In this work we introduce a new class of multivariate stochastic volatility models using a latent multifactor structure. Latent factors can follow distinct dynamic structures - stationary autoregressive processes, first- and second-order random walks, and long memory processes. The combination of different dynamic structures in the latent factors makes it possible to capture short and long memory processes, and also the impact of shocks with permanent effects on the conditional variance structure. We perform Bayesian inference of parameters, latent factors and predictions using Bayesian estimation using Integrated Nested Laplace Approximations (INLA), making the method computationally efficient and scalable. We applied this method to analyze a portfolio of cryptocurrencies, and the results of the in-sample and out-of-sample analyzes indicate the importance of combining short- and long-memory processes to capture the dynamics of this market.
Complete tool for submission, evaluation and publication of proceedings for scientific events.
