Multi-Source Point-by-Point Dataset for Tennis Match Outcome Prediction

This dataset is a comprehensive and high-quality resource tailored for tennis match outcome prediction research, integrating multi-source data spanning nearly three decades. Its data sources are diverse and authoritative, including official match records from ATP (Association of Tennis Professionals) and WTA (Women's Tennis Association), digitized historical archives (converted from unstructured materials like paper score sheets via OCR and manual proofreading), fine-grained features extracted through video analysis of key matches (e.g., Grand Slam finals), and high-precision sensor data from third-party partners such as Hawk-Eye. The dataset adopts a point-by-point recording granularity, with core technical features including PointWinner (the winner of each point), Speed_KMH (ball speed), P1DistanceRun and P2DistanceRun (running distances of Player 1 and Player 2), ServeWidth (serve direction, categorized into 5 types), and ServeDepth (serve depth). It contains a total of 20,562 records, which are divided into a training set (16,450 records) and a test set (4,112 records) using a time-series-aware subsampling strategy—matches are split into three chronological phases, with 80% random sampling in each phase to preserve temporal continuity. Prior to model training, the data undergoes strict preprocessing: missing values and outliers are checked and handled (none were found in this dataset), and continuous features like ball speed and running distance are standardized using Z-score to ensure uniform feature weights. This dataset effectively captures dynamic match characteristics and player technical performance, providing solid support for training and validating tennis match prediction models (e.g., the XGBoost model optimized by NSGA-II), and can also be used for in-depth analyses of critical points and score trends.