In this article, we develop a new general inference method for selecting learning models. The method relies upon a specific hold-out cross-validation, which takes into account the dependency within the data. This allows us to retrieve the model that best fits the learning strategy of a single individual. The novelty of our approach lies on the choice of the testing set, both in the experimental design and in the data analysis. This individual approach is then applied to two category learning models (ALCOVE and Component-cue) on data-sets manipulating presentation order, after verification of the reliability of our method. We found that both models performed equally well during transfer, but Componentcue best fits the majority of participants during learning. To further analyze these models, we also investigated a potential relation between the underlying mechanisms of the models and the actual types of presentation order assigned to participants.