This paper aims at understanding modular organization in multivariate dynamical data. In contrast to information-theoretic approaches, the authors start from the complementary point of view of statistical modeling and prediction of dynamical systems. They arrive at the conclusion that modularity is not necessarily an objective property of a system’s organization but rather is inferred by cognitive systems as it can simplify learning and lead to gains in predictive power. This conclusion may prove useful for constructivist approaches as the paper establishes in formal ways that learning agents may perceive modularity and correlations among variables in their environments, even when such variables are actually dependent on others.