Linking Words in Economic Discourse

AROMÍ JOSÉ DANIEL

CABA

Seminario; Seminario de Economía; 2018

CAF

Indicators of unstructured information in the press are developed using a word vector representation model. The information content of these indicators is assessed through business cycle prediction tasks. Word vector representations are trained using the GloVe model (Pennington et al. 2014) and a corpus covering 90 years of press content. The representations are shown to learn meaningful associations in economic context. These associations are exploited to develop indicators of uncertainty. In-sample and out-of-sample forecasting exercises show that the indicators contain valuable information regarding future economic activity. The combination of indices associated to different subjective states (e.g. uncertainty, fear, pessimism) is shown to result in further gains in information content. Alternative text analysis techniques previously proposed in the literature are not seen to capture as much information.