Sports Analytics for Football League Table and Player Performance Prediction Victor Chazan ~ Pantzalis ‘The Data Mining and Analytics research group Schoo! of Science and Technology Tnxernational Hellenic University Thermi, Greece [email protected] gr Abstract-Common Machine Learning applications in sports analytics relate to player injury prediction and prevention, potential skill or market value evaluation, as well as team or player performance prediction. This paper focuses on football. Scope is long-term team and player performance prediction, reliable prediction of the final league table for certain leagues is presented, using past data and advanced statistic. Other predictions for team performance included refer to whether team is going to have a better season than the last one. Furthermore, we approach detection and recording of personal skills and statistical categories that separate an excellent from an average central defender. Experimental results range between ‘encouraging to remarkable, especially given that predictions were based om data available at the beginning ofthe season. Keywords-Sports Analytics, Performance Prediction. Machine Learning (ML), Data Mining, Classification, Regression. 1. InTRopucTION ‘Sports analytes isthe use of historical data and advanced statistics to measure performance, make decisions and predictions regarding perfomance and outcomes, in order to gain an advantage over ‘competitors [I] Performance prediction is the commonest task in sports analytics. Sport analysts process data regarding players and teams with an intended goal: the prediction of match results, toumament winners or tam and individual player efficiency Forecasts may be related to short-term ot long-term events For that reason, diverse methods and algorithms have been deployed, Clubs use sophisticated devices and software (ie GPS. tracking systems) to gather and analyze data generated by players during ming sessions and metches, They process these data to use for short-term decision making and long-term organization development Also, extensive analysis of all data availble is a prerequisite for betting companies. Finally, fans are also very interested in advanced statistics and bow they affect football For all the above reasons, the use of sports analytics has increased during the last few years Football vas selected for our research because of the abundance of statistical categories and historical data, its fame, as well as the simplicity of its rules and of national championships formats. On the other hand, there are special Aifiulties, which make football long-term prediction challenging “The abundance of online data regarding football 1s an asset, but requires filtering and proper data for team and player performance prediction, Unfortunatly, this isnot always easy. Additionally, team tnd player performance can be aflected by incidents not depicted in the data collected; a team is rated higher than it should be when their ‘opponents underperform. A player might have a low rating performance when coming into action ate a serious injury Finally, the nature of football makes statistieal recording of match events as well as player and team rating, an ambiguous process ‘978-1 -6654.2228-4/20831,00 62020 IEEE Christos Tors ‘The Data Mining and Analytics research group School of Seience and Technolo Invernational Hellenic University Thermi, Greece ctjonis@ihu edugr Following the same pattern, performance prediction is not easy and long-term performance prediction is even tougher, but also. not sufficiently studied until nov ‘Nevertheless a iti show in this paper, iis possible, upto a certain Tevel, to make some long-term predictions, especially for team performance Our prediction is relatively good. mainly with regards to the champion and the teams that win European qualification. What ‘makes this research interesting is that the prediction is performed before the beginning of the season, with no official matches played, conly with historical data and the information gathered during the summer break Another novelty of this paper is that advanced ‘statistics fiom previous seasons ae used for prediction. ‘The remaining of this paper is structured as follows. Section Il reviews the erature providing background information, Section II defines the problem and details our approach, Section IV provides experimental results, evaluated in Section V, and Section VI concludes with directions for further work. I. BackGRouND Reep is believed to be the first British notational analyst. He published a statistical analysis of pattems of play in football, along ‘with B. Benjamin in 1968, using 578 matches between 1953 and 1967 During the last 20 years, sophisticated techniques, algorithms, and tools for sports analysis were developed, while articles and papers related to sports analytics are constantly being. published Match outcome prediction is an interesting topic in Sports Analytics, Researchers approach the problem ffom diferent angles. A simple, solid but also obsolete prediction strategy is to predict the number of goals scored by the two teams. ‘The fist model sufficient for predicting the result of a match was created in 1997 by Dixon and Coles. The mode is considered classic tnd was able to extract probabilities forthe goals scored in a match, {lowing Poisson distribution [2] During the last years, researchers, focused on directly predicting wins, draws and losses, nsead of trying to predict goals scored or points ‘won. Various Machine Leamming (ML) algorithms were implemented in order to discover the most discriminating factors that separate the ‘winning from the losing side, Lago-Penas etal. concluded in shots on goal, crosses, mateh location, ball possession and opponent team ability. based on a ranking system [3]. Harrop and Nevill supported that the best predictor is pass accuracy, followed by the number of shots, the numberof passes and dribbles (the fewer the beter) and the venue of the match [4]. Mao et al. claimed that the features that provide the most postive effects are shots on goal, shot accuracy, tackles and acrals wor [3], ‘Tax and Joustra employed a set of factors from public data and used dimensionality reduction technigues, such as Principal Component Analysis (PCA), along with ML algorithms (Naive Bayes and Multilayer Perception) to predict the Dutch football championship, ‘Auhorzesleerees ute ied a: SAKTHIS M. Downcaded on December 18,202 at 17:00:23 UTC from IEEE Xsloe. Resticions spo © scanned ith OREN Scanner‘They achieved an accuracy of almost 55% in their predictions and proved that a hybrid model, combining public data and betting odds could improve accuracy [6] ‘Neural Networks (NNs) have also been used for predietion in football. McCabe and Tievathan dealt with four different sports. Using data from 02-08 anda Multilayer Perception, trained with Back Propagation, equipped with conjugative-gradient algorithms, they tried to predict match results. The NNs had 20 input layer nodes, 10 hidden layer nodes and 1 output layer node. The Same features were tused for every spor, Football had the worst average prediction performance of S4 6% [7]. Then, Hucaljuk and Rakipovie concluded that Ns performed better than any other MIL. technique they used [8] ‘Goddard, in 2005, compared the two methods, i. modeling the goals scored vs modeling win-draw-lose match result and concluded that a hhybrid- model achieves the best prediction performance. He also was fone of the first to use variables other than previous match results. He leveraged features like the importance of individual matches, seographical disuance between the two opponents and more For the Win-draw-lose method he used an ordered probit regression model and exploited a database of English match results for the past 25 years. He also included in his work a comparison of his predictions ‘ith the betting odds of the matches and concluded that achieving a Positive betting return over time is possible [9]. Tnaddition, many remarkable advanced statistics have emerged during the last decade, such as Expected Goals (xG), Packing, Defensive Coverage. Sequences and more. xG is a statistical measure of the quality of chances created and conceded (Expected Goals Against) XG probabilstically assign a score from 0 to 1 to each chance based on several variables, Shot quality evaluation is usually achieved by ning NNs over large dataseis of shots. xG are caleulated for individual players, but also cumulatively for the whole team. The ‘model eliminates some of the randomness ofthe actual goals scored and gives beer insights into team performance. XG has not avoided criticism, but there have been certain eases that the method was Implemented with great success. In 2016, Fagels otal. used xG trying to build a model to classify each scoring opportunity into a scoring probability. They leveraged geospatial data and implemented various clasifcation techniques ‘They also indicated that XG could be further used for evaluating players and seasons, but they warned that probability estimates of goal Scoving opportunities may suffer from high standard deviation [10] ‘Apart ftom works on predictive analysis, there are various intresting researches refering to comparative analysis. The main components compared are wins and losses. It appears that a noteworthy attribute that most researchers point out is efficieney' Efficiency is defined as the number of goals divided by the number of shots. Shots on goa, pass aceuracy, quality of the opponent team, venue of the match and, tall possession also seem to be significant variables [11]. Bekris et al. used a different approach: they compared matches with at most one-goal difference (i.e. short range) to matches with at least three-goals difference (ie. wide range) They found out that wide range winners outplayed their opponents in ball possession percentage, number af passes, “one +s ane” ducls won, number of Shots, number of shots di target and shooting accuracy. Contrarwvise, ‘those findings do not stand for short range matches, which are more sensitive to luck [12]. Researchers have also used the concept of rating. Rating is a single ‘number which is used to describe the strength ofa team in comparison to other teams at the time. lamous rating system isthe ELO Rati Which was used by Hvattum and Arnizen | 13]. They used ELO Rating dierences between teams as covariates in ordered logit regression ‘models Constantinou and Fenton used the pi-ratings that they had eatlier invented for model validation, trying to. make long-term predietion over team performance [14] Predicting the outcome of a match is important, but maybe not as important as the prediction of team performance forthe season Tis obvious that its very hard to predict the long-term performance of a team and itis much harder to predict its performance by comparison with the performance of other teams. Limited work has been done on this challenging task so far One of the most intriguing but also almost unexplored scopes isthe prediction of a championship’ final table. ‘Van Haaren and Davis emphasized on the difficulty to predict the exact position ofa team in the final table, because it depends on the final position of every other team [15] Another obstacle for their ‘method was the number of matches that ended in @ draw. Ranking systems used for simulating match results have difficulties in predicting draws. This resulted in high variance on the predicted frumber of points for each team However, they indicated two substantial metrics needed Tor evaluating the quality of the predicted final tables: the pereentage of correctly predicted relative positions ‘nd the Mean Squared Error (MSE) regarding positions. Obersione developed a multiple regression model, ending up with 6 independent variables which he assessed to be sufficient for predicting the final league table of EPL in terms of points, instead of accurate positions [16]. He also used F distribution to compare means of ‘multiple samples (ie one-way analysis of variance) to investigate ‘which pitch actions differentiate the four best teams fom all the others inthe league. He managed to achieve outstanding results| ‘here have been some interesting works focused on the financial strand of football clubs, rather than pitch performances. Kringstad and. Olsen used data from the Norwegian league and focused on the relationship between financial strength and sporting outcome [17]. “They presented some mixed results: evidence suggested that budgeted revenue was a success indicator, but only for bottom-half teams, ‘while static and dynamic regression models they implemented supported the notion of budgeted revenues being a driver of sporting ‘outcome, They concluded that focus on athletics is still vital as money ‘sa significant factor of success, but only toa certain extent Coates et al used data from every team that participated in Major League Soecer (MLS) in USA during 2008-13. They examined the relationship between salary level and dispersion with football success They revealed that while the wage bill of team has a positive effect on success, salary inequality has a negative effect om success. In that way they proved that cohesion i essential in Footbal 18} Cintia et al. used pass-based performance indicators and other efficient metres, Ike the Pezzali score. The signification of this settc lies oa the fact that it rewards teams effective on both sides of the pitch, ie in offensive skills and in defensive duties. It is formulated as follows Pezzaliscore(team) = ~esttceam)_ ttompetepponen| (4) Tueemprstcam)* goaltopponend ‘They simulated matches from four major leagues and claimed that they achieved superb results, a they predicted match outcome with an accuracy of almost 60% They also found that the final rankings inthe simulated championships were very close to the true sankings Nevertheless. some teams had a considerable ranking eror, which was explained by very high or very low Pezzali score. Finally, they marked the simplicity of their models and encouraged researchers 10 ‘work with more complex models as they reckoned that there i room for improvement in accuracy [19], Constantinou and Fenton, studying predictive accuracy in long-term team performance, proposed a method which they called smart-data 14]. ‘They exploited sternal factors which might influence the sengih of a team (ie, managerial changes, European qualification, rely promoted teams ete.) With those Factors they buat new ones, such as “rue team strength”. “expected performance” and more. Theit goal was to predict the final table in terms of points won by each team. They achieved great results, managed to single out ceriain external factors that boost or worsen a team performanee, focusing on the quality oftheir data, not on the quantity. Football passes are important actions. Cakmak et al introduced a metric, named "Pass Effectiveness” [20]. They based pass evaluation ‘upon mathematic grounds. Pass effectiveness is being extracted from, ‘Auharzesleerees ute ied a: SAKTHIS M. Dwncaded on December 18,202 at 17:00:23 UTC rom IEEE Xsloe. Resticions spo © scanned ith OREN Scanner‘the combination of five other measurable pass metrics: gain ofa pass, pass advantage, goal chance, decision time and pass effectiveness of the next pass Passing networks is also a very intriguing subject, players are represented as nodes of a network. while passes between two players ae represented as edges between the players. The edges are weighted based on the number of passes being exchanged between players Cinta et al. leverage passing networks in several papers, in order 10 predict football matches outcome, but also final league tables. They ‘concluded that networks are more efficient for long-term predictions for whole competitions [21] Grund analyzed a dataset of 283,259 passes and applied mixed-effeets modeling 10 76 repeated ‘observations ofthe interaction networks and performance of 23 soccer teams. He proved that best performing teams were characterized by networks with high intensity and low centralization [22] Spatiotemporal data are significant in sports analytics. The advances fin image processing made the analysis of positional data alot easier Borrie et al suggest that temporal pattern analysis will lead to a \desper understanding of sport performance. They detected temporal pattems to find similar pass sequences within matches [23] Player performance prediction is also interesting. Nsolo et al investigated the attributes which best predict the suecess of individual players, based on their position, and evaluated different ML. algorithms regarding prediction performance. They focused on top players ofthe top five European leagues and evaluated players based ‘on different attributes for each position. They concluded that forwards tend to have higher performance ratings than other players, so maybe ‘mare advanced metrics should be applied on defensive players [24] Previously, we used past data for long-term performance prediction: ‘we estimated how many goals cerian player will score ina season and the number ofa player's shots during each individual match, We also predicted the playing positions of a set of players according to theitattbutes [25] We also predicted the best NBA defender as Well as the MVP for2 years [26] Sitb etal. presented a set of $4 performance criteria, over diferent playing positions in order to evaluate the performance of players, ‘consider each player's natural position and the tatieal formation that the team deployed ina match [27] Finally, Pappatardo et al analyzed player performance ftom 18 different competitions for several years and presented PlayeRank, a ‘data-driven famework. The dataset contained 31 million matches and 21 thousand players. PlayeRank was found to outperform competitive Predictive algorithms. They also discussed what distinguishes top layers from others and discovered patterns for excellent performances One of the limitations was that PlayeRank does not consider off-ball actions, like pressing, The authors also emphasized fom the fact that an improved version ofthe framework should be able to leverage data ftom other sources, like wearables, GPS and video ‘tracking data [28] IL PROBLEM DEFINITION, APPROACH FOLLOWED A. Problem Definition Long-term performance prediction for teams or individual players are fields requiting exploration Not only coaches, but also sports agents and bookmakers are interested in how teams or players perform during, 1 season compared to previous ones, What i discussed inthis section 's the context ofthis problem, Also, the objectives ofthe research are se ‘The unique components of football matches make long-term predictions very difficult, only few goals are scored pet match, Also, ‘there is no clear changeover between the instantaneous change of Possession and transition between offense and defense Moreover. player positions and tactics are not fixed and finally, the game has & ‘continuous flow, which complicates recording of game events [16] (Our research focuses on statistics fom the previous season and. historical data. Also, some financial data (ic. transfer spending, team, salaries) are exploited to contribute tothe team evaluation process. ‘The novelty of this research is that advanced metres were used, such as xG and Pezzali score to predict next season performance before the season begins, not after matches have alteady been played and recorded. Additionally, attackers are usually graded higher than defenders, even if they are not always more influential in team, strategy. So, regarding to player evaluation, this research atompts to ‘dently skills and features suitable, that make good defenders. 1B. Approach Followed ‘This section showeases the flow of evens taking place before we ean get any meaningful experimental results, as well as the way the data ‘were aequited and their preprocessing The block diagram whieh summarizes the a ‘is depicted in Fig. 1 2 Oats Figure 1: Block diagram of the process followed forthe experiments. At first, the appropriate data had to be found. There are lot of web pages thal concain information and statistics regarding football ‘matches and events, The data refer to both teams and playets. Some of the data were accessed and collected manually, especially when that ‘was easy. However, some of them were seraped from the internet using various scraping tools. Finally, a free database from an expired fonline competition had been downloaded and used for the experiments. The database contains data fiom thousands of players and is exacted flom a famous manager simulation game. It demonstrates player ratings for several football skills. Players ate rated by domain experts After the process of data acquisition, there was a large database which ‘needed to be organized. The database was split ito different csv files, according 10 what data were essential for each experiment. Then, the csv files were uploaded to jupyter the software that was maunly used. for data processing Naturally. the data firstly needed to be preprocessed They were checked for null values, duplicates, noise ete Python was used to clean the data and build the models Then, data transformation and. data reduction took place to keep only the appropriate features for cach classification or regression, Finally, results were evaluated in terms of accuracy, eror rates and bias involved. They were also being compared to results of other similar reseaiches to estimate the Value produced by them, IV. EXPERIMENTS AND RESULTS A. Istexperiment: Team Performance Prediction “The fist experiment is divided in two parts: The fist part ean be escribed a5 follows: Having a dataset with every team from four ‘important European football national leagues, with more than 40 features for every team for each of the last four years (2015-18), predict whether a certain team is going to have a better or Worse season than the previous year in terms of points. Every previous season is used as training set and the final season (ie. 2017-18) is ‘Auharzesleerees ute ied a: SAKTHIS M. Dwncaded on December 18,202 at 17:00:23 UTC rom IEEE Xsloe. Resticions spo © scanned ith OREN Scannersed as test set. It is handled as a binary classification problem and the evaluation is conducted by measuring Accuracy? as follows ber of teams wih correct perfromance prediction "humber of total tears @ Accuracy? = “Then, for the second part of the experiment, another method is presented; using almost the same features as inthe fist part, a model \was built, to simulate every match ofthe 2018-19 season for te same championships (.¢ 380 matches per championship). Then. the vital points colleced by teams are accumulated in order to predict the final league standings The predicted leaguc table is compared tothe actual league table and the evaluation s conducted by calculating the Root “Mean Squared Error (RMSE; forthe championship ause = eG») 0) where ‘isthe numberof teams participating in the championship + Si ste predicted pois fr the th team + yj ithe actual points forthe th tan Also, every model is evaluated fr is ability to predic the outcome of matches played ‘The evaluation metric 18 Accuracy, defined as follows: uber of games with corey predicted outcome ‘uber f total games ‘AccuracyM a) “The Features used forthe experiment are the ones that were considered. mone relative to team performance. Those features can be divided in thtee categories 1 Past data generated during the last ive years. This mainly refers to performance indicators ftom previous seasons (eg team average points) 2. Team statistical features from the season that has just ended (e \wins, 4G, shots, possession percentage, Pezzali seore and more) 3. Data not measurable by team performance (e.g. financial). These altubutes are generated during the summer break, so most of them ‘ate independent from the previous season, but very likly to have an impact on the new season's performance Finally, in the dataset, there isthe target attribute. It is binary and. corresponds to whether the team is going to have a better or worse season than the previous one in terms of points wen. Afler data preprocessing, some attributes were removed fom the original datasets, being trelevant with the research or noisy, aiding limited value to the outcome, Those were team statistics, like card, interceptions, offsides, fouls ete ‘The first problem to handle was that not every team of the previous ‘championship takes part in the new one, there are teams that are relegated and teams that are promoted. It is meaningless to have historical data about newly promoted teams, because the data would refer to a different league than the one studied So for the newiy Promoted teams, some adjustments had to be made. Indicatively. calculating the average team points of the last five seasons, if a team ‘was playing in a lower division during that time, the points of the bottom league team were assigned to it. “Those adjustments eaused certain problems; newly promoted teams do rot necessarily have the same strength as teams that have just got relegated. Thus, the way they are described by features and atiibutes assigned to them might not be representative of their actual status Furthermore, th three newly promoted teams are all assigned with the same values for the corresponding variables, which is not efficient Therefore, the validity ofthis method is questionable ‘The data were spit into tain and test set. Then, multiple classifiers were used to classify the test set teams into two classes (ve. beter season / worse season) Grid search was used for mode! tuning and. 10-fold Cross Validation was used for testing the effectiveness of the ‘model. A feature importance graph was also deployed to track the ‘most valuable features; difference between goals and team XG in the previous championship tuned out tobe the most important features (On the other hand, managerial change was not deemed an important performance indicator. However, it must be noted that the atuibute ‘sed in this research does not factor what circumstances caused the ‘managerial change in the club, Similar researches inthe future may eal with this issue (On the mode! build, Random Forest was the classifier that achieved the highest accuracy, with more than 70% AccuracyP and. with standard deviation less than 10% For the second part ofthe experiment, more databases were used. That data pertained to the results of every match ofthe four championships presented in the first part of the experiment. Every unnecessary atuibute was again removed and datasets were merged with the datasets of the first part ofthe experiment. This process resulted ino a ‘ew dataset, which contained every match fom a football season with its fall time result and with statistical, financial, and historical data about the two teams involved in each match Naturally, a problem came up: some of the teams participating in championship matches lack any data, because they are newly promoted. Thus, there were some missing values inthe dataset t0 be hhandled properly. by the following method. Newly promoted teams were considered to be the weakest ones in the league and were assigned the maximum, the minimum or the mean value of the costesponding atuibute. according to the nature of each aturbute “The next step was to combine the attabutes of home and away team by subtracting the corresponding pauts. Some attributes fiom the frst par of the problem were excluded from the second one as the subtraction was not meaningful. Finally. every team was encoded using dummy variables and the first three seasons of each national championship were concatenated. ‘The dataset was split into a training set and a validation set. Last season was Kept separaely ftom others, the target (.e. the full-time result) was hidden and was used as a test set. Training/alidation set consisted of 1140 rows (380 rows for the test set) and almost 40 tributes (tam dummy variables were not included) Standardization of the data, parameter tuning, and Cross Validation techniques were used again, exactly as previously, Multiple classifiers were deployed to predict the match outcome and therefore the Teagues’ final standings Team market value percentage, expected ppomis and non-penalties xG turned out to be the most important features, but not by a big margin from ather attribute. ‘During this process two problems came up: the first problem was that almost every classifier used had the tendency to favor big teams over smaller ones The other problem was that most of the models built faced difficulties in predicting draws. Despite the drawbacks, the results achieved could be considered promising. They are comparable to resalts fom similar researches, ‘while the advantage of this research is that the experiments can be concluded at the beginning of the season, with no official matches played and recorded. The best AscuracyMI for the outcome of the ‘matches was 37% for the English Premier Division and the smallest RMSE for team points was 9, achieved for Spanish La Liga French Ligue | produced the worst results, both in terms of AccuracyM and. RMSE. The results from each league are presented in the following, Tables to 4 The best result foreach league is noted with green color and the worst one with red “Table 1: Results from the English Premier League CLASSIFIER ‘AccuracyM RMSE ‘Naive Bayes 35 1 Decision Tree a5 129 Random Forest 36 133. KNN zy 153 ‘SVM (rbf kernel 34 13 SYM (poly kernel) 7 i XGRoost 32 173 ‘Auharzesleerees ute ied a: SAKTHIS M. Dwncaded on December 18,202 at 17:00:23 UTC rom IEEE Xsloe. Resticions spo © scanned ith OREN Scanner“Table 2: Results from the Spanish La Liga. predicting these teams’ performance It correctly. predicted the CLASSIFIER ‘Accuracy RMSE] champion, but also the ranking of the frst six tars inthe league Naive Bayes 7 237 In this example, SVM with polynomial kemel succeeded not to Decision Tree 0 Ta9 overestimate the top teams (Le problem whieh was offen observed Random Forest 7 177 Throughout most of the clasiiers) but on the other hand ERN a iy overestimated the bottom teams instead. One other problem was it SYM (rbt kernel) 1 ae ie redetng ravs, very ev ma oucomes wee rnd predicted as “draw 2h ete ete = TET Despite ther divergence and how sll or big AccuraeyM and RMSE wore in every case, most of the classifiers correctly predicted the Table 3: Results from the Italian Serte A tar ter toa ree gulcaes Lal hand ie mek eae here MSE] then shar uid fr Champions’ Leap, as shove in Tube iatve Bayes E Results for the relegated teams were also acceptable. Euopa League Decision Tree at aE teams were the exception, asthe prediction accuracy was poor Random Forest 40 laa KNN a7 ia7 “Table S: Accuracy tn predicting champion eam that won European ‘SVM (ror kernel) 32 7 ‘qualification and teams relegate. ‘SVM (poly kernel sm Taz Premier | La [Serie | GS] Oyo XGBoost a 145, League | Liga A i ‘Championship | 7,9, 71% | $7% | 57% 64% “Table 4: Results from the French Legue 1 ‘Winner CLASSIFIER ‘AccuracyM RMSE European sox | vom | xx | ax | 75% Naive Bayes 2 252 Qualification Decision Tree 39 176 ‘Champions 79% 86% | 71% | 57% Ta% ‘Random Forest 45 2a League ‘KNN 39 209 Eorope 38% 29% 0% 29% SVM (rbf kernel) a3 222 League [SVM (poly kernel rn 7 aa | sm | 10% | am ‘NGBoast a zie Its obvious that SVM with polynomial kernel is observed to steadily achieve good results in every league studied, so itis regarded as a benchmark for this research ftom now on, Overall, the best result in tems of RMSE was achieved fiom the Spanish La Liga, where the classifier predicted the final league table with surprisingly high accuracy. given the few attributes used, In Fig. 2, the real vs predicted Teague tables are shown and compared ACTUAL TABLE PREDICTED TABLE, 1. Baredona a 1. Barcelona ® 2 Alielco Madkid 76 2.allico Madd 75 ‘3 Real Madkia os 3. Real Madiid “4 4 Valencia. 6 44 Valencia 3 5 Sevilla 9 5. Sevila 7 6 Getafe 2° 6 Getafe a 7 Espanyol 3 7. Real Betis 37 5 Allie Bilbao 53 8. Eibar 37 9, Real Sociedad 50 9. Cet Vigo a 10, Real Beit 50 10, Vilarreal 55 11 Alves 50 1 Allee Bilbao 54 12, Biber a 12. Real Sociedad 54 13. Leganes 4 13. Legines 4 14. Vilareal “4 14. Espanyol 3 15, Levante 4 45, lives 3 16. Ceta Vigo a 16. Levante a 17, Valdolid a 17, Valladlid 31 18 Girona 7 18 Girona st 19, Fiuesea x 19. Vallecano 50 20, Valecano| 2 20, Hse ° Figure 2: Spanish La Liga 2018-19 actual vs prediced table Green fonts are used for teams that won European qualification through Champions League, blue for teams that won European {qualification through Europa League and red for teams relegated after the end of the season. The classifier has done an outstanding job in Finally, a8 far as AccuraeyM. is eoncemed, another aspect of the experiment is the following: Instead of using the previous three seasons a training set and the last season as validation set. use the first 10 match days. of 2018-19 season as training set and the remaining 28 match days as test set In that case, AccuraeyM of the Spanish La Liga rose from 51% to 70%, as seen in Fig 3. Therefore, it ts shown that present scason’s data can boost the accuracy of the ‘model in a very beneficial manner eet fe Le % aah) Figure 3 Accuracy im predicime maich results after 10 match days ‘from the Spanish La Liga have been analyzed. 1B. 2 experiment: Player Performance Prediction ‘This experiment focuses on individual players, specifically’ central defenders. In rating systems, there is a bias toward forwards and attacking midfielders, Goals are considered the most important clement of football, so defenders’ contribution to a team is usually underestimated. Consequently, there is very limited research on central defenders. Additionally, while it is easy to rate attacking players, according to the goals, key passes and assists, it is not ‘Straightforward what makes good eentral defender. “The purpose ofthis research 1s to examine the characteristics and the statistics for central defenders in comparison with their season rating fand deeide which of them contribute more to distingush central defender as top class player. ‘Authorized icerse ue ete tm: SAKTHI S M_ Downicaded on Decomber 18,202 at 17.0023 UTC trom IEEE Xplore. Restictons apy © scanned wth OnENScnner“The data collected refer to player atributes, playing postions and some demographic features. The database was narrowed down to 59 players, as ony central defenders, playing in English Premier League and having participated in at least 10 league matches for the 2016-17 season were selected. “The next step was to collect season statistics for those players. The main focus was on statistics regarding defensive player actions, but also, some team statistics were collected; despite demanding to build a ‘model based on player performance, it must be acknowledged that a footballer's team has an impact both on his statistics but also en his overall rating “The initial approach to the problem was to normalize every numeri value of the dataset so every attribute's range was transformed to range 0 1o 1. Then a multiple regression model was built with every possible feature. Despite the simplicity of this approach, some useful carly conclusions were drawn regarding to which features contribute ‘more to central defenders’ competency. It scems that forthe examined dataset, inerveptions are the mest important characteristic, flowed by team overall rating, as expected. Players’ best atsibutes turned out to be thei jumping teach, versality, acceleration and fist touch on. the ball Anather approach that was followed was to spit the datasets features ito three categories 1. Player characteristies and atributes 2. Player statistics 3. Team statistics. ‘Again, the target was to build three multiple linear regression models (ie player attributes based, and statistics based), but with fewer Independent variables than in the fitst approach. The method used for the implementation of this part of the experiment was backward climination. For the first category of features, the final model was built with seven features, which seem to be the most influential for a central defender ‘The five assumptions of linear regression were also verified for this model, thete was an indication of lieavty in the model, Also, the expectation (mean) of residuals was almost 2ero and it appeared that there was no (perfect) multicollinearity between features Additionally, by performing a Breusch-Pagan test, i was proven that there is no heteroscedastcity in the model, Nevertheless, the final assumption was not verified. The Durbin-Watson test gives a value ‘much lower than 2. which implies that there was positive autocorrelation between features, Also, the Resquared and. the adjusted R-squared were relatively low (under 05). However, ‘considering that dependent and independent variables emerged from two different sources, the results could he deseribed as encouraging ‘The features ofthe second category (all derived fiom the same source) were the independent variables, while player rating (also derived fiom the same source) was, again, the dependent variable. This time, the final model consisted of 12 features, after Backward Elimination, with very low P-values, while, as seen in Figure 4, R-squared was 0 867 and adjusted R-squared was 0 833, a vast improvement from the first ‘model Figure 4: Model bul with player statistics as independent variables. Additionally. all five assumptions of linear regression were met Linearity of the model was obvious, as seen in Fig. 5. The expectation (mean) of residuals was found almost zero and there was no (perfect) ‘multicollinerity between the features, The Breusch-Pagan test gave a p-value of 0.44, so there was no heteroscedastcity and the Durbin— Watson test gave a value of 191, so there was almost no autocorrelation between the features Figure 5: Linearity ofthe second model “The third set of features (ie team statistics) did not help to build a satisfactory model. There was an indication that the only of those Variables worth noting is “TeammRating” It was decided to incorporate “TeamRating” in the second model, in order to exploit that feature Indeed, by updating the model, adding “TeamRating” as its 13 feature, there has been a slight improvement to the model; R-squared rose t0 0.907 and adjusted R-squared rose to 0 88 In conclusion, summarizing the results of all models deployed, the ‘most critical’ attributes and game actions for predicting, the performance of a eeniral defender can be described in the following list It must be highlighted that attacking skills are not absent fram the list. following the way modem defenders are expected to play Interceptions, Clearances, Acrials Won, Tackles, Jumping reac, Versaili Acceleration, First touch on tll Age. Passing Vision, Determination, Strensth Professionalism and ability to perform well in important matches, Inernational Caps, Minutes Played, Fouls, Inaccurate short passes, Key passes, Goals, “Team's rating V._ Discussion ‘This section reviews and discusses our approach and results. Problems. that came up during the process and the solutions given are debated Results of the experiments are evaluated and threats to validity are mentioned, 10, The first problem encountered was the abundance of data It was practically impossible 10 use every ftee online data acquired, so data ‘Auharzesleerees ute ied a: SAKTHIS M. Dwncaded on December 18,202 at 17:00:23 UTC rom IEEE Xsloe. Resticions spo © scanned ith OREN Scannerselection was challenging. Fortunately, the models produced from the datasets were not computationally intensive, so the approach followed was to include as many attributes relative tothe research as possible, Jn order not to miss out important information. Later, during the feature selection phase, some less important attributes were removed, Conversely. the acquisition of data substantial for research on football analytes was very difficult Data regarding player injuries and data, from wearable devices are mostly defined as private personal details ‘Thus, there are no such free online data to be used for the ‘experiments Another problem was the handling of newly promoted teams, as Statisties from previous seasons were generated for a lower division, so they could not be used. Concerning those teams, values were automaticaly assigned to some variables. That was a necessity, but in certain circumstances the predicted team performance did not meet some teams" real potential Additionally, most models were biased in favor of big clubs An attribute that could be used as a penalization fietor in cases of ‘overestimation could balance out the aforementioned bias ‘A similar problem was encountered because ofthe models” difficulty ‘in predicting draws. ‘The solution behind that problem usually lies on ‘the proper usage of cost sensitive classifiers or by tuning the weights of lasses, In both eases, the proposed solutions were tested. Even though they resolved the issues that were deployed for. they both filed to extract better results than the ones aleady achieved. Hence, they were not ‘included in the madels Every championship has its own particularities, so rules extracted from one league do not necessarily apply to others. Therefore, despite the feasibility of building a universal model, proven by experimental results, the exploration of the differences between leagues would probably provide beter research opportunities. This issue also reflects the second experiment. The database used consisted of defenders based only in England, because it would be inefficient to inelude players ffom different lees. Domain expert opinion was used in rating player atributes Scout reports, despite generally describing well enough player ability and. potential, have often been misleading, while intentional tampering With ratings and attributes should not be ruled out Additionally, finaneial-based data usually suffer from inaccuracies and cannot be fully rusted Unfortunately, football matches are not affected only by team ability and player skills. There are some external factors that cannot be predicted. Luck is an imponderable factor Long term injuries of Important players are also part of the game. “Strange” results in matches where one or both teams are notin real need of vitory are often observed. Finaly, betting odds inevitably have an influence on ‘match outcome. All those drawbacks, which can be viewed as threats to validity, prove that long-term sports prediction is very demanding and may not always provide meaningful results Nevertheless, the results of the experiments conducted in our research ean be described {as good or even impressive in certain occasions. ‘AccuracyP level for the fist part of the first experiment can be described as satisfactory, given the fact that it is a long-term prediction with no official mateh data and statistics available A professional expert could exploit the experimental results, along with his own intuition and make certain decisions. ‘The main achievement of the research is the second part of the experiment, where the models used predicted some femous champions’ final tble with great azeuracy. Also, classifier ae able 10 predict almost 2 out of 3 match outcomes when the model is applied ‘nthe midst of the season. Consequently, this implementation can be vastly used for betting purposes under certain circumstances Provably, planning a profitable betting strategy based on experimental results and -apparently— in some human expertise, is possible Finally, the second experiment suceeeds into locating a set of auibutes and skills that a central defender must improve in ocder to be considered a top class player. OF course, every player is different and. has hisher own playing style, but it would be very useful for coaches. to have a specific targeting when traning @ player. Long-term player prediction performance could also be a huge contsibution to fantasy sports games, The experiment resulted in a vanety of features ‘Unsurprisingly, some of them were the main defending actions and attributes, but in an interesting manner, some were also found to be tacking actions or attacking atributes. VI. CoNeLUSIONS AND FUTURE Work A. Conclusions In this research, two fundamental cases of sports analytics were studied: team performance prediction and player performance prediction. For the frst experiment, the goal was to predict how each team of four important European leagues would perform during the 2018-19 season. The data available were only historical data (from 2015 ‘onwards) and information about team actions (tansfers, managerial changes et.) during the summer of 2018, just before the beginning of the szason, Two approaches were followed to address this issue For the first approach, the target was to classify teams in those that ‘would perform better than last season and those that would perform ‘worse than last season in terms of points collected. Results could be described as satisfactory, but not impressive, as Accuracy? of the classifiers deployed reached the level of 70%. In this approach, no distinction between the examined championships was made, 50 the model used could be deseribed as universal ‘Another approach for team performance prediction achieved remarkable results the idea was to simulate every match of the season and classify their results as home win, draw or away win. At the end, cach team's poinis were accumulated, and a predicted league final table was extracted. The eflectiveness of the model was measured ‘with two metres AccuracyM of the predicted match outcomes and. [RMSE of predicted vs actual team points in the league table. The highest AccuraeyM achieved was 57% for the English Premier League and the lowest RMSE was 9 for the Spanish La Liga ‘Additionally, the champion was correctly predicted in 64% of the times and the teams that won European qualification were correctly predicted in 75% of the times. Also, ths time, the four championships ‘were separately studied and differences between them were evident. (Our results are very satisfactory and comparable to results of similar researches, Regarding prediction of match outcome, Tax and Joustra achieved 56% accuracy [6], while McCabe and Trevathan achieved 54.6% accuracy [7] Joseph et al. achieved their best result sing Baysian Networks, with 592% accuracy [29] and Eggels et al achieved 54% aceuracy [10] Cinta et al. predicted match outcome ‘with 60% accuracy and team points with 9.1 RMSE [19] Our results have the advantage of being obiained without any current official satch data available ‘Additionally, applying prediction after using the first ten match days. fof the season as a training set was suggested as an alterative In that case, AcvuraeyM of predieted match outcomes was impressively raised 10 70%. “The second experiment was about defining which attributes and match actions are mainly influencing a eentral defender’ match rating. The dataset consisted of 39 central defenders having played at least 10 ‘matehes forthe English Premier League 2016-17 season. The method. used was Multiple Linear Regression with Backivard Elimination and the evaluation metres were R-squared and adjusted R-squared Findings were noteworthy, as for a quite satisfying 0.907 Resquared and 0.88 adjusted R-squared, thirteen features were proved to be statistically significant. Classic defensive actions like interceptions and clearances were amongst them. along with player attributes more Suitable for defenders, such as jumping reach and strength The interesting part was that some attacking skill, such as passing. and ‘Auharzesleerees ute ied a: SAKTHIS M. Dwncaded on December 18,202 at 17:00:23 UTC rom IEEE Xsloe. Resticions spo © scanned ith OREN Scanner