A comparative evaluation of machine learning and deep learning for sports predictions in table tennis

Abstract

The surge in sports data availability has opened avenues for exploring the application of machine learning (ML) and deep learning (DL) techniques in predicting match outcomes as classification tasks across various sports. While mainstream sports have garnered considerable attention in this regard, with a plethora of ML and DL techniques being explored, there remains a lack of research on niche sports like table tennis. In contrast to mainstream sports, which often have limited data availability due to their shorter seasons, table tennis emerges as an ideal yet unexplored candidate due to its significant volume of match data, enabling a meaningful comparison between ML and DL techniques. This study bridges this research gap by employing a range of ML and DL models to predict table tennis match outcomes and explores their financial implications in the sports betting domain, and subsequently extends the body of knowledge on the ML vs DL debate on tabular data. Leveraging an extensive dataset of 123,107 table tennis matches, various ML and DL classification models were evaluated, including techniques novel to this domain. Results indicate that while ML models outperform complex DL models, feed-forward neural networks show comparable performance and emerge as promising candidates, especially when considering the benefits of transferable player embeddings to ML models. By assessing both algorithmic and monetary performance, this research offers a comprehensive understanding of predictive modeling in table tennis matches, laying a solid foundation for future studies. The monetary results demonstrate state-of-the-art returns, further validating the efficacy of the approach.