Business Cycle Narratives

Abstract

This article quantifies the epidemiology of media narratives relevant to business cycles in the US, Japan, and Europe (euro area). We do so by first constructing daily business cycle indexes computed on the basis of the news topics the media writes about. At a broad level, the most influential news narratives are shown to be associated with general macroeconomic developments, finance, and (geo-)politics. However, a large set of narratives contributes to our index estimates across time, especially in times of expansion. In times of trouble, narratives associated with economic fluctuations become more sparse. Likewise, we show that narratives do go viral, but mostly so when growth is low. While narratives interact in complicated ways, we document that some are clearly associated with economic fundamentals. Other narratives, on the other hand, show no such relationship, and are likely better explained by classical work capturing the market's animal spirits.

Econometrics Meets Sentiment: An Overview of Methodology and Applications

“…Larsen and Thorsrud (2018) use a graphical Granger causality modeling framework to gain insights into the network of economically relevant news topics. Every node in the graph represents a sentiment/topic time series.…”

“…This is due to both the changing mix of the news sources as well as the actual impact of the news. Interestingly, Larsen and Thorsrud (2018) find that narratives mostly go viral during downs in the business cycle, albeit for a duration of only a few months.…”

“…Does the sentiment impact the outcome variable directly or indirectly, and through what mechanism? Larsen and Thorsrud (2018) partition their network of sentiment/topic variables into more and less exogenous variables. Variables are considered exogenous if they have predictive power for other topics but are not (often) predicted themselves.…”

Forecasting Japanese inflation with a news-based leading indicator of economic activities

“…For example, Shapiro, Sudhof, and Wilson (2018) develop a US news sentiment index, based on 16 major US newspapers from 1980 to 2015, which is found to be strongly correlated with the state of contemporaneous business cycles. In a similar vein, by combining quarterly GDP data and the daily news topic variables in mixed-frequency dynamic factor models, business cycle indexes are constructed by Thorsrud (2018) for Norway and by Larsen and Thorsrud (2018) for the US, Japan and Europe (the euro area), respectively. While we share the motivation behind these studies, our methodology of constructing business cycle indexes has several distinct features that diverge from their analysis.…”

“…We make use of this survey structure and consider the supervised learning of topics to determine whether the sentences describe either current or future economic conditions. Our use of supervised learning to estimate the sentence topic is in contrast with the analyses of Thorsrud (2018) and Larsen and Thorsrud (2018) that are based on the Latent Dirichlet Allocation (LDA) model, one of the most popularly used unsupervised topic models. For our purposes, supervised learning seems more appropriate since the topic of our interest, namely, future economic conditions, is di¢ cult to discover by means of unsupervised topic model such as LDA.…”