Is there macroeconomic predictive power in Swedish business news?

Sammanfattning: This thesis explores if, and how, what is written in the newspaper can be used to forecast macroeconomic variables such as inflation, unemployment and consumption. A large data set consisting of articles from the largest Swedish business newspaper is transformed using different methods from the Natural Language Processing field. The focus lies on topic modelling by Latent Dirichlet Allocation as well as sentiment analysis. The newspaper data are represented as a combination of the distribution over topics covered in the newspaper as well as sentiment scores of prominent articles. The data representations of the newspaper are created, explored and later used to predict the movement of the economic variables using Lasso regression, a method automatically selecting important input variables. The newspaper data on stand-alone basis have not been shown to have predictive power for these macroeconomic variables. But, when allowing the model to also be trained on the lagged economic variable interesting observations are made. The predictive performance is improved by the newspaper data, in comparison to only using the lagged economic variable. This is the case for expected inflation, CIPF-inflation and unemployment.