Macroeconomic Forecasting and Market Analysis with Newspaper Articles

The overarching theme of this dissertation is whether a particular type of alternative data, namely text data, can improve forecasts of key macroeconomic aggregates or help explain the financial market dynamics. Large-scale text data have become available only relatively recently and have attracted substantial attention in the empirical macroeconomic and financial literature. My central research question concerns which dimensions of news data are informative for macroeconomic forecasting and financial analysis, and how these dimensions should be adapted to the variable of interest. The existing literature has predominantly focused on three types of text-based indicators. The first strand constructs sentiment measures that quantify the positive or negative tone of news articles and shows that such measures can predict macroeconomic activity. A second strand uses topic indicators, capturing variation in the intensity with which different themes are covered in the media. A third, more recent strand combines these approaches by assigning sentiment to individual topics, and therefore produces topic-specific sentiment measures that reflect both the thematic focus and the tone of news coverage. My work adopts this hybrid perspective but departs from standard off-the-shelf methods by asking whether commonly used text dimensions are appropriate for the specific forecasting question at hand. Rather than relying directly on existing lexicons or unsupervised topics, I adapt sentiment analysis and topic modelling so that the resulting indicators capture the economically relevant information for the variable being forecast, whether GDP, consumption, investment, or stock returns.