ROBERT MOSKOVITCH

Professional road cycling is a team sport where cyclists serve in different tactical roles, yet most predictive models focus solely on individual performance. This paper introduces VeloRost, a Bayesian dual-skill framework that separately models cyclists’ capabilities as leaders and supporting helpers. Using the TrueSkill rating system, we develop three methods for quantifying helper contributions and aggregate them into a roster strength score combined with each cyclist’s leader skill to predict race outcomes. We evaluated our framework through direct ranking, utilizing skill estimation and statistically enhanced learning, across seven seasons of cycling data. Results demonstrate that modeling helper skills significantly outperforms the state-of-the-art method, achieving NDCG@10=0.443, highlighting the important role of helpers in race outcomes. Further analysis of the distribution patterns of helper contributions across race profiles for each method provides insights into when each weighting method excels, offering practical guidance for context-dependent method selection.