Comparing modern machine learning approaches and different forecast strategies on Grand Slam tennis tournaments

In this article, different modern machine learning and regression approaches for modeling and prediction of tennis matches in Grand Slam tournaments are investigated. Our data include information on 5013 matches in men's ssGrand Slam tournaments from the period 2011 to 2022. The investigated methods focus on modeling the probability of the first-named player to win the respective match. Moreover, different features are considered including the players' age, the ATP ranking and points, bookmakers' odds, Elo rating, and two additional age variables, which take into account the optimal age of a tennis player. We compare the different regression approaches to modern machine learning approaches with respect to various performance measures. Moreover, we also investigate different forecast strategies. First of all, a cross-validation-type strategy for all matches between 2011 and 2021. We also use an “expanding window” strategy by continuously updating the training data to analyze the predictive performance of the approaches on the tournaments from 2022. Finally, a “rolling window” strategy is used with only 3 years of tournaments as training data. We then select small subsets of best models with largest average ranks and investigate those in more detail by the help of interpretable machine learning techniques.