System identification aims to build models of dynamical systems from data.\nTraditionally, choosing the model requires the designer to balance between two\ngoals of conflicting nature; the model must be rich enough to capture the\nsystem dynamics, but not so flexible that it learns spurious random effects\nfrom the dataset. It is typically observed that the model validation\nperformance follows a U-shaped curve as the model complexity increases. Recent\ndevelopments in machine learning and statistics, however, have observed\nsituations where a "double-descent" curve subsumes this U-shaped\nmodel-performance curve. With a second decrease in performance occurring beyond\nthe point where the model has reached the capacity of interpolating - i.e.,\n(near) perfectly fitting - the training data. To the best of our knowledge,\nsuch phenomena have not been studied within the context of dynamic systems. The\npresent paper aims to answer the question: "Can such a phenomenon also be\nobserved when estimating parameters of dynamic systems?" We show that the\nanswer is yes, verifying such behavior experimentally both for artificially\ngenerated and real-world datasets.\n