Alternative Data and ML for Macro Nowcasting

Abstract

Worldwide macroeconomic data suffer from three fundamental problems - high dimensionality, a staggered release schedule, and poor data quality. Nowcasts are a popular set of tools that address the first two problems, and the advent of alternative or Big Data offers a chance to address the poor data quality. In this chapter, I provide an overview of nowcasting techniques, discuss the need for an ex-ante hypothesis to guide alternative data selection, and compare typical alternative datasets to traditional data on several quality dimensions such as timeliness and granularity. Finally, I present a case study that establishes that search data can statistically and economically significantly improve US government employment data along the timeliness and accuracy dimensions - a novel result. The case study nowcasts revisions to Non-Farm Payrolls (NFP) three months in advance of the government data, proves these revisions are news and not noise in the framework of Mankiw et al. (1984), controls for Wall Street analyst predictions, and finds that machine learning techniques such as random forest and elastic net provide a substantial improvement over traditional linear regression methods.

Keywords: Nowcasting, Alternative data, Big Data, Machine Learning, AI, NFP, jobs report

JEL Classification: E01, C5, E32, E37, E47, E50, E60, M20, G10, J20,