The emergence of the first Fitness-Fatigue impulse responses models (FFMs) have allowed the sport science community to investigate relationships between the effects of training and performance. In the models, athletic performance is described by first order transfer functions which represent Fitness and Fatigue antagonistic responses to training. On this basis, the mathematical structure allows for a precise determination of optimal sequence of training doses that would enhance the greatest athletic performance, at a given time point. Despite several improvement of FFMs and still being widely used nowadays, their efficiency for describing as well as for predicting a sport performance remains mitigated. The main causes may be attributed to a simplification of physiological processes involved by exercise which the model relies on, as well as a univariate consideration of factors responsible for an athletic performance. In this context, machine-learning perspectives appear to be valuable for sport performance modelling purposes. Weaknesses of FFMs may be surpassed by embedding physiological representation of training effects into non-linear and multivariate learning algorithms. Thus, ensemble learning methods may benefit from a combination of individual responses based on physiological knowledge within supervised machine-learning algorithms for a better prediction of athletic performance. In conclusion, the machine-learning approach is not an alternative to FFMs, but rather a way to take advantage of models based on physiological assumptions within powerful machine-learning models.