2-Set (test set) that reflect the data collected

from two artificially manipulated matches are used.

2.1 Benford’s Law

Benford’s Law is an observation that numerical

values that can be observed in our daily lives appear

in accordance with certain rules. When we look at

datasets such as the population numbers, death rates,

passwords, and lengths of rivers, the probability of

the first digit being number 1 is approximately 31%,

while the numbers 5 and 9 appear as the most

significant digits 8%, and 5% of the time, respectively.

This shows that the lower numbers are observed more

frequently than the higher numbers (Figure 1).

Benford’s Law was first discovered in 1881 by an

American mathematician, Simon Newcomb.

Newcomb noticed that the earlier pages in logarithm

tables were much more worn than the other pages,

and realized that the smaller digits were more likely

to appear than the larger digits as naturally occurring

real-life numbers. It was an empirically interesting

discovery, but it lacked mathematical and logical

Figure 1: The distribution of first digits according to

Benford's law.

explanation, and was not accepted as a law. It was

again noted in 1938 by an American physicist Frank

Benford, who observed a phenomenon that supported

Newcomb's claim in naturally occurring collections

of data. He analyzed 20 unrelated domains, such as

river lengths, population numbers, and also a number

of magazines and as a result, the probability of the

first digit being number 1 appeared as 31%, while

19% began with number 2. He measured the

probabilities for the occurrence of the digits ranging

from 1 to 9 and announced the results. This

observational phenomenon was later named after him

as the Benford law (Nigrini, 2012).

The Benford’s Law reveals the probability

distribution of naturally occurring numbers, and as

artificial numbers do not follow this law, we can

identify human intervention by verifying the numbers

observed. For example, numbers that reflect people’s

thoughts and purposes, such as the phone numbers,

postal codes, and the price of goods, do not follow the

Benford’s law. Therefore, we believe the law can be

applied in this study. We conjecture that the numbers

observed in matches played with a specific intention,

such as match-fixing, might differ from the data

recorded at the games that were played to win.

2.2 Machine Learning

Machine learning algorithms are classified into

supervised learning, unsupervised learning, and

reinforcement learning. Supervised learning utilizes

input and labeled data and accurate results of input

data are achieved using the training data. A program

icSPORTS 2018 - 6th International Congress on Sport Sciences Research and Technology Support

126

is generated based on the analysis of these patterns,

enabling it to predict the results of the incoming input

data. Figure 2 shows the process of analyzing the

training data and creating the evaluation model to

verify the test data. Here, the extract features are the

data format expressed and processed for the algorithm

to judge. Machine learning algorithm generates a

model that predicts an output based on an input. In the

end, the output created based on the input can be

predicted as a result of the learning carried out by the

machine learning algorithm. In this study, an analysis

of the deep learning method of supervised learning is

used.

2.3 Research Data

This study used a number of rallies observed in tennis

matches for analysis. A rally can be seen in a tennis

match where a ball is played with a net, such as a

tennis ball, and it refers to a process where the ball

passes over to the opposing side of the court and then

returns. In other words, the number of rallies indicates

the number of times the ball has passed over to the

other side of the court and one round trip is referred

to as one rally. However, in this study, we used the

number of times that the opponent passes the ball for

a more accurate analysis and named it as h-rally. The

h-rally data is collected by implementing the

following two methods.

2.3.1 Data Collection from ATP and WTA

Leagues

The matches studied for the data collection were the

tennis matches held in the ATP and WTA leagues

from January 2016 to March 2018. We selected and

collected data from 220 ATP and WTA matches that

were broadcasted television or were available online.

The data was recorded by researchers who watched

every selected game, and the h-rally that occurs for

each score was directly observed and collected. The

dataset collected here is referred to as “Dataset A”.

2.3.2 Data Collection from Fixing Game

We artificially manipulated practice games by asking

one of the players with similar skills to lose the match,

and we measured the outcomes. The experiments

were conducted twice on H and S tennis courts in

Ansan, South Korea on April 21 and May 13, 2018.

Four players, C, K, J, and P participated in twelve

games in total, six times on each day. We made a prior

arrangement with player C to artificially lose the

game while players K, J, and P remained unaware of

player C’s intention to fix the game throughout the

entire experiment. In the end, h-rallies recorded from

all six matches that player C participated were

collected as a dataset for fixed games. The datasets

collected on the first and second days of experiments

were labeled as “Dataset B1” and “Dataset B2”,

respectively.

Figure 2: Supervised Learning workflow.

Detection of Possible Match-ﬁxing in Tennis Games

127

2.4 Analysis Method

The collected data was analyzed as follows. First, we

verified whether the data by applying x

test statistics

followed by the Benford’s law. Second, we used

“Dataset A” collected from ATP and WTA as a

training set and “Dataset B1” that was collected from

artificially manipulated matches as a test set. Third,

we used additional dataset, “Dataset B2”, for

validating our model.

3 RESULTS AND DISCUSSION

Previous studies that adopted the Benford’s Law

analysis simply concluded that it can be used in

detecting the occurrence of match-fixing. Therefore,

this study investigated whether the distribution of the

number of ralleys follows the Benford’s Law. We

also explored whether the expected frequencies of

ralleys measured through the Benford’s Law can be

used in detecting match-fixing attempts.

3.1 Benford’s Law Analysis

Figure 3 graphically reflects the comparison of Data

Set B, collected through the artificial match-fixing,

and Data Set A, gathered from ATP and WTA,

against Benford’s Law. Because there are not enough

confirmed cases of match-fixing, obtaining data from

confirmed fixed matches was not simple. The

statistical analysis of these values produces slightly

different results. Table 1 shows that normal games

follow Benford’s Law. The data for that can be found

in “Dataset A”. The data from fixed matches found in

“Dataset B” shows that it doesn't follow Benford’s

Law. Data from sets A and B were not compared with

each other but with the ‘Probability for Benford’s

Law’. The results of the data in Table 1 was verified

by the use of chi-square value. These results differ

from those of a previous study that applied Benford’s

Law to fixed badminton matches (Choi and Park,

2017). In other words, it signifies that unlike in

badminton games, the relationship between the

number of rallies and the Benford’s Law is not

statistically significant in tennis matches. This

suggests that it is difficult to judge whether a game is

fixed based only on the number of rallies made in

each tennis match.

Figure 3: Distribution of Data Set.

icSPORTS 2018 - 6th International Congress on Sport Sciences Research and Technology Support

128

3.2 Detecting Possible Match-fixing

In the above analysis, we confirmed that the number

of rallies recorded in a tennis game with a high

probability of match-fixing is different from the

expected frequency predicted by the Benford’s Law

(Figure 3). Based on lessons learned, we develop a

machine-learning-based model to detect whether a

given game is fixed or not. The model used for this

study contained an h-rally frequency of first digit

feature. The performance results of the proposed

model are presented in the following section.

First, the learning process was performed by

applying 113 sets that account for the 50% of

“Dataset A” gathered from ATP and WTA, and

“Dataset B1” collected from fixed matches. We apply

the artificial neural network as a classifier. The nine

expected frequencies of Benford’s Law were used in

the model. The number of layers of the artificial

neural network was set to 10. The training process is

presented in Figure 4. The model A was obtained by

applying the above learning method.

Second, the evaluation data was composed of 113

games that amount to the remaining 50% of “Dataset

A” and collected “Dataset B2”. The corresponding

test results (game number and the possibilit of its

match-fixing) are shown in Figure 5. Note that the last

three games (no. 111, 112, and 113) were the match-

fixed ones. As shown in Figure 5, the last three sets

of values show a near 1.0 value (.948, .972, and .941).

The other 110 games are identified as normal ones.

This shows that our proposed model can

successfully deliver a positive outcome for detecting

all three fixed matches.

4 CONCLUSIONS

In this study, Benford’s Law and machine learning

were used to examine the possible detection of match-

fixing. Benford’s Law is an empirical rule and a

phenomenon that occurs every day in nature.

Experimental results of this research show that the

numbers collected from fixed games and the expected

frequencies predicted by Benford’s Law are different.

However, it is not easy to simply say that a game is

fixed “simply because it does not follow Benford’s

law”. Therefore, in this study, we made a detection

model for match-fixing through machine learning and

confirmed the results by analyzing the test data.

Experimental results show that fixed games can be

detected by the proposed model implemented through

machine learning.

“Dataset B” used in this study is collected from

artificially fixed matches and it cannot be guaranteed

that a similar set of data will appear in actual games.

However, improved results will be obtained by

constructing the judgment model for machine

learning more precisely and using other features in

addition to the expected frequencies of the Benford’s

Law used in this study.

Table 1: Comparison of Benford's Law and Data Set A, B.

first digits(d) Probability for Benford's law Data Set A Data Set B

30.10 33.97313

36.29547

17.61 16.84369 25.93724

12.49 11.55616 14.35712

9.69 9.180898 8.49764

7.92 7.869342 5.915024

6.69 6.082734 2.804777

5.8 4.957067 3.526798

5.12 4.667906 1.638434

4.58 4.110236 0.827492

n 220 6

mean 58.21 59.26

sd 47.77 67.75

.439 9.210

p .999 .324

Detection of Possible Match-ﬁxing in Tennis Games

129

Figure 4: Training result.

Sports such as horse racing and motorboat racing

play a major role in sports betting and strict

monitoring and control are being systematically

implemented to prevent match-fixing. However,

organizations that make profit illegally through

match-fixing are still making attempts to manipulate

games while evading the surveillance network. Since

the match-fixing attempts are carried forward

covertly, the cases exposed by the monitoring

agencies are very rare, which implies that the match-

fixing is occurring even at this moment.

Figure 5: Testing result.

Sports games classify their results as data. Until

recently, it was very difficult to detect match-fixing

attempts and therefore, punitive measures were

difficult to implement. The lack of punishment has

inevitably made the eradication of match-fixing more

challenging. However, various scientific analysis

techniques that can determine match-fixing attempts

through simple data analysis are being studied

extensively. If these methods can be successfully

implemented in the sports policies, match-fixing

attempts could get completely eradicated as a result.

The initial prevention of match-fixing attempts will

be able to help sports advance a step further.

ACKNOWLEDGEMENTS

We would like to thank the anonymous reviewers for

their helpful comments on an earlier version of this

paper.

REFERENCES

Asser Institute., 2014. Study on risk assessment and

management and prevention of conflicts of interest in

the prevention and fight against betting-related match

fixing in the EU 28. Asser Institute. Luxembourg.

Aquilina, Dawn, and Chetcuti, Angelo. 2014). Match-

fixing: the case of Malta. International Journal of Sport

Policy and Politics. 6(1). 107-128.

Bag, Parimal Kanti. and Saha, Bibhas., 2011. Match-fixing

under competitive odds. Games and Economic

Behavior. 73(2011). 318–344.

Boeri, Tito. and Severgnini, Battista., 2011. Match rigging

and the career concerns of referees. Labour Economics,

18(2011). 349–359.

Carpenter, Kevin, 2013. Match-Fixing: The Biggest Threat

to Sport in the 21st Century?. Sweet and Maxwell’s

International Sports Law Review. 2012(2). 13-24.

Choi, C. H., and Park, J. H., 2017. The Possibility of

Detecting Match-fixing: Benford’s Law in Sports Data.

The Korean Jounml of Measurement and Evaluation in

Physical Education and Sport Science. 19(1). 69-89.

Department For Culture Media And Sport: UK., 2010.

Report of the Sports Betting Integrity Panel.

Department For Culture Media And Sport. London.

European Commission, 2012. Match fixing in sport.

http://ec.europa.eu/assets/eac/sport/library/studies/stud

y-sports-fraud-final-version_en.pdf

European Commission, 2014. Expert Group on Match

Fixing - State of Play on the fight against match fixing.

European Commission. Luxembourg.

European Commission, 2018. Match fixing.

https://ec.europa.eu/sport/policy/organisation-of-sport/

match-fixing_en

FIFA, 2017. FIFA ACTIVITY REPORT 2017. FIFA. Zurich.

icSPORTS 2018 - 6th International Congress on Sport Sciences Research and Technology Support

130

Forbes, 2013. How The 92nd-Ranked Tennis Player In The

World Earns A Comfortable Living. https://www.

forbes.com/sites/miguelmorales/2013/08/26/aces-into-

assets-how-michael-russell-has-made-a-profitable-care

er-in-the-demanding-world-of-pro-tennis/#f585f52475

Hill, Declan, 2010. The Fix: Soccer and Organized Crime.

McClelland and Stewart. Toronto.

Hill, Declan, 2013. The Insider's Guide to Match-Fixing in

Football. Anne McDermid and Associates Limited.

Toronto.

Hill, T. P., 2011. Benford’s Law Blunders. The American

scientist. 86(4). 358-363.

HM Government., 2017. UNITED KINGDOM ANTI-

CORRUPTION STRATEGY 2017-2022. HM

Government. London.

Huffington post, 2016. Tennis Match-Fixing Claims

Involve Wimbledon And French Open Competitors.

https://www.huffingtonpost.co.uk/2016/01/18/tennis-

match-fixing-wimbledon-french-open-competi tors_n_

9 0 06 842.html

IOC., 2016. Rules for the Application during the Games of

the XXXI Olympiad in 2016 in Rio de Janeiro of the

Articles 7, 9 and 10 of the IOC Code of Ethics and of

the Olympic Movement Code on the Prevention of the

Manipulation of Competitions. https://stillmed.

olympic.org/media/Document%20Library/OlympicOr

g/Documents/Code-of-Ethics/2016/Rules-on-the-Prev

ention-of-Manipulation-Rio-2016.pdf#_ga=2.1854964

80.867491679.1524447313-1193877039.1524447313

Katsarova, Ivana, 2016. Match-fixing: Issues and policy

responses. European Parliamentary Research Service.

2016(April). 1-10.

Lastraa, Reynald, Bellb, Peter, and Bonda, Christine, 2018.

Sports betting and the integrity of Australian sport:

Athletes' and non-athletes' perceptions of betting-

motivated corruption in sport. International Journal of

Law, Crime and Justice. 52(2018). 185–198.

Moriconi, Marcelo, 2018. The official football match-

fixing prevention discourse as a cognitive limitation

(the cases of Iberian countries). Soccer and Society.

19(2). 271-287.

Nigrini, Mark, 2012. Benford’s Law. WILEY. New Jersey.

Rodenberg, Ryan M., and Feustel, Elihu D., 2014. Forensic

Sports Analytics Detecting And Predicting Match-

fixing In Tennis. The Journal of Prediction Markets.

8(1). 77-95.

The Korea Times, 2012. Over 10 footballers under probe

for match-rigging scam. http://www.koreatimes.co.

kr/www/nation/2018/04/113_87763.html

TIU, 2018. TENNIS INTEGRITY UNIT Annual Review

2017. TENNIS INTEGRITY UNIT. London.

Detection of Possible Match-ﬁxing in Tennis Games

131