Agenda

Sylvia Frühwirth-Schnatter - Wirtschaftsuniversität Wien

Time-varying parameter (TVP) models are a popular tool for handling data with smoothly changing parameters. However, in situations with many parameters the flexibility underlying these models may lead to overfitting models and, as a consequence, to a severe loss of statistical efficiency. This occurs, in particular, if only a few parameters are indeed time-varying, while the remaining ones are constant or even insignificant. As a remedy, hierarchical shrinkage priors have been introduced for TVP models to allow shrinkage both of the initial parameters as well as their variances toward zero.

The present talk reviews various recent approaches of introducing shrinkage priors for TVP models. Bitto et al (2019), for instance, introduced the (hierarchical) double Gamma prior and discussed efficient methods for MCMC inference and predictive analysis. The talk also discussed recent extensions of the double Gamma prior such as the triple Gamma prior (which includes the horseshoe prior as a special case) and the spike-and-slab double Gamma prior (which includes the spike-and-slab Lasso prior as a special case).

For illustration, hierarchical shrinkage priors are applied to EU area inflation modelling based on the generalized Phillips curve and to a Cholesky stochastic volatility model, modelling multivariate financial time series of stock returns from the DAX. The results clearly indicate that shrinkage priors reduce the risk of overfitting and increase statistical efficiency in a TVP modelling framework.

(based on joint work with Angela Bitto,  Annalisa Cadonna and Peter Knaus, Vienna University of Economics and Business)