Deciphering professional forecasters' stories : analyzing a corpus of textual predictions for the German economy

We analyze a corpus of 564 business cycle forecast reports for the German economy. The dataset covers nine institutions and 27 years. From the entire reports we select the parts that refer exclusively to the forecast of the German economy. Sentiment and frequency analysis confirm that the mode of the textual expressions varies with the business cycle in line with the hypothesis of adaptive expectations. A calculated "uncertainty index" based on the occurrence of modal words matches with the economic policy uncertainty index by Baker et al. (2016). The latent Dirichlet allocation (LDA) model and the structural topic model (STM) indicate that topics are significantly state- and time-dependent and different across institutions. Positive or negative forecast "surprises" experienced in the previous year have an impact on the content of topics.