Abstract: According to Shiller (2017), economic and financial narratives often emerge as a con- sequence of their virality, rather than their veracity, and constitute an important, but understudied driver of aggregate fluctuations. Using a unique dataset of news- paper articles over the 1950-2019 period and state-of-the-art methods from natural language processing, we characterize the properties of business cycle narratives. Our main finding is that narratives tend to consolidate around a dominant explanation during expansions and fragment into competing explanations during contractions. We also show that the existence of past reference events is strongly associated with increased narrative consolidation. (C63, D84, E32, E7)
Keywords: Natural Language Processing, Machine Learning, Narrative Economics.