Statistical flaws of the fitness-fatigue sports performance prediction model

Optimizing athletic training programs with the support of predictive models is an active research topic, fuelled by a consistent data collection. The Fitness-Fatigue Model (FFM) is a pioneer for modelling responses to training on performance based on training load exclusively. It has been subject to several extensions and its methodology has been questioned. In this article, we leveraged a Bayesian framework involving biologically meaningful priors to diagnose the fit and predictive ability of the FFM. We used cross-validation to draw a clear distinction between goodness-of-fit and predictive ability. The FFM showed major statistical flaws. On the one hand, the model was ill-conditioned, and we illustrated the poor identifiability of fitness and fatigue parameters using Markov chains in the Bayesian framework. On the other hand, the model exhibited an overfitting pattern, as adding the fatigue-related parameters did not significantly improve the model's predictive ability (p-value > 0.40). We confirmed these results with 2 independent datasets. Both results question the relevance of the fatigue part of the model formulation, hence the biological relevance of the fatigue component of the FFM. Modelling sport performance through biologically meaningful and interpretable models remains a statistical challenge.