A0411
Title: Copula estimation with flow copula models
Authors: Bolin Liu - Ludwigshafen University of Business and Society (Germany) [presenting]
Oliver Grothe - Karlsruhe Institute of Technology (Germany)
Maximilian Coblenz - Ludwigshafen University of Business and Society (Germany)
Abstract: Flow copulas are introduced as a new copula class based on the change of variables formula. Theoretical properties such as the universal expressive power, i.e. any well-behaved absolutely continuous copula can be modeled by a flow copula, are shown. Furthermore, constructions of flow copula models based on the normalizing flow technique and its training procedures are presented. For this, a customized model structure is developed that guarantees copula properties such as uniform margins, which enables the estimation of a flow copula model from data. The learned flow copula model can then be used to estimate the copula density or to generate synthetic data. In simulation studies, it is shown that the presented flow copula models not only can represent various dependence properties such as tail dependence and asymmetry but also provide comparable results to other non-parametric copula estimation methods. Furthermore, the proposed model is applied to real data.
