Techno-Science 5:2 (2022) 57-62

Techno-Science
Review
Article

Scientific Journal of Mehmet Akif Ersoy University

www.dergipark.gov.tr/sjmakeu

A SURVEY ON FOOTBALL PLAYER PERFORMANCE AND VALUE

ESTIMATION USING MACHINE LEARNING TECHNIQUES
Vehbi Hakan SAYAN1 , Emrah HANÇER2* ,
1 Management Information Systems, Süleyman Demirel University, Turkey
2 Department of Software Engineering, Burdur Mehmet Akif Ersoy University, Turkey

ARTICLE INFO ABSTRACT

Article History The popularity of games like FIFA and Football Manager has attracted millions of players. Data
Received : 07/12/2022 plays an increasingly prominent role in these games. In other words, these games contain data
Revised : 30/12/2022 from all soccer players worldwide, which can be used to simulate soccer games. Therefore, a lot
Accepted : 30/12/2022 of experts and a lot of tools are involved in producing football data. It is not only the production
Available online : 31/12/2022 of this data that has attracted researchers, but also the analysis of it since useful and consistent
Keywords information can be extracted from it to design real-world football applications. In this study, we
Football Video Games, FIFA, Football introduce a survey of studies which focus on the prediction of player value/performance using
Manager, Machine Learning machine learning techniques. As far as we know, there is no survey in the literature that
specifically addresses this topic.

1. INTRODUCTION

As a sport branch, football spread across the world during the 20th century and became popular across the globe. Football
became a major entertainment industry in the 21st century. In addition to sponsorship deals and TV deals, this expansion
affected the market value of football players as well. A player’s market value can be considered an estimate of how much a
team would be willing to pay to get him to sign a contract, regardless of the actual deal [1]. The market value of a player is
resulting from various factors like talent, popularity, skill, playing style, efficiency, etc. [2]. Therefore, players’ values are still
highly judgmental in the present-day football world. It is common for even the most richly resourced clubs in the world to
transfer players for fees they regret later on. At the same time, a player who attracts high transfer fees for his small club can
become the club's savior. For football clubs to be successful, both financially and competitively, transfer decisions are
becoming increasingly crucial in this competitive environment. A machine learning algorithm may assist coaches and clubs
in deciding which players to transfer. However, due to the difficulty of gathering detailed information about players, there
has been limited research in football analytics using machine learning techniques.

Transfermarkt.com is the leading online source that uses crowd estimation to determine the value of players on the transfer
market. A number of football-related data can be found on the site, such as results, scores, transfer rumors, football news
and estimates of market value for most professional football leagues [4]. The method used in Transfermarkt.com involves
members of the website estimating the players' values, which are determined by special members, known as mentors, based
on the estimates of the other members. However, this method has following limitations. Firstly, inexperienced members can
lead to inaccurate estimations. In addition, it is generally appropriate for well-known players. Thirdly, the value of players
is not frequently updated since the system requires feedback from mentors. As an alternative to Transfermarkt.com, video
games like FIFA and Football Manager (FM) are also considered as sources of data for football analytics. The use of video
games as an alternative source of data has been gaining traction since 2014. It is an asset to have so much online data about
football. However, it needs to be filtered and analyzed properly in order to make predictions about player value. Moreover,
the process is not always straightforward. Additionally, incidents not captured in the data may influence a player's
performance; for example, poor performances by opponents may cause a player to be rated higher than he deserves.

_____
* Corresponding Author: [email protected]

To cite this article: SAYAN V.H., HANCER E., (2022). A Survey on Football Player Performance and Value Estimation Using Machine Learning
Techniques, Techno-Science, vol. 5 no. 2- p. 57-62

e-ISSN 2651-3722 © 2018 Burdur Mehmet Akif Ersoy University (Turkey). All rights reserved.
 Hancer E., Sayan V.H., Techno-Science 5:2 (2022) 57-62

Various studies have been conducted with football data in the literature. The following are some areas of study: football
player value estimation, match result prediction, predictions made using football video data, crowdsourcing for player value
estimation, social networking analysis for player value prediction, etc. In this survey, we consider studies concerning player
performance and value estimation with data based on FIFA video games. To the best of our knowledge, there is no survey in
the literature that specifically addresses this topic.

The remainder of the paper is as follows. Section 2 describes the overall view of FIFA datasets. Section 3 presents the
corresponding studies with discussions. Section 4 summarizes the corresponding studies.

2. STUDIES ON PLAYER PERFORMANCE AND VALUE ESTIMATION

FIFA is a worldwide video game developed by Electronic Arts (EA). In order to keep the data up-to-date, EA employs 30 data
producers and 400 data contributors. Furthermore, the database is reviewed or modified by 6000 SOFIFA talent scouts and
reviewers. The FIFA Soccer Games provide a comprehensive and coherent scout of players from all over the world. The game
includes all the major leagues in the world with over 17000 players, 30 leagues, and 700 clubs. Simulated FIFA players are
categorized into 14 positions and rated on 29 different skills, with each skill being evaluated on a 0-to-100 scale. As well as
attributes such as mental abilities, physical abilities, field position, and other characteristics (cardinal variables) such as
nationality, flag, etc, the data also includes monetary values, wages, and release clauses of football players. Some studies
conducted on FIFA-based data are described as follows.

Lee et al. [5] examined the key factors affecting the transfer fees of top soccer players around the world based on FIFA data
analysis. An improved LightGBM model was proposed by optimizing its hyper parameters using a Tree-structured Parzen
Estimator (TPE) algorithm. A Shapley Additive Explanations (SHAP) algorithm was used to identify prominent features.
Various baseline regression models (including linear regression, lasso, elastic net, and kernel ridge regression) and gradient
boosting models without hyper parameter optimization were compared against the proposed method. The optimized
LightGBM model outperformed the others.
Table 1. Summary of Articles
Data
Reference Research Purpose Source Features Modelling Technique Metrics
Lee et al. Football player value FIFA, 124 features Linear Regression, Lasso Regression, Kernel MAE ,RMSE
(2022) estimation using WhoScored. Rigde Regression, Elastic Net, Gradient
bayesian ensemble com Boosting Decision Tress, LightGBM
approach
Chavan and Applying machine FIFA 43 attributes, Naive Bayes, Decision Tree, Support Vector Accuracy,
Dondio learning techniques for 17589 players Machines, XGBoost, K Nearest Neighbours Precision,
(2019) football player selection Recall, F1
Al Asadi and Machine learning with FIFA Base model 29 Random Forest Classifier Accuracy,
Taşdemir balancing and feature features, Precision,
(2021) selection techniques decreased to 10 Recall
Al Asadi and Football player value FIFA 7 features Linear Regression, Multiple Linear MAE, RMSE,
Taşdemir estimation using Regression, Decision Trees, Random R2
(2021) machine learning Forests
techniques
Yaldo and Football Player wage FIFA 40 features Additive Regression, Decision Table, MAE, R
Shamir calculation, and analysis Nearest Neighbor with a weighted
(2017) according to different condition, K*, Locally Weighted Learning
leagues, different with Naïve Bayes and Linear regression
positions by using classifiers, Random Committee, Random
machine learning Trees and Random Subspace
techniques
Pariath et al. Football player FIFA 36 features, Linear Regression, PCA MAE, RMSE,
(2018) performance prediction 21280 players (4 R2, Median
using machine learning dataset for 4 Absolute
player positions
Error
created, 6000-
9000 players)
Yigit et al. Football player value Football 55 features, Validation Set, Cross Validation, Ridge MSE
(2020) estimation using Manager, 5316 players Regression, Lasso Regression, Principal
machine learning www.trans Component Regression and Partial
techniques fermarkt.c Least Squares, Decision Tree, Random
om Forests and Extreme Gradient
Boosting, Ensemble Model
(0,7*Decision Tree + 0,3*Regression)

58
 Hancer E., Sayan V.H., Techno-Science 5:2 (2022) 57-62

Table 2. Summary of Articles

Research
Reference Purpose Data Source Features Modelling Technique Metrics
Apostolou Player position, FIFA - Random Forest, Multi Confusion Matrix,
and Tjortjis number of goals Layer Perceptron Number of goals,
(2019) for season and Classifier, Linear Number of shots
number of shots Support Vector Classifier
per match
calculation
Rajesh et al. Team formation FIFA 36 features Naïve Bayes, Random Accuracy, F1,
(2020) and different Forest, Decision Tree, Jaccard Similarity
statistics by Support Vector
machine learning Classification, one
proposed prediction
algorithm. K-Means
He et al. Football player's www.transfermarkt.com, 100 features Lasso Regression MSE, Developed
(2015) performance and WhoScored.com, for value 84 Metric
market value European Football for voting
Database and Guardian
voting system
Stanojevic Player market www.transfermarkt.com, 45 features, Random Forests, RMSE, R, Median
and estimation and InStat 12858 Gradient Boosting Trees Error
Gyarmati comparison with players Regression and
(2016) real values using generalized linear
machine learning models
methods
Rao and Team formation FIFA 30 features, Artificial Neural Accuracy,
Shrivastava with machine 17800 Network, Multinomial Precision, Recall,
(2017) learning players Logistic Regression, F1
algorithms Random Forests
Uzochukwu Football player www.pesstatsdatabase.co 30 features Neural Networks Metrics
et al. selection with m developed and
(2015) machine learning average status
calculated
Behravan Football player FIFA 55 features, Support Vector MSE, RMSE, MAE,
and Razavi estimation with in different Regression MAPE, R2
(2020) particle swarm numbers
optimization and after APSO
support vector
regression
Cotta et. al Clustering for FIFA 19 features Linear Regression, PCA,
(2016) soccer analytics K-Means -

A methodology was proposed by Chavan and Dondio [6] to locate the most appropriate player within a short period of time
for a player who moved to another club. In the preprocessing stage, data cleaning and feature engineering processes were
applied to prepare the data for further process. In the second stage, data transformation (e.g. scaling and binary to numeric
attribute conversation) was performed to transform the data into an appropriate format for a classifier. In the final stage, a
variety of classification methods, including support vector machine (SVM), linear discriminant analysis (LDA) and k-nearest
neighbors (KNN) were applied using the 10-fold cross-validation technique to determine the most suitable player instead
of a player moving to another club. According to the results, SVM achieved the best scores.

Al-Asadi and Tasdemir [7] introduced a methodology to characterize football players in terms of nine field positions. In the
first stage, data resampling techniques were applied to address class-imbalance problems in the data. In the second stage,
feature selection techniques were conducted to deal with high-dimensionality. Finally, a classification process was
conducted using a variety of classifiers. Although feature selection and resampling techniques have a significant impact on
the classification performance, the methodology requires improvements to achieve high prediction performance.

In another study, Al-Asadi and Tasdemir [8] proposed a methodology to estimate the market value of players. In the first
stage, redundant features were eliminated, and missing values were filled. Moreover, the positions of football players were
categorized as three categories, namely forwarder, midfielder and defender. In the second stage, Pearson correlation

59
 Hancer E., Sayan V.H., Techno-Science 5:2 (2022) 57-62

analysis was conducted to identify the relationships between features. Finally, linear regression process was conducted by
using a variety of techniques, such as linear regression, decision trees and random forests. The results showed that random
tree-based models outperformed linear models in terms of predicting the market value.
Yaldo and Shamir [9] proposed a methodology to determine the wages of football players depending on their skills. A
comprehensive analysis was conducted using 8 different regression methods. Specifically, overpaid and underpaid players
were determined using predicted and actual values. Skills that differentiate between overpaid and underpaid players were
determined by LDA. Moreover, Relief method was used in order to determine the skills that have the highest impact on
salary.

Parath et al [10] introduced a methodology to determine the performance levels of players. In the first stage, an explanatory
data analysis was performed using scraping tools. As a result, players were grouped in four fundamental categories in terms
of their field positions, attackers, midfielders, defenders and goalkeepers. Each category was individually considered for the
determination of player performance level. In the second stage, principal component analysis (PCA) was performed to
eliminate adverse impact of noisy features on the learning process. In the last stage, linear regression was individually
conducted on the reduced dataset for each specific category. According to the results, the prediction performance of the
methodology was promising.

Yigit et al [11] introduced a linear regression model to estimate player values. The regression model was built on the
following attributes: ability, age, multiplication of age and concentration, and multiplication of determination and technique.
Ridge regression, lasso regression, and tree-based regressors were selected to carry out the regression process. The results
showed that the introduced model performed well in terms of predicting a player’s value.
Apostolou and Tjortis [12] tried to predict the field position, the number of goals scored for a season and the number of
shots during a match. For the performance verification, they selected two top football players and then made comparisons
according to their data. However, the generalized performance evaluation scores were not presented in this study; therefore,
it is not possible to make analysis concerning the generalization of the proposed method.

There were three objectives Rajesh et al [13] considered: speeding up the selection process of players, facilitating the
selection of brand ambassadors, and creating a dream team. It was concluded from statistical analysis and analytical
comparisons that a nationality with high potential and higher overall performance was well suited to an international career,
and a nationality with moderate potential and performance was more likely to succeed in domestic leagues. Moreover,
players between the ages of 21 and 26 were considered to be the most profitable when considering age and performance
together. In terms of nationality, England and its neighbors were ranked first, followed by America and Africa.

He et al [14] focused on discovering the relationship between the market value and the player performance. To achieve this,
regression models were proposed to predict the market value and the player performance using Lasso regression. While the
dataset used in the market value prediction had 100 attributes and 37 forward players who transferred in the same year,
the dataset used in the player performance prediction had 40 forward players with 84 attributes. The results showed that
each field position brought its own characteristics, resulting significant difference between the players. According to a
particular regression model, a good forward player should have the following attributes: few fouls, shots and goals in penalty
area, shots on target, goals from out of box, dribble successfully and assists.

Stanojevic and Gyarmati [15] proposed a methodology for data-driven player market value determination. A comparison
was made between the Transfermarkt market value estimates (TMVE) and the proposed performance-driven market value
estimates (PDMVE). There are three types of features: performance features, player information features and team
information features. To select the appropriate hyper-parameters, a small validation set (ten percent of the training data)
was used as well as a grid search. The error rate of gradient-boosting trees was slightly lower than that of other methods.
Moreover, the top 10 and top 10 undervalued players were listed according to PDMVE. In all metrics, PDMVE achieved a
better result than TMVE by 4-7%.

Rao and Shrivastava [16] introduced a methodology to determine the field position of players. In the first stage, data was
collected and then a feature selection process was applied. In the second stage, 14 positions were mapped to 3 predefined
classes, namely attack, mid and defense. In the last stage, three classifiers, namely multi-layer neural networks, random
forest and multinomial logistic regression were performed through GridSearchCV. According to the results, mid and defense
positions were generally correctly determined by all three classifiers, while the prediction performance was low for the
attack position. The reason is that fewer players play in the attack position. In the case of players with more than one
position, the prediction performance was not satisfactory.

Uzochukwu and Enyindah [17] introduced a neural network model to determine characterized attributes of players,
involving resistance, speed, technique and physical status. To determine each characterized attribute, some attributes were
taken into consideration as inputs for the neural network. After processing neural network models, the average status was
calculated by taking the mean of resistance, speed, technique and physical status. The coach will reject the player if the

60
 Hancer E., Sayan V.H., Techno-Science 5:2 (2022) 57-62

average status is below 50; if it is 50, the coach can select the player, but he will not be in the starting eleven; if it is above
50, the coach selects the player and he will be in the starting squad. Although neural networks produced promising results
in this study, a more comprehensive analysis is required.

Behravan and Razavi [18] introduced a two-stage approach to estimate football players’ value. In the first stage, an automatic
clustering method based on particle swarm optimization (PSO) was applied to group the data. Accordingly, the data was
divided into four groups. Depending on the attributes, the generated groups were goalkeepers with 29 features, strikers
with 32 features, midfielders with 28 features and defenders with 30 features. In other words, the proposed method is
capable of automatically determine which attributes are important for each cluster. In the second stage, a hybrid version of
PSO support vector regression was applied to estimate player values. According to results, PSO performed better than recent
evolutionary techniques.

Cotta et al. [19] tried to classify players in order to gain insights into how teams succeed. Firstly, players were divided into
three categories: defender, midfielder and attacker. Mean overall attribute for each player was calculated by using 8 years
football data from 2007-2014. Since players can change during a season, the top 20 overall players were selected for each
category. Using linear regression, a player was predicted to improve or deteriorate for each attribute over the years. And by
looking at this data it was seen how attributes change in each team over the years for each category. By following this
methodology, they explained one big shocking result in football history, namely Germany's 7-1 win over Brazil in the 2014
FIFA World Cup. In this manner and then by applying PCA and K-means, they were able to explain also one of the biggest
success stories in football history, FC Barcelona between 2008 and 2012.

3. CONCLUSION

In this study, we introduce a survey of recent studies on predicting the performance and value of football players. All studies
are considered in terms of research purpose, data source, number of features and players, machine learning techniques,
evaluation metrics and additional processes. We also summarize the reviewed studies in Table 1 and 2. It can be extracted
from Table 1 and 2 that most of the studies are based on regression models, while the remaining studies consider the data
as a classification problem. For the subject of estimating player values, these regression studies are understandable.
Furthermore, there has been a growing interest in the development of robust sport analysis systems in recent years. To
achieve robust football analytic systems, the following issues need to be considered: 1) standard datasets should be provided
for researchers to conduct consistent experiments; 2) more comparative studies should be provided to see which machine
learning techniques work well in a problem; 3) it is necessary to determine which analyses are crucial for football analytics.
Therefore, developing robust sport analysis systems is still an open issue.

In the future, we also plan to develop a methodology using machine learning techniques to estimate player values and
categorize players according to their skills.

REFERENCES

[1] Herm S., Callsen-Bracker H.M., Kreis H. (2014). When the crowd evaluates soccer players’ market values: Accuracy
and evaluation attributes of an online community, Sport Management Review, Vol 17, No:4, p.484-492, DOI:
10.1016/j.smr.2013.12.006
[2] Singh P., Lamba P. 2019). Influence of crowdsourcing, popularity and previous year statistics in market value
estimation of football players. Journal of Discrete Mathematical Sciences and Cryptography, Vol 22, No 2, p. 113-
126, DOI: 10.1080/09720529.2019.1576333
[3] Kirschstein T., Liebscher S. (2019). Assessing the market values of soccer players – a robust analysis of data from
German 1. and 2. Bundesliga, Journal of Applied Statistics, Vol. 46, No 7, p. 1336-1349, DOI:
10.1016/j.ejor.2017.05.005
[4] Müller O., Simons A., Weinmann M.. (2017). Beyond crowd judgments: Data-driven estimation of market value in
association football. European Journal of Operation Research, Vol. 263, No:2, p.611-624, DOI:
10.1016/j.ejor.2017.05.005
[5] Hanso L., Tama B.A., Cha M. (2022). Prediction of Football Player Value using Bayesian Ensemble Approach. Preprint
to be submitted. DOI:10.48550/arXiv.2206.13246
[6] Chavan A., Dondio P. (2019). Recruitment of Suitable Football Player by using Machine Learning Techniques. Msc.
Research Project
[7] Al Asadi M.A., Taşdemir Ş. (2021). Empirical Comparisons for Combining Balancing and Feature Selection Strategies
for Characterizing Football Players Using FIFA Video Game System. IEEE ACCESS, Vol 9, p. 149266-149286, DOI:
10.1109/ACCESS.2021.3124931
[8] Al Asadi M.A., Taşdemir Ş. (2021). Predict the Value of Football Players Using FIFA Video Game Data and Machine
Learning Techniques. IEEE ACCESS, Vol 10, p. 22631-22645, DOI: 10.1109/ACCESS.2022.3154767

61
 Hancer E., Sayan V.H., Techno-Science 5:2 (2022) 57-62

[9] Yaldo L., Shamir L. (2017). Computational Estimation of Football Player Wages. International Journal of Computer
Science in Sport, Vol 16, No 1, p. 18-38, DOI: 10.48550/arXiv.2206.13246
[10] Parath R., Shah S., Surve A., Mittal J. (2018). Player Performance Prediction in Football Game, 2nd International
conference on Electronics, Communication and Aerospace Technology (ICECA2018), p. 1148-1153, DOI:
10.1109/ICECA.2018.8474750
[11] Yiğit A.T., Samak B., Kaya T. (2020). Football Player Value Assessment Using Machine Learning Techniques, Springer
Nature Switzerland, p. 289-297, DOI: 10.1007/978-3-030-23756-1_36
[12] Apostolou K., Tjortjis C. (2019). Sports Analytics algorithms for performance prediction 10th International
Conference on Information, Intelligence, Systems and Applications (IISA), DOI:10.1109/IISA.2019.8900754
[13] Rajesh P., Alam M., Tahernezhadi M., (2020). A Data Science Approach to Football Team Player Selection. IEEE
International Conference on Electro Information Technology (EIT), p. 175-183,
DOI:10.1109/EIT48999.2020.9208331
[14] He M., Cachucho R., Knobbe A. (2015). Football Player’s Performance and Market Value. Proceedings of the 2nd
workshop of sports analytics, European Conference on Machine Learning and Principles and Practice of Knowledge
Discovery in Databases (ECML PKDD)
[15] Stanojevic R., Gyarmati L. (2016). Towards data-driven football player assessment. IEEE 16th International
Conference on Data Mining Workshops, p. 167-172, DOI: 10.1109/ICDMW.2016.0031
[16] Rao V., Shrivastava A. (2017). Team Strategizing using a Machine Learning Approach. International Conference on
Inventive Computing and Informatics (ICICI 2017), p. 1032-1035, DOI: 10.1109/ICICI.2017.8365296
[17] Uzochukwu O. C., Enyindah P. (2015). A Machine Learning Application for Football Players’ Selection. International
Journal of Engineering Research & Technology (IJERT). Vol. 4, No 10, p.459-465, DOI : 10.17577/IJERTV4IS100323
[18] Behravan I., Razavi S.M. (2020). A novel machine learning method for estimating football players’ value in the
transfer market. Soft Computing, Vol. 25, p 2499-2511,DOI:10.1007/s00500-020-05319-3
[19] Cotta L., Benevenuto F., Vaz de Melo P., Loureiro A. (2016). Using FIFA Soccer video game data for soccer analytics,
Workshop on large scale sports analytics, DOI: 10.1145/1235

Techno-Science Paper ID: 1215552

62