Tennis match outcome prediction using temporal directed graph neural networks

We present the first application of a graph neural network for tennis match outcome prediction. Using MagNet, an existing spectral graph neural network for directed graphs, we construct temporal directed graphs by representing players as nodes and surface-specific historical match outcomes as edges. The model is trained and evaluated using a dataset of Grand Slam, ATP Masters 1000, and two ATP 500 events from 2007 to the conclusion of the US Open in September 2024. Following hyperparameter optimisation, a tuned model on the out-of-sample data achieves comparable predictive accuracy (66.0%) to the benchmark weighted Elo rating system (65.6%). Many recent advancements in tennis match prediction have focused on incremental improvements to the Elo rating system, such as incorporating margin of victory and surface-specific adjustments. Our research shifts this paradigm by demonstrating that graph neural networks, which inherently capture complex relational and temporal dynamics, offer a powerful alternative for pairwise comparison tasks such as tennis match prediction.