Integrated nested Laplace approximations for threshold stochastic volatility models

Abstract

The aim is to implement the integrated nested Laplace approximations (INLA), known to be very fast and efficient, for estimating the parameters of the threshold stochastic volatility (TSV) model. INLA replaces Markov chain Monte Carlo (MCMC) simulations with accurate deterministic approximations. Weakly informative proper priors are used, as well as Penalizing Complexity (PC) priors. The simulation results favor the use of PC priors, specially when the sample size varies from small to moderate. For these sample sizes, PC priors provide more accurate estimates of the model parameters. However, as sample size increases, both types of priors lead to similar estimates of the parameters. The estimation method is applied to six series of returns, including stock market, commodity and cryptocurrency returns, and its performance is assessed, by means of in-sample and out-of-sample approaches; the forecasting of one-day-ahead volatilities is also carried out. The empirical results support that the TSV is the model that generally fits the best to the series of returns and most of the times ranks the first in terms of forecasting one-day-ahead volatility, when compared to the symmetric stochastic volatility model.