Can We Predict Business Cycles With Natural Language Processing?

Detta är en D-uppsats från Handelshögskolan i Stockholm/Institutionen för nationalekonomi

Sammanfattning: For centuries, many economists have attempted to solve the puzzle of business cycles; what explains them, and is it possible to predict them? The great amount of electronically available textual data, together with the recent advancements in unsupervised natural language processing are bound to offer new ways of analysing these perennial questions. Inspired with the emergence of narrative economics as a new field of economic analysis, I propose a novel research design to computationally extract business cycle sentiment from textual data on economic expectations. Firstly, using word vectorisation and embedding techniques on a large corpus of news articles on economic boom and bust, the computer learns to identify expansionary and contractionary sentiment. Secondly, Latent Dirichlet Allocation is used to categorise economic expectation news into topics, and Shannon's Entropy of the topic distribution enumerates the broadness of discourse. The insights from both approaches are used to construct two indices -- Relative Sentiment and Narrative Consensus Index -- to foresee business cycle turning points and predict realisations of U.S. Gross Domestic Product growth. Correlations between current news content and future realisations of U.S. GDP growth as well as several other key macroeconomic time series are established. The predictive power of the indices is evaluated. The Relative Sentiment Index is found to be strongly related to current and short-term future macroeconomic outcomes, and to identify NBER U.S. business cycle turning points with a lead of up to five months. The research method of the thesis offers inspiration for further computational narrative research in macroeconomics, and the conclusions provide preliminary evidence of the potential relevance of economic storytelling for future macroeconomic outcomes.

  HÄR KAN DU HÄMTA UPPSATSEN I FULLTEXT. (följ länken till nästa sida)