Predicting tennis match outcomes mid-game using machine learning on psychological and physical data

Le résumé fourni par la source

Predicting game outcomes has significantly garnered the interest of researchers in recent years. The role of player performance is integral in-game analytics, impacting the interpretation and results of the analysis. Our work presents an AI for Science (AI4Sci) method to use real-time data from each game point to determine essential feature values, formulate and assess the impact of psychological momentum, and employ machine learning methodology on mid-match data for predicting the game’s victor. The data source is from Wimbledon and US Open games from 2017 to 2022, a total of 1592 games, and utilize 363 games of 2023 to evaluate their forecasting ability. We first obtained weights through information entropy and defined psychological momentum, and then 3 best classifiers, random forest, CatBoost, and Logistic Regression, were detected to assess the features. Additionally, we implemented a soft voting ensemble method integrating the Random Forest and CatBoost classifiers. All four models achieve over 90% accuracy and F1-score, with the soft voting classifier performing the best (accuracy: 97.5%, F1 score: 97.4%). These models achieve predictive accuracies above 70% using the first 25% data of a game.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.