Title: Comparing linear and non-linear dynamic factor models for large macroeconomic datasets
Authors: Alessandro Giovannelli - University of Rome Tor Vergata (Italy) [presenting]
Abstract: A non-linear extension for macroeconomic forecasting is proposed by using a large dataset based on a dynamic factor model (DFM). The main idea is to allow the factors to have a non-linear relationship to the input variables using the methods of (i) kernel and (ii) neural networks principal component analysis. We compare the empirical performances of these methods with (iii) the standard principal-component model introduced by Stock and Watson in 2002, conducting a pseudo forecasting exercise based on a Euro Area macroeconomic dataset composed by 834 monthly variables spanning the period January 1996 - September 2017. Using a rolling window for estimation and prediction, the results obtained from the empirical study suggest that (i) and (ii) have the same forecasting performances of (iii) for both Industrial Production and Inflation, but (i) significantly outperforms (iii) for the Unemployment Rate. Moreover, there is no difference with respect to previous results if we consider the pre-crisis period. However, during the crisis and subsequent recovery, we observe a slight improvement of (ii) with respect to (i) for Industrial Production and Inflation while (i) is the best model for Unemployment Rate.