Title: Nowcasting US GDP in real time: A Bayesian mixed-frequency latent-threshold model with stochastic volatility
Authors: Boriss Siliverstovs - KOF ETHZ (Switzerland) [presenting]
Abstract: A novel approach is proposed for short-term forecasting of economic variables sampled at heterogeneous frequencies by adapting a Bayesian Latent-Threshold Dynamic model to the mixed-frequency setting. As argued before, introducing latent thresholds into dynamic multiple-regressor models helps handling the associated curse of dimensionality by inducing dynamic sparsity in estimated coefficients. This is especially important in models with mixed-frequency data where blocking of higher-frequency data into their lower-frequency counterparts leads to parameter inflation rapidly growing with the frequency mismatch. As a matter of fact, this approach can be considered as a useful alternative to commonly used Bayesian shrinkage applied to large mixed-frequency models. We illustrate the usefulness of the proposed model by nowcasting US GDP growth using historical real-time data vintages.