Inferência Bayesiana para modelos de volatilidade estocástica baseados em mistura de escala da distribuição normal assimétrica

Abstract

This dissertation aims to evaluate and compare the performance of the No-U-Turn Sampler (NUTS) algorithm, implemented in the Stan software, in estimating the parameters of stochastic volatility models with leverage based on scale mixtures of the skew-normal distribution. These SV models can simultaneously capture important features of financial return series, such as leverage effect, heavy tails, and asymmetry. The results of simulation studies show that, according to bias and root mean squared error (RMSE) measures, the NUTS algorithm performs well. When comparing the NUTS sampling approach with that of the stochvol package, we observe that stochvol has faster execution times, but NUTS outperforms it in terms of effective sample size. Additionally, we propose the use of the Leave-Future-Out Cross-Validation (LFO-CV) technique for selecting stochastic volatility models and evaluate the performance of information criteria and cross-validation techniques for model selection. Finally, we apply the developed methodology to real return series.