Performance analysis of nowcasting of GDP growth when allowing for conditional heteroscedasticity and non-Gaussianity

Abstract

ABSTRACT The nowcasting performance of autoregressive models for GDP growth are analysed in a setting where the error term is allowed to be characterized both by conditional heteroscedasticity and non-Gaussianity. Standard, publicly available, quarterly data on GDP growth from 1979 to 2019 for six countries are employed: Australia, Canada, France, Japan, the United Kingdom and the United States. In-sample analysis suggests that when homoscedasticity is assumed, support is provided for non-Gaussian error terms; the estimated degrees of freedom of the t-distribution lie between two and seven for all countries. However, allowing for both conditional heteroscedasticity and t-distributed innovations, results indicate that conditional heteroscedasticity captures the fat-tailed behaviour of the data to a large extent. Results from out-of-sample analysis show that point nowcasts are hardly affected by taking conditional heteroscedasticity and/or non-Gaussianity into account. For the density nowcasts, it is found that accounting for conditional heteroscedasticity leads to improvements for Australia, Canada, Japan, the United Kingdom and the United States; allowing for non-Gaussianity seems less important though. This result is robust to which measure is used for assessing density nowcasting performance.