Intransitive Player Dominance and Market Inefficiency in Tennis Forecasting: A Graph Neural Network Approach

Considerable effort has also been devoted to developing highly accurate models for forecasting match outcomes (Wunderlich and Memmert, 2021). Tennis is a sport well-suited to predictive modelling, with dense tournament schedules generating extensive historical data. The official ranking systems of the Association of Tennis Professionals (ATP) and Women's Tennis Association (WTA) have been shown to exhibit some predictive power for match outcomes (Clarke and Dyte, 2000; Klaassen and Magnus, 2003), but there are notable limitations: for example, ranking points accumulate over a 52-week period, without decay, which can mask recent changes in player form; while match-specific factors, such as surface type, tournament progression difficulty, and margin of victory in individual matches, are overlooked. Some well-known methods have been applied to tennis and modified to accommodate these factors, such as a Bradley-Terry model with surface-specific adjustments (McHale and Morton, 2011) or Elo rating systems that incorporate margin of victory (Kovalchik, 2020; Angelini et al., 2022). Bookmakers are considered the most accurate predictors of match outcomes (Kovalchik, 2016), with sophisticated models that adjust odds based on betting patterns and proprietary methods. Yet, despite the multi-billion dollar betting industry, one limitation that persists is the poor consideration of intransitivity (van Ours, 2025). Intransitivity is analogous to rock-paper-scissors. In tennis, it occurs where player A tends to defeat B, B defeats C, yet C defeats A, violating the assumption of transitive dominance.