Title: Model specification for Bayesian neural networks in macroeconomics
Authors: Karin Klieber - Oesterreichische Nationalbank (Austria) [presenting]
Niko Hauzenberger - University of Salzburg (Austria)
Florian Huber - University of Salzburg (Austria)
M. Marcellino - Bocconi University (Italy)
Abstract: Relations in macroeconomic data are often nonlinear and subject to structural breaks. This is commonly captured through appropriate models. However, by choosing a specific model, the researcher takes a strong stance on the nature and degree of nonlinearities. This gives rise to substantial model and specification uncertainty. We develop Bayesian neural networks that remain agnostic on the precise form of nonlinearities. Our model flexibly adjusts to the complexity of the dataset. This is achieved through Bayesian regularization techniques that adequately select the appropriate network structure without the necessity for using cross-validation. To investigate the degree and nature of nonlinearities in macroeconomic data, we train our neural network to four commonly used datasets in macroeconomics and finance. Our empirical results suggest that for cross-sectional data, a linear approximation works well in predictive terms, whereas for time series data, nonlinearities are important and especially so during turbulent times.