Measuring economic uncertainty using news-media textual data

We develop a news-media textual measure of aggregate economic uncertainty, defined as the fraction of Financial Times articles that contain uncertainty-related keyphrases, at frequencies from daily to annual, from January 1982 to April 2014. We improve on existing similar measures in several ways. First, we reveal extensive and irregular duplication of articles in the news database most widely used in the literature, and provide a simple but effective de-duplication algorithm. Second, we boost the uncertainty ‘signal strength’ by 14% through the simple addition of the word “uncertainties” to the conventional keyword list of “uncertain” and “uncertainty”, and show that adding further uncertainty-related keyphrases would likely constitute only a second-order adjustment. Third, we demonstrate the importance of normalising article counts by total news volume and provide the first textual uncertainty measure to do so for the UK. We empirically establish the plausibility of our measure as an uncertainty proxy through a detailed narrative analysis and a detailed comparative analysis with another popular uncertainty proxy, stock returns volatility. We show the relationship between these proxies is strong and significant on average, but breaks down periodically. We offer plausible explanations for this behaviour. We also establish the absence of Granger causation between the measures, even down to daily (publication) frequency.