Paper deep dive
Predicting Tennis Serve directions with Machine Learning
Ying Zhu, Ruthuparna Naikar
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 92%
Last extracted: 7/20/2026, 10:28:16 AM
Summary
This paper presents a machine learning method for predicting professional tennis players' first serve directions using data from the Match Charting Project. The study analyzes features such as previous point outcomes, fatigue, and performance anxiety to predict serve directions (wide, body, down-the-T) for deuce and ad sides. The model achieves an average prediction accuracy of 49% for male players and 44% for female players, suggesting that top players use mixed strategies and that contextual information is crucial for anticipatory reactions.
Entities (10)
Relation Signals (6)
Match Charting Project â usedby â Machine Learning Method
confidence 95% ¡ We used the data from the Match Charting Project [15]... to help understand the players' serve decisions, we have developed a machine learning method...
Mixed-Strategy Nash Equilibrium â explains â Serve Selection
confidence 92% ¡ found that the theory of mixed-strategy Nash equilibrium can largely explain the top playersâ selection of serve directions.
Contextual Information â importantfor â Anticipatory Reactions
confidence 90% ¡ contextual information is perhaps more important for returnersâ anticipatory reactions than previously thought.
Performance Anxiety â influences â Serve Decision
confidence 90% ¡ Performance anxiety could influence a playerâs decisions.
Fatigue â influences â Serve Direction
confidence 90% ¡ fatigue might be a factor in choosing serve directions.
Random Forest â achievesaccuracyfor â Novak Djokovic
confidence 85% ¡ Table 1 shows Random Forest accuracy for Novak Djokovic is 0.50.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Serves, especially first serves, are very important in professional tennis. Servers choose their serve directions strategically to maximize their winning chances while trying to be unpredictable. On the other hand, returners try to predict serve directions to make good returns. The mind game between servers and returners is an important part of decision-making in professional tennis matches. To help understand the players' serve decisions, we have developed a machine learning method for predicting professional tennis players' first serve directions. Through feature engineering, our method achieves an average prediction accuracy of around 49\% for male players and 44\% for female players. Our analysis provides some evidence that top professional players use a mixed-strategy model in serving decisions and that fatigue might be a factor in choosing serve directions. Our analysis also suggests that contextual information is perhaps more important for returners' anticipatory reactions than previously thought.
Tags
Links
- Source: https://arxiv.org/abs/2602.22527v1
- Canonical: https://arxiv.org/abs/2602.22527v1
Trouble viewing inline? Open PDF directly â
Full Text
34,517 characters extracted from source content.
Expand or collapse full text
Predicting Tennis Serve Directions with Machine Learning Ying Zhu and Ruthuparna Naikar Georgia State University Atlanta USA yzhu@gsu.edu rnaikar1@student.gsu.edu Abstract. Serves, especially first serves, are very important in professional tennis. Servers choose their serve directions strategically to maximize their winning chances while trying to be unpredictable. On the other hand, returners try to predict serve directions to make good returns. The mind game between servers and returners is an important part of decision-making in professional tennis matches. To help understand the playersâ serve decisions, we have developed a machine learning method for predicting professional tennis playersâ first serve directions. Through feature engineering, our method achieves an average prediction accuracy of around 49% for male players and 44% for female players. Our analysis provides some evidence that top professional players use a mixed-strategy model in serving decisions and that fatigue might be a factor in choosing serve directions. Our analysis also suggests that contextual information is perhaps more important for returnersâ anticipatory reactions than previously thought. Keywords: Tennis serve directions â ¡ Machine learning â ¡ Prediction. 1 Introduction At the beginning of each point in a tennis match, the serving player needs to decide where to direct the serve. There are three serve directions: wide, body (to the returner), and down-the-T (to the middle of the court) for each serve. Each player makes about 100 such decisions in a typical professional tennis match. These are important decisions because serves, especially first serves, are crucial in professional tennis matches [10, 12, 14]. A fast and well-placed first serve gives the server a big advantage. For example, based on ATP statistics [2], Novak Djokovic (world No. 1 for much of the last seven years) has won 76.5% of his first service points and 53.4% of his second service points. In other words, without his first serves, Djokovic was only slightly better than his opponents. The player who serves (the server) chooses the serve directions strategically to maximize the winning chances. On the other hand, the player who returns the ball (the returner) will try to predict serve directions by analyzing serve patterns, which is especially important for the fast first serves. The average first serve speed is about 180â200 km/h for male tennis professionals and about 150â170 km/h for female professionals. With less than 0.5 seconds to react to such powerful serves, a returner needs a fast physical reaction and reasonably accurate anticipation. Previous research suggested that tennis playersâ anticipatory responses are informed by both the kinematics of serve motion and contextual information [19, 18] but the nature of such anticipatory responses is still being debated [3]. The mind game between a server and a returner makes it an interesting case for studying human decision-making in a highly competitive environment. Some economists have used game theories to analyze professional playersâ serve patterns, with mixed results [21, 5, 16, 4, 1]. As more tennis data are available, Wei et al. [22] have developed a machine learning method to predict serve directions by analyzing the Hawkeye data, with a prediction accuracy of 27.8%. In this paper, we present our work on predicting professional tennis playersâ first serve directions by machine learning. Our machine learning model uses human annotated professional tennis match data, without video analysis or Hawkeye data. Through feature engineering and model tuning, we achieved an average prediction accuracy of about 49% for male players and about 44% for female players. Our results provide more insights into the behavior of both tennis servers and returners. Our analysis provides some evidence that top professional players use a mixed strategy for choosing serve directions and that fatigue might affect such decisions. In addition, our work shows that it is possible to achieve reasonable serve direction predictions solely based on contextual information, suggesting that contextual information is perhaps more important for returnersâ anticipatory reactions than previously thought. Professional tennis players rarely reveal their on-court decision-making process. Since many decisions are made intuitively, a player may not even be aware of their own tendencies and patterns. Building machine learning models to predict serve directions may help us gain insights into their decision-making process as well as the differences between players. Players and coaches may also use the machine learning model to evaluate the predictability of their serves. Our study shows that some playersâ serves are more predictable than others. Some players may be more predictable serving from the ad side than from the deuce side and vice versa. Such information can guide players to fine-tune their own games or study their opponentsâ games. 2 Related Work Tennis data have been used in many academic research projects. Here we focus on the analysis of tennis serve directions. Some economists have used game theories to study the optimal strategy for choosing the serve directions. Walker and Wooders [21] analyzed ten professional tennis matches and found that the theory of mixed-strategy Nash equilibrium can largely explain the top playersâ selection of serve directions. They noted that top players tended to switch strategies frequently, resulting in serial dependence and higher predictability. Hsu et al. [5] revisited Walker and Woodersâ work using a broader data set and found no significant evidence of serial dependence. However, Spiliopoulos [16] analyzed the data from the Match Charting Project [15] and found some top male players had higher serial dependencies in serve directions than others. More recently, Gauriot et al. [4] analyzed the Hawkeye data from over 3000 tennis matches played at the Australian Open and confirmed the finding by Walker and Wooders [21]. However, Anderson et al. [1] analyzed the data from the Match Charting Project [15] and rejected a key implication of a mixed-strategy Nash equilibrium, that the probability of winning a service game is the same for all serve directions. They argued that the dynamic programming strategy is more efficient than the mixed strategy. Wei et al. [22] analyzed the Hawkeye data from three years of the Australian Open menâs draw and developed a method to predict serve directions. They considered 14 serve directions (seven for the deuce side and seven for the ad side) and used machine learning techniques (e.g., Random Forest) to make predictions. Their input parameters are score, player style, and opponent style. A playerâs style is the distribution of the playerâs serve count in the 14 directions. Their highest prediction accuracy is 27.8%. Our work is similar to Wei et al. [22] in that we both try to predict tennis serve directions using machine learning techniques. The difference is that we use the Match Charting Project data [15] instead of the Hawkeye data. We use a different set of features, and we adopt the commonly used six serve directions (wide, body, and down-the-T for the deuce or ad side) rather than 14 directions. Kovalchik and Reid [7] analyzed the Hawkeye data from singles matches at the 2015 to 2017 Australian Open and built a taxonomy of shots via clustering. They reported the overall distributions of serve directions for the deuce and ad side separately but did not predict serve directions for individual players. Tea and Swartz [17] analyzed the ball tracking data from the 2019 and 2020 Roland Garros tournaments, which contain 82 menâs and 81 womenâs matches. They used Bayesian Multinomial Logistic Regression to build a predictive model of serve directions. They found discernible differences between male and female players and between individual players. Their model can output predictive distributions of serve directions. An example with Roger Federerâs data was discussed, but the model prediction accuracy for individual players was not reported. Whiteside and Reid [23] used machine learning (k-means clustering) to analyze the Hawkeye data of tennis serves to study the optimal landing locations for aces. They found three key elements related to serve aces: direction relative to the returner, closeness to the lines, and speed. De Leeuw et al. [8] used subgroup discovery to find the characteristics of won service points for a specific professional tennis player. They found that more points were won if the player avoided hitting a backhand after the serve. These two studies are not relevant to our work on predicting serve directions. Now we will look at serves from a returnerâs perspective. Returning first serves in professional tennis is one of the most difficult tasks in sports, and yet professional players have been able to return most first serves. Many analysts do not believe that fast physical reaction alone is enough to explain the many successful returns because the reaction time is less than 0.5 second. These players must have learned to read the serves with reasonable accuracy. However, the exact nature of this anticipatory behavior is still unknown [3]. The most widely examined source of anticipatory information has been the kinematics of serve motion. In other words, a returner might be able to predict the serve directions from reading the serve motion before the ball is hit. But professional players are trained to disguise their serve directions by maintaining the same serve motion. Therefore, reading the kinematics of serve motion is difficult. Some studies showed that contextual information might be useful for predicting the serve directions [19, 18]. Our work is related to this subject because our results suggest that it is possible to make reasonably accurate predictions solely based on contextual information. 3 Basic Information about Tennis Serves A tennis match is divided into sets, games, and points. A tennis court is laterally divided into two sides: the deuce side and the ad side. The two players take turns serving for each game. The serving player serves from the deuce side and ad side alternatively. At the start of each point, the serving player has two chances to serve. If the first serve fails, the player can make a second serve. A player usually makes a faster but more risky first serve and a slower but safer second serve. The player can serve toward anywhere in the service box, but there are generally three directions: wide, body (toward the returner), and down-the-T (toward the middle of the court). In this study, we only consider the first serves because the first serves give the server a significant advantage. Therefore, in the discussions below, the word âserveâ means âfirst serveâ by default. 4 Data We used the data from the Match Charting Project [15]. This open-source project provides detailed point-by-point and shot-by-shot data for thousands of professional tennis matches. Unlike the Hawkeye data, this data set is created by a group of volunteers watching tennis match videos and manually entering coded shot-by-shot data, including the serve directions and outcomes. An Excel script then derives additional information from the shot-by-shot data, such as the score, who is serving, and rally length for each point. The human-coded serve direction for each point is the ground truth for our model training and testing. The data set we analyzed contains 3424 matches of 655 male players and 1916 matches of 422 female players. However, most players only have one or a few matches in the database. Therefore, we only run the analysis for a selected group of players with at least 30 matches. We processed and analyzed the data using Python-based tools such as Pandas [9], Sklearn [13], SciPy [20], etc. The original data set contains errors, such as missing values in match data, duplicate or incorrect match IDs, match IDs in the point data set but are not in the match data set, and data entry errors in some shot-by-shot codes. So we spent a lot of time on data cleaning and transformation. The original data set primarily contains the scores and shot-by-shot descriptions for each match. Several previous works [16, 1] also used the Match Charting data, but they did not use machine learning. 5 Feature Engineering The performance of a machine learning program depends on the selected features. We went through an iterative process of extracting, selecting, and testing features. In addition to the original features in the Match Charting data set, we also derived many new features from the original point-by-point and shot-by-shot data set. For example, we calculated the number of serves a player made toward each direction, identified critical points, calculated how many shots a player had played before each point, and estimated how much a player had run in the last point, etc. Many features were tested and rejected. The features discussed in this paper are the ones that currently generate the highest prediction accuracy. We are still working on feature engineering. New features may be added, and some of the existing features may be modified or removed in the future. We try to select features likely to influence a playerâs serve decisions. In addition to predicting serve directions, we also want to see if certain features are more important in making such decisions, which may help us gain insights into the playersâ decision-making process. We will discuss each of the selected features in the subsections below. 5.1 Outcome of Previous Points We believe that the outcome of previous points would influence the selection of serve directions. We assume that a professional player would have a rough idea of how many points he or she has served toward each direction and how many points are won. This is the assumption made by several previous works [1, 4, 21]. The analysis by Spiliopoulos [16] also showed the serial dependency of serve directions on the previous pointâs serve direction and outcome. The following features are calculated and used in our machine learning model. ⢠For each server and each point, our program calculates the count of the serves made toward each direction, from the beginning of the match to the previous point (3 features). These parameters are similar to the âprior styleâ parameters in Wei et al. [22], but we only use three directions for the deuce or ad side while Wei et al. used 7 directions for each side. Although a player may not be able to remember the exact count of serves made toward each direction, the player should have a rough idea of the counts for the last several service games. Therefore, a variation of this feature is the counts of serve directions from a certain number of service games prior to the current point rather than from the beginning of the match. But it is not easy to determine how many prior service games should be considered. ⢠For each server and each point, our program also calculates the count of serves the server made toward each direction and won, from the beginning of the match to the previous point (3 features). Similarly, a variation of this feature is to count only for a certain number of service games before the current point, not from the beginning of the match. ⢠For each point, the program calculates the percentage of âgoodâ first serves the server made toward each direction, from the beginning of the match to the previous point (3 features). These are the so-called âserve percentageâ for each direction. A professional player should have a reasonably accurate understanding of their current serve percentage. ⢠For each point, the program records the winner of the previous point (1 feature). 5.2 Fatigue As the match progresses, both players become more and more tired. We want to examine whether the level of fatigue is a factor in choosing serve directions. It is reasonable to assume that a player will exploit the opponentâs fatigue in choosing serve directions. For example, serving wide is likely to make an opponent run more because a wide serve opens up the court more. The playerâs own fatigue may also affect serve directions. Because the net is lower in the middle, serving to the T may require less jumping. The following features are used to estimate fatigue in our machine learning model. ⢠For each point, our program calculates the cumulative run indexes for two players from the beginning of the match to the previous point (2 features). They indicate how tired each player is from the running and hitting before each serve. Because the Match Charting data [15] contains detailed shot-by-shot information, including the shot type (e.g., forehand, backhand, slice, volley, overhead), shot direction (i.e., to-deuce-side, to-middle, or to-ad-side), and the depth of each shot (e.g., shallow and deep), our program can infer the playerâs court position when they hit a particular shot. Based on that information, the program can estimate how much a player ran for each point and calculate a ârun index.â It is more accurate than the shot (rally) count because it includes running. This run index is not as accurate as the Hawkeye data for measuring running distance, but it is a consistent estimate from point to point. Since Hawkeye data are not publicly available, it is difficult to get more accurate measures. 5.3 Performance Anxiety Because a tennis match does not have a time limit, scoreboard pressure is the primary source of a playerâs performance anxiety. Such anxiety could influence a playerâs decisions. For this reason, Wei et al. [22] used scores in their machine learning model. But our method is different. Instead of using scores, we calculate an index of each playerâs performance anxiety. Our work is based on the OCC model [11] of emotion, which is the standard model in affective computing. Based on the OCC model, anxiety is influenced by hope, fear, and uncertainty. A playerâs feeling of uncertainty is related to the gap between the two playersâ scores. The smaller the gap, the higher the uncertainty. If the score is tied, the uncertainty is the highest. If one player is very close to winning, the uncertainty is very low. Due to tennisâ hierarchical score structure, there are three levels of uncertainty. The gap between the set scores influences the match-level uncertainty. The gap between the game scores influences the set-level uncertainty. The gap between the point scores influences the game-level uncertainty. A playerâs feeling of hope depends on how close the player is to winning. If a playerâs score is close to the winning score, the playerâs hope is high. Again, there are three levels of hope: match-level hope, set-level hope, and game-level hope. A playerâs feeling of fear depends on how close the player is to losing. There are three levels of fear: match-level fear, set-level fear, and game-level fear. For example, if a player leads by a significant margin and serves for the set point, the playerâs hope is high, fear is low, and uncertainty is low, resulting in a relatively low level of anxiety. On the other hand, if the scores are very close near the end of a match, such as in the final set tiebreak, each player will have high hope, high fear, and high uncertainty, resulting in high levels of anxiety for both players. In our model, a performance anxiety index is calculated separately on the game, set, and match level based on the following equation (3 features): â=â(â+)performance\ anxiety=uncertainty*(hope+fear) (1) The overall performance anxiety is the sum of game, set, and match level anxiety indices (1 feature): â=â+â+ââoverall\ anxiety=game\ anxiety+set\ anxiety+match\ anxiety (2) In the equations, we do not consider fear to be the negative of hope. Therefore, fear does not reduce hope, and vice versa. This is because hope and fear can coexist in most situations. For example, in a close tiebreak game, a player is close to both victory and defeat at the same time. In such cases, both strong hope and strong fear can coexist. 5.4 Other Features We also considered other factors that may influence the serve decisions, such as court surfaces [6] and the opponentâs handedness. For example, a player might serve differently against a lefty opponent. 6 Machine Learning We ranked the players by the number of matches they have in the data set and analyzed players with at least 30 matches in the data set. Due to the space limit, we only present the results for ten male players and ten female players. The players are selected based on their current ranking and significant achievements. But the results for other players are generally consistent with those presented here. We applied the following machine learning models to our data set: Multinomial Logistic Regression, Decision Tree, Random Forest, Support Vector Machine (Multiclass Classification), and Neural Network. We also applied Bagging classifier, Ada Boost classifier, and XGBoost classifier, but the results are no better than the models mentioned above, so we do not present their results here. For Random Forest, we used 200 trees with a maximum depth of 150. For the Bagging classifier, we used 50 estimators. For the Ada Boost classifier, we used 70 estimators. For the XGBoost classifier, the K value is 10. For the neural network, we use sklearn.neural_network.MLPClassifier() with two hidden layers (200, 100). We train our models individually for each selected player. For each player, we randomly split the data into a training set (70%) and a testing set (30%), and we use the same training set and testing sets for the different machine learning models. For each player, our program selects all the points that this player serves, and predicts the first serve direction using the features discussed in Section 5. The prediction is then compared with the actual first serve direction coded by the person who entered the data for the Match Charting project. The prediction accuracy is calculated based on all the points in the testing data set. The results of the analysis are presented in Tables 1â4. In the tables, LR stands for Multinomial Logistic Regression, DT stands for Decision Tree, RF stands for Random Forest, SVM stands for Support Vector Machine, and N stands for Neural Network. Table 1: First Serve Direction Prediction Accuracy for the Deuce Side Serves (men) First name Last name LR RF DT SVM N MEAN Novak Djokovic 0.47 0.50 0.45 0.46 0.47 0.47 Roger Federer 0.46 0.49 0.44 0.47 0.47 0.47 Nick Kyrgios 0.55 0.52 0.50 0.55 0.54 0.53 Daniil Medvedev 0.45 0.49 0.48 0.47 0.46 0.47 Andy Murray 0.56 0.53 0.58 0.52 0.54 0.55 Rafael Nadal 0.52 0.50 0.43 0.52 0.52 0.50 Dominic Thiem 0.55 0.47 0.43 0.55 0.55 0.51 Stefanos Tsitsipas 0.49 0.48 0.46 0.47 0.47 0.47 Stan Wawrinka 0.46 0.49 0.45 0.45 0.48 0.46 Alexander Zverev 0.46 0.42 0.42 0.47 0.46 0.44 MEAN 0.50 0.49 0.46 0.49 0.50 0.49 Table 2: First Serve Direction Prediction Accuracy for the Ad Side Serves (men) First name Last name LR RF DT SVM N MEAN Novak Djokovic 0.49 0.50 0.44 0.49 0.48 0.48 Roger Federer 0.56 0.52 0.53 0.59 0.56 0.55 Nick Kyrgios 0.50 0.49 0.45 0.48 0.48 0.48 Daniil Medvedev 0.53 0.51 0.53 0.49 0.60 0.53 Andy Murray 0.48 0.45 0.44 0.48 0.47 0.46 Rafael Nadal 0.54 0.51 0.45 0.54 0.54 0.52 Dominic Thiem 0.58 0.54 0.52 0.58 0.58 0.56 Stefanos Tsitsipas 0.47 0.45 0.42 0.48 0.47 0.46 Stan Wawrinka 0.49 0.49 0.43 0.49 0.47 0.47 Alexander Zverev 0.45 0.47 0.44 0.44 0.44 0.45 MEAN 0.51 0.49 0.47 0.51 0.51 0.50 Table 3: First Serve Direction Prediction Accuracy for the Deuce Side Serves (women) First name Last name LR RF DT SVM N MEAN Victoria Azarenka 0.47 0.47 0.40 0.49 0.48 0.46 Ashleigh Barty 0.51 0.45 0.45 0.51 0.48 0.48 Angelique Kerber 0.37 0.37 0.33 0.40 0.36 0.36 Anett Kontaveit 0.41 0.43 0.39 0.39 0.39 0.40 Garbine Muguruza 0.47 0.45 0.40 0.47 0.45 0.45 Naomi Osaka 0.49 0.41 0.41 0.48 0.45 0.45 Karolina Pliskova 0.44 0.45 0.40 0.43 0.44 0.43 Maria Sakkari 0.44 0.44 0.37 0.47 0.47 0.44 Iga Swiatek 0.41 0.43 0.36 0.43 0.44 0.41 Serena Williams 0.48 0.49 0.44 0.49 0.49 0.48 MEAN 0.45 0.44 0.40 0.46 0.45 0.44 Table 4: First Serve Direction Prediction Accuracy for the Ad Side Serves (women) First name Last name LR RF DT SVM N MEAN Victoria Azarenka 0.46 0.41 0.34 0.47 0.43 0.42 Ashleigh Barty 0.51 0.46 0.45 0.46 0.46 0.47 Angelique Kerber 0.61 0.58 0.48 0.61 0.60 0.58 Anett Kontaveit 0.43 0.42 0.42 0.43 0.42 0.42 Garbine Muguruza 0.39 0.40 0.38 0.37 0.40 0.39 Naomi Osaka 0.44 0.38 0.36 0.45 0.47 0.42 Karolina Pliskova 0.43 0.42 0.42 0.46 0.44 0.43 Maria Sakkari 0.50 0.51 0.45 0.50 0.53 0.50 Iga Swiatek 0.49 0.40 0.37 0.49 0.43 0.44 Serena Williams 0.47 0.47 0.45 0.47 0.48 0.47 MEAN 0.47 0.44 0.41 0.47 0.47 0.45 We also calculated the serve direction distributions for each player. Our results are similar to those reported by Tea and Swartz [17]. The serve direction distributions vary from player to player. For example, Djokovicâs serve directions are more evenly distributed, while Federer tended to serve much less to the body. 7 Discussion From Tables 1 and 2, we can see that our machine learning models achieved an average 49% prediction accuracy for the deuce side serve directions and 50% accuracy for the ad side serve directions for the selected male players. Adding other male players will bring the average percentage slightly lower to around 48%. From Tables 3 and 4, we can see that our machine learning models achieved an average 44% prediction accuracy for the deuce side serve directions and 45% accuracy for the ad side serve directions for the selected female players. Adding other female players will bring the average percentage slightly lower to around 43%. From the tables, we can also see that prediction accuracy is generally consistent among different machine learning methods. We found only one published work by Wei et al. [22] that reported a serve direction prediction accuracy (27.8%). However, it is difficult to compare our prediction accuracy with theirs because Wei et al. used seven serve directions per side while we used the more traditional three directions per side. This is because we based our analyses on different ground truths. Wei et al. used the Hawkeye data as ground truth, and they could divide the serve directions into smaller groups. We used human-observed serve directions as our ground truth, and our data only has three serve directions per side. The features we used are also quite different from the features used by Wei et al. We conducted a feature importance analysis for the Decision Tree model. We found that the most important features are the cumulative counts of first serves made to each direction, the run index of the server in the previous point, and the first serve percentage for each direction. While the prediction accuracy varies for each player, these three features are consistently among the most important. This provides some indirect evidence that they might also be the important factors a player considers when choosing serve directions. The importance of cumulative counts of first serve directions and first serve percentage are consistent with the mixed-strategy findings by Walker and Wooders [21], Spiliopoulos [16], and Gauriot et al. [4]. But as far as we know, the importance of server fatigue (run index) in choosing the serve direction has not been discussed in previous work. Finally, our results show that it is possible to achieve reasonably accurate prediction of serve directions solely based on contextual information such as the outcome of previous serves, performance anxiety, and fatigue. This may help explain why professional players are able to return most of the very fast first serves. Many analysts do not believe that fast physical reaction alone is enough to explain the many successful first serve returns because a returner must react to a first serve in less than 0.5 seconds. These players must have learned to read the serves with reasonable accuracy. Although the exact nature of this anticipatory behavior is still unclear [3], the most widely examined source of anticipatory information has been the kinematics of serve motion. In other words, a returner might be able to predict the serve directions from reading the serve motion before the ball is hit. But professional players are trained to disguise their serve directions by maintaining the same serve motion, making it difficult to read. Some studies showed that contextual information might be useful for predicting the serve directions [19, 18]. Our work suggests that contextual information is perhaps more important for returnersâ anticipatory reactions than previously thought. 8 Conclusion and Future Work We have described our machine learning methods for predicting professional tennis playersâ first serve directions. Through feature engineering, our method achieves an average prediction accuracy of around 49% for male players and 44% for female players. Our feature importance analysis provides some indirect evidence that the top professional players seem to use a mixed-strategy model in choosing serve directions, which is consistent with some previous works [21, 16, 4]. However, the importance of server fatigue in choosing the serve direction has been a new discovery. Our work also suggests that contextual information is perhaps more important for returnersâ anticipatory reactions than previously thought. We are continuing our work on feature engineering to improve prediction accuracy. We will also test using Brier Score as a measurement of prediction accuracy. We also plan to apply our method to applications in other highly competitive situations. References Anderson et al. [2021] Anderson, A., Rosen, J., Rust, J., & Wong, K.P. (2021). Disequilibrium play in tennis. Working papers, Georgetown University. https://ideas.repec.org/p/geo/guwopa/gueconwpa~21-21-07.html ATP [2022] Association of Tennis Professionals (ATP). (2022). ATP Stats. https://w.atptour.com/en/stats AvilĂŠs et al. [2019] AvilĂŠs, C., Navia, J.A., Ruiz, L.M., & MartĂnez de Quel, O. (2019). Do expert tennis players actually demonstrate anticipatory behavior when returning a first serve under representative conditions? A systematic review including quality assessment and methodological recommendations. Psychology of Sport and Exercise, 43, 16â26. Gauriot et al. [2020] Gauriot, R., Page, L., & Wooders, J. (2020). Expertise, gender, and equilibrium play. Tech. rep., New York University Abu Dhabi. http://w.johnwooders.com/papers/Expertise,Gender,andEquilibrium.pdf Hsu et al. [2007] Hsu, S.H., Huang, C.Y., & Tang, C.T. (2007). Minimax play at Wimbledon: Comment. The American Economic Review, 97, 517â523. Hughes & Clarke [2021] Hughes, M., & Clarke, S. (2021). Surface effect on elite tennis strategy. In T. Reilly, M. Hughes, & A. Lees (Eds.), Science and Racket Sports I (p. 288â294). Taylor & Francis. Kovalchik & Reid [2018] Kovalchik, S., & Reid, M. (2018). A shot taxonomy in the era of tracking data in professional tennis. Journal of Sports Sciences, 36, 2096â2104. de Leeuw et al. [2020] de Leeuw, A.W., Hoekstra, A., Meerhoff, L., & Knobbe, A. (2020). Tactical analyses in professional tennis. In Proceedings of the Joint European Conference on Machine Learning and Knowledge Discovery in Databases (p. 258â269). Springer. McKinney [2010] McKinney, W. (2010). Data structures for statistical computing in Python. In Proceedings of the 9th Python in Science Conference (p. 56â61). Mecheri et al. [2016] Mecheri, S., Rioult, F., Mantel, B., Kauffmann, F., & Benguigui, N. (2016). The serve impact in tennis: First large-scale study of big Hawk-Eye data. Statistical Analysis and Data Mining, 9, 310â325. Ortony et al. [1988] Ortony, A., Clore, G.L., & Collins, A. (1988). The Cognitive Structure of Emotions. Cambridge University Press. OâDonoghue & Brown [2008] OâDonoghue, G.P., & Brown, E. (2008). The importance of service in grand slam singles tennis. International Journal of Performance Analysis in Sport, 8, 70â78. Pedregosa et al. [2011] Pedregosa, F., et al. (2011). Scikit-learn: Machine learning in Python. Journal of Machine Learning Research, 12, 2825â2830. Rioult et al. [2015] Rioult, F., Mecheri, S., Mantel, B., Kauffmann, F., & Benguigui, N. (2015). What can Hawkeye data reveal about serve performance in tennis? In Proceedings of the Machine Learning and Data Mining for Sports Analytics Workshop (ECML/PKDD) (p. 36â45). Sackmann [2022] Sackmann, J. (2022). The Match Charting Project. Available at http://w.tennisabstract.com/ (accessed 2022/06/19). Spiliopoulos [2016] Spiliopoulos, L. (2016). Randomization and serial dependence in professional tennis matches: Do strategic considerations, player rankings and match characteristics matter? Judgment and Decision Making, 13, 413â427. Tea & Swartz [2022] Tea, P., & Swartz, T.B. (2022). The analysis of serve decisions in tennis using Bayesian hierarchical models. Annals of Operations Research, 2021, 1â16. Vernon [2020] Vernon, G. (2020). Decision making in tennis: Exploring the use of kinematic and contextual information during anticipatory performance. Ph.D. thesis, Victoria University, Australia. Vernon et al. [2018] Vernon, G., Farrow, D., & Reid, M. (2018). Returning serve in tennis: A qualitative examination of the interaction of anticipatory information sources used by professional tennis players. Frontiers in Psychology, 9, 895. Virtanen et al. [2020] Virtanen, P., et al. (2020). SciPy 1.0: Fundamental algorithms for scientific computing in Python. Nature Methods, 17, 261â272. Walker & Wooders [2001] Walker, M., & Wooders, J. (2001). Minimax play at Wimbledon. The American Economic Review, 91(5), 1521â1538. Wei et al. [2015] Wei, X., Lucey, P., Morgan, S., Carr, P., Reid, M., & Sridharan, S. (2015). Predicting serves in tennis using style priors. In Proceedings of the 21st ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (p. 2207â2215). ACM. Whiteside & Reid [2016] Whiteside, D., & Reid, M. (2016). Spatial characteristics of professional tennis serves with implications for serving aces: A machine learning approach. Journal of Sports Sciences, 35, 648â654.