A1446
Title: Computationally tractable factor copula models
Authors: Bahareh Ghanbari - RMIT University (Australia) [presenting]
Pavel Krupskiy - University of Melbourne (Australia)
Laleh Tafakori - RMIT University (Australia)
Yan Wang - RMIT University (Australia)
Abstract: Gaussian factor models are extensively employed to characterise multivariate dependence structures. Nevertheless, their applicability in high-dimensional settings is limited by restrictive tail behaviour and strong distributional assumptions. In particular, they fail to capture asymmetric and tail dependence structures commonly observed in real data. To overcome these limitations, a novel class of factor copula models is proposed that combines the tractability of Gaussian factor models with the flexibility of copula-based dependence modelling. The proposed framework represents each observed variable as a mixture of a latent Gaussian factor and a residual component, where the mixing weights are random loadings. Dependence among the residuals is introduced through an exchangeable one-factor copula driven by an additional latent variable, allowing for flexible tail dependence and asymmetric structures. Both continuous and discrete specifications of the random loadings are considered, leading to analytically tractable expressions for the correlation matrix. Theoretical properties of the proposed model are established and it is shown that the latent factor can be consistently recovered as both the sample size and dimension increase. Simulation studies demonstrate the accuracy and robustness of the estimation procedure in high-dimensional settings. An empirical application illustrates the models ability to capture complex dependence patterns, particularly in the tails.
