A0845
Title: Forecasting with time-varying order-invariant structural vector autoregressions
Authors: Tomasz Wozniak - University of Melbourne (Australia) [presenting]
Annika Camehl - Erasmus University Rotterdam (Netherlands)
Abstract: Recent developments suggest that heteroskedastic structural vector autoregressions forecast better when their contemporaneous effect matrix is time-varying or invariant to the ordering of the variables in the system. However, combining these two features is challenging because order-invariant specifications can be estimated when the model is identified by time-varying volatility or non-normal shocks, a feature that is difficult to ensure in time-varying models. A new forecasting model is proposed that combines fast-moving stochastic volatility with a Markov-switching structural matrix following persistent regimes. This model enables the identification of structural matrices through heteroskedasticity and non-normality within each regime. Additional flexibility is provided by estimating an overfitting number of regimes and extensive hierarchical prior structures. It is demonstrated that both features, time variation and order invariance of the structural matrix, contribute to improvements in the forecasting precision of macroeconomic systems with up to 20 variables.
