Mixed-frequency Data Estimation And Forecasting Based On Bayesian Framework

Macro forecasting models based on conventional time series analysis mainly focus on scaled equation(s),which means that variables deserve to be incorporated into the same frequencies under classical framework.However,many time series variables in reality are sampled at different frequency,which makes it relatively difficult to obtain information from variables sampled at high frequencies while making analysis.How-ever,models simply concentrating on variables sampled at low frequency may cause potential information losses.Consequently,it is intuitively more informative to incor-porate economic variables at different frequencies and how to efficiently incorporate these variables into analysis has been popular among academic research recently.In such a case,conventional estimation method like MLE or pseudo-MLE will not be suitable for estimation.Weighting aggregation is still the major method adopted in present research as to solving mixed-frequency series despite this method still faces the potential risk of over-parametrization when the number of variables in model are relatively large.Fur-thermore,it is necessary to incorporate stochastic volatilities into mixed-frequency data model since increasing volatilities and uncertainty in reality.Such a kind of extension will make MLE or pseudo-MLE no longer fit estimation.As a result,it dose make sense to establish a mixed-frequency data model with stochastic volatility and design related estimation method through using the newly-proposed MCMC method under Bayesian econometric framework.This thesis will briefly introduce some background knowledge about Bayesian econo-metrics and related fundamental theories sustaining the proposed research in the first part.Furthermore,this thesis will concentrate on how to establish a mixed-frequency data model with stochastic volatility and how to estimate this model via interpolating Acceptance-Rejection algorithm and Metropolis-Hastings algorithm in Gibbs sampling loop under Bayesian econometric framework.A simulation example will be used as to demonstrate the effectiveness of the proposed estimation method.Finally,this thesis will use the established mixed-frequency data model to forecast the GDP growth rate in U.S.and another example in China as to demonstrate the improved accuracy and timeliness in comparison to benchmark model.