Model Confidence Sets for Nowcast Procedures

Abstract

Nowcasting methods have become a crucial tool for central banks and investors due to their timeliness and ability to make 'on the spot' predictions. However, despite their popularity, there has been little research into statistical methods for the comparison of different nowcasts across multiple horizons. This paper aims at filling this gap. We introduce the concept of nowcast procedures which describe the potentially changing set of 'optimal' models across different nowcast horizons. Formally, our methodology modifies the Model Confidence Set (MCS) to a multi-horizon setting. We thereby address the two dimensions of the multiple testing problem which arise when comparing nowcast accuracy both across models and horizons. We provide guidance to practitioners in using the outputs of the multi-horizon MCS for (i) computing nowcast combination weights based on horizon-specific optimal models, and (ii) testing nowcast monotonicity with multiple models. An empirical application investigates the performance of various model-based predictions alongside a survey of professional nowcasts of the U.S. real GDP and consumption growth rates. Unlike previous studies, we are able to robustly determine the point at which survey nowcasts should be replaced by nowcast model predictions which use more up-to-date information.

Keywords: Nowcasting, Multiple Model Comparison, Model Confidence Set, Bootstrap

JEL Classification: C12, C22, C52, C53