News Embeddings for Sector Timing – A textual similarity based approach to determining trending sectors in news coverage


Asset Pricing, Empirical Capital Market Research



Prof. Dr. Dominik Wolff Hans Christian Schmitz Prof. Dr. Dirk N. Dr. Berhard L.


In this study, we present an approach to quantify the relevance of daily newspaper articles  for individual companies and their corresponding sectors. Based on the textual content of daily newspaper, we utilize sentence and paragraph embeddings to compare news articles from daily newspapers over a time span of 15 years with company descriptions from Form 10-K filings. We measure the relevance of a news article as the similarity between the article and the company’s description. In contrast to prevalent approaches in the literature, our relevance estimation does neither rely on stock price data nor on a predefined mapping between news articles and companies. We evaluate the proposed methodology on several strategies on sector exchange traded funds and on the level of individual stocks. We find that the trading strategies based on our relevance estimation approach yield up to 15.38% out-of-sample annualized return after transaction costs.