Tennis-Prediction

The goal of this project is to predict the outcome of a tennis match using the data of both players and ML models.
The data used comes from Jeff Sackmann's repository.

• Installation
• Usage
  • Data Loading
  • Machine Learning modelling
  • Encoding Matches
• License

Installation

To clone the repository, with the data you need to also clone the submodules:

git clone --recurse-submodules https://github.com/VincentAuriau/Tennis-Prediction.git

Usage

You can find examples in /examples:

Data Loading

Loading players statistics at match time and match outcome: Example

from data.data_loader import matches_data_loader
data_df = matches_data_loader(path_to_data="submodules/tennis_atp")

data_df contains let you access information about players (statistics prior to the match) along statistics of the match. A basic example statistic: the victory percentage of the best ranked player in a match, depending on players rankings.

Number of ATP main matches depending on players rank	Victory % of best ranked player

It can be easily used to also compute players statistics over their carreer, and/or at match time. Here is a simple example with Stan Wawrinka:

Stan's Victory % in main ATP matches	Stan's career aces % diff with adversary

Here is an example of a data row:

id	tournament	tournament_level	tournament_date	tournament_surface	round	best_of	match_id	Winner	Score
atp_matches_qual_chall_2003_5427	San Benedetto CH	C	20030811	Clay	SF	3	20030811	0	2-6 7-5 7-5

Base Match Statistics:

• id and match_id: unique identification of the matches
• tournament: name of the tournament
• tournament_level: Category of the tournament 'G' (Grand Slams) 'M' (Masters 1000s), 'A' (other tour-level events), 'C' (Challengers), 'S' (Satellites/ITFs), 'F' (Tour finals), 'D'( Davis Cup)
• tournament_date: date
• tournament_surface: surface 'Grass', 'Clay', 'Carpet', 'Hard'
• round: tournament round of the match 'F' (finale), 'SF' (semi-finale),
• best_of: number of won games needded (3 or 5)
• Winner: index of the winner: 0 (Player1) or 1 (Player2)
• Score: finale score

Additional match statistics:

• elapsed_minutes: Duration of the match
• aces_nb_x: Number of aces of player x
• doublefaults_nb_x: Number of doublefaults
• svpt_x: Number of serve points
• 1stIn_1: Number of first serves made
• 1stWon_1: Number of first-serve points won
• 2ndWon_x: Number of second-serve points won
• SvGms_1: Number of serve games
• bpSaved_1: Number of break points saved
• bpFaced_1: Number of break points faced

Example of match statistics:

Name_1	ID_1	Ranking_1	Ranking_Points_1	Ranking_History_1	Best_Rank_1	Birth_Year_1	Versus_1	Hand_1	Last_Tournament_Date_1	Height_1	Matches_1	Matchs_Clay_1	Matches_Carpet_1	Matches_Grass_1	Matches_Hard_1	Victories_Percentage_1	Clay_Victories_Percentage_1	Carpet_Victories_Percentage_1	Grass_Victories_Percentage_1	Hard_Victories_Percentage_1	Aces_Percentage_1	Doublefaults_Percentage_1	First_Save_Success_Percentage_1	Winning_on_1st_Serve_Percentage_1	Winning_on_2nd_Serve_Percentage_1	Overall_Win_on_Serve_Percentage_1	BreakPoint_Face_Percentage_1	BreakPoint_Saved_Percentage_1	last_rankings_1	last_ranking_points_1
Stan.Wawrinka	104527	184	114	{20030616: [387, 68], 20030707: [363, 74], 20030714: [348, 79], 20030721: [303, 99], 20030811: [284, 114]}	284	19850328	[]	R	20030721	183	[['V', 'atp_matches_qual_chall_2003_3466'], ['D', 'atp_matches_qual_chall_2003_3481'], ['D', 'atp_matches_2003_4049'], ['V', 'atp_matches_2003_4315'], ['D', 'atp_matches_2003_4328'], ['V', 'atp_matches_2003_4773'], ['D', 'atp_matches_2003_4782'], ['V', 'atp_matches_qual_chall_2003_5408'], ['V', 'atp_matches_qual_chall_2003_5419'], ['V', 'atp_matches_qual_chall_2003_5424']]	['V', 'D', 'D', 'V', 'D', 'V', 'D', 'V', 'V', 'V']	[]	[]	[]	60	60	0	0	0	3.41880341880342	4.27350427350427	64.957264957265	54.985754985755	15.6695156695157	70.6552706552707	11.3960113960114	7.69230769230769	[303, 9999, 9999, 9999, 9999, 9999, 9999, 9999, 9999, 9999, 9999, 387]	[99, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 68]

Name_2	ID_2	Ranking_2	Ranking_Points_2	Ranking_History_2	Best_Rank_2	Birth_Year_2	Versus_2	Hand_2	Last_Tournament_Date_2	Height_2	Matches_2	Matchs_Clay_2	Matches_Carpet_2	Matches_Grass_2	Matches_Hard_2	Victories_Percentage_2	Clay_Victories_Percentage_2	Carpet_Victories_Percentage_2	Grass_Victories_Percentage_2	Hard_Victories_Percentage_2	Aces_Percentage_2	Doublefaults_Percentage_2	First_Save_Success_Percentage_2	Winning_on_1st_Serve_Percentage_2	Winning_on_2nd_Serve_Percentage_2	Overall_Win_on_Serve_Percentage_2	BreakPoint_Face_Percentage_2	BreakPoint_Saved_Percentage_2	last_rankings_2	last_ranking_points_2
Martin.Vassallo Arguello	103506	125	296	{19990201: [817, 13], 20000710: [398, 61], 20000731: [354, 75], 20000807: [377, 70], 20010625: [459, 48], 20010709: [405, 61], 20010813: [391, 68], 20010820: [374, 72], 20010827: [342, 88], 20010917: [291, 117], 20010924: [286, 122], etc...}	123	19800210	[]	R	20030804	183	[['V', 'atp_matches_qual_chall_1999_380'], ['D', 'atp_matches_qual_chall_1999_393'], ['V', 'atp_matches_qual_chall_2000_3972'], ['V', 'atp_matches_qual_chall_2000_3988'], ['D', 'atp_matches_qual_chall_2000_3996'], ['D', 'atp_matches_qual_chall_2000_4725'], ['D', 'atp_matches_qual_chall_2000_4758'], ['V', 'atp_matches_qual_chall_2001_3699'], etc...]	['V', 'D', 'V', 'V', 'D', 'V', 'V', 'D', 'V', 'V', 'D', 'V', 'D', 'V', 'V', 'D', 'V', 'V', 'V', 'D', 'V', 'D', 'V', 'V', 'V', 'V', 'D', 'D', 'V', 'V', 'V', 'D', 'V', 'D', 'V', 'D', 'D', etc...]	['D', 'V', 'D', 'V', 'D']	['D']	['D', 'D', 'D', 'V', 'V', 'V', 'V', 'D', 'V', 'D', 'V', 'D']	61.0294117647059	63.5593220338983	40	0	50	4.82456140350877	5.26315789473684	61.4035087719298	46.4912280701754	18.859649122807	65.3508771929825	9.64912280701754	5.70175438596491	[157, 136, 165, 158, 9999, 9999, 204, 198, 197, 188, 204, 233]	[247, 304, 220, 232, 0, 0, 164, 172, 177, 188, 167, 137]

Player statistics before the match:

• Name_x: Name of the player
• ID_x: ID of the player
• Ranking_x: ATP ranking of the player. For all rankings, 9999 means unranked.
• Ranking_Points_x: Number of ATP points
• Ranking_History_x: All recorded rankings
• Best_Rank_x: Best reached ATP rank
• Birth_Year_x: Birth year
• Versus_x: Dictionnary containing all match outcomes agains other players
• Hand_x: Hand used to play 'R', 'L' or 'U' for unknown
• Last_Tournament_Date_x: Date of the last previous tournament attended
• Height_x: Height
• Matches_x: History of outcomes of previous matches
• Matchs_Clay_x: History of outcomes of previous matches on clay
• Matchs_Carpet_x: History of outcomes of previous matches on carpet
• Matchs_Grass_x: History of outcomes of previous matches on grass
• Matchs_Hard_x: History of outcomes of previous matches on hard
• Victories_Percentage_x: Victory percentage over all player ATP matches
• Clay_Victories_Percentage_x: Victory percentage over all player ATP matches on clay
• Carpet_Victories_Percentage_x: Victory percentage over all player ATP matches on carpet
• Grass_Victories_Percentage_x: Victory percentage over all player ATP matches on grass
• Hard_Victories_Percentage_x: Victory percentage over all player ATP matches on hard
• Aces_Percentage_x: Aces percentage over all player ATP matches
• Doublefaults_Percentage_x: DoubleFaults percentage over all player ATP matches
• First_Save_Success_Percentage_x: First save success percentage over all player ATP matches
• Winning_on_1st_Serve_Percentage_x: Winning on first serve percentage over all player ATP matches
• Winning_on_2nd_Serve_Percentage_x: Winning on second serve percentage over all player ATP matches
• Overall_Win_on_Serve_Percentage_x: Overall winning percentage on serve over all player ATP matches
• BreakPoint_Face_Percentage_x: Overall breakpoint face percentage over all player ATP matches
• BreakPoint_Saved_Percentage_x: Overall breakpoint saved percentage over all player ATP matches
• last_rankings_x: Five previous recorded ATP rankings
• last_ranking_points_x: Five previous ATP ranking points recorded

Machine-Learning modelling

Train/Testing on matches outcome [Example].

A generic function lets you evaluate your model with a train/test scheme without much work. Your model only needs a scikit-learn like signature. By playing with the years, columns to use in modelling and models & hyperparmaters, you can easily create your own best-performing model.

from sklearn.ensemble import RandomForestClassifier
from evaluation.train_test import train_test_evaluation

test_score = train_test_evaluation(
    train_years=[2020, 2021],
    test_years=[2022, 2023],
    model_class=RandomForestClassifier,
    model_params={"n_estimators": 2000, "max_depth": None},
    match_features=[],
    player_features=["Ranking"],
    encoding_params={},
    additional_features=[],
    save_path="./results",
    save_all_results=False
)

print("Test Score", test_score)

Models and hyperparamters can easily be compared with the file results.csv saved in save_path.

Accuracy of different models

If the argument save_all_results is set to True, the whole csv of test data is saved. It helps to get more in-depth analysis of results

Model precision compared with best ranked player wins strategy	Model precision depending of players ranks

Encoding matches

Example In order to represent history of a player, one can use MatchEncoders:

from history_modeling.encoding_model import PCAMatchEncoder

model = PCAMatchEncoder(num_pca_features=2)
model.fit(data_df, transform_data=True)
X_r, match_info = model.predict(data_df, transform_data=True)

2D representation of match outcome:

License

The repository is under the MIT License, you can freely use any part as you like.
If you find this repository useful, you can cite it and add a star ⭐ !