EcoSta 2022: Start Registration
View Submission - EcoSta2022
A0537
Title: Inference in stochastic volatility models with variational sequential Monte Carlo Authors:  Yuliya Shapovalova - Radboud University (Netherlands) [presenting]
Abstract: Applications of stochastic volatility models, in particular in the multivariate case, are limited due to high computational time requirements by state-of-the-art methods such as particle Markov chain Monte Carlo methods. When the number of latent states in stochastic volatility models is large, inference with this class of methods becomes practically infeasible. Variational methods in recent years have shown great potential in large-scale applications. However, those relying on linearization techniques in the case of stochastic volatility models may still result in large biases even in the univariate case. We consider a recently published work on a new approximating family of distributions, the variational sequential Monte Carlo (VSMC), with application to stochastic volatility models. The advantage of this new inference method is in its flexibility: the posterior can be approximated arbitrarily well while maintaining efficient optimization of the parameters. Using a case study of stochastic volatility models, we evaluate the potential of VSMC in terms of scalability and precision and place it in the literature in comparison to other methods varying from particle MCMC to INLA.