Multivariate time-series modeling with generative neural networks

Generative moment matching networks (GMMNs) are introduced as dependence models for the joint innovation distribution of multivariate time series (MTS). Following popular copula–GARCH approach modeling dependent MTS data, a framework based on GMMN–GARCH is presented. First, ARMA–GARCH utilized to capture serial within each univariate marginal series. Second, if number large, principal component analysis (PCA) used dimension-reduction step. Last, remaining cross-sectional modeled via GMMN, main contribution this work. GMMNs highly flexible and easy simulate from, which major advantage over approach. Applications involving yield curve foreign exchange-rate returns demonstrate utility approach, especially in terms producing better empirical predictive distributions making probabilistic forecasts.