A1282
Title: The flow copula class
Authors: Bolin Liu - Ludwigshafen University of Business and Society (Germany)
Oliver Grothe - Karlsruhe Institute of Technology (Germany)
Maximilian Coblenz - Ludwigshafen University of Business and Society (Germany) [presenting]
Abstract: A new class of copulas, the flow copula, is introduced based on the change of variables formula and Sklar's theorem. Properties of the flow copula are shown, including the universal expressive power of the class, i.e., that any absolutely continuous copula can be modelled by a flow copula. Furthermore, practical constructions of flow copula models that rely on the normalizing flow technique are presented, and a specific training procedure is proposed that guarantees the copula properties, such as uniform marginal distributions. The proposed method can be used not only to estimate a flow copula model from data without prior knowledge of the copula family but also to generate synthetic data efficiently. It is shown through simulation studies and real data applications that the presented flow copula models can represent various dependence properties such as tail dependence and asymmetry. Moreover, they show results comparable to those of other non-parametric copula estimation methods.
