Müller, HenrikRieger, JonasHornig, Nico2021-03-252021-03-252021-03http://hdl.handle.net/2003/4009710.17877/DE290R-21974In this paper, we present a new dynamic topic modeling method to build stable models and consistent time series. We call this new method RollingLDA. It has the potential to overcome several difficulties researchers, who use unsupervised probabilistic topic models, have grappled with: namely the problem of arbitrary selection, which is aggravated when models are to be updated with new sequences of data. RollingLDA is derived by combining the LDAPrototype approach (Rieger, Jentsch and Rahnenführer, 2020) with an implementation that uses preceding LDA results as an initialization for subsequent quarters, while allowing topics to change over time. Squaring dual-process theory, employed in Behavioral Economics (Kahneman, 2011), with the evolving theory of Economic Narratives (Shiller, 2017), RollingLDA is applied to the measurement of economic uncertainty. The new version of our Uncertainty Perception Indicator (UPI), based on a newspaper corpus of 2.8 million German newspaper articles, published between 1 January 2001 and 31 December 2020, proves indeed capable of detecting an uncertainty narrative. The narrative, derived from the thorough quantitative-qualitative analysis of a key-topic of our model, can be interpreted as collective memory of past uncertainty shocks, their causes and the societal reactions to them. The uncertainty narrative can be seen as a collective intangible cultural asset (Haskel and Westlake, 2017), accumulated in the past, informing the present and potentially the future, as the story is being updated and partly overwritten by new experiences. This concept opens up a fascinating new field for future research. We would like to encourage researchers to use our data and are happy to share it on request.enDoCMA Working Paper;6uncertaintynarrativeslatent dirichlet allocationbusiness cyclesCovid-19text miningcomputational methodsbehavioral economics004070310working paper