Vorhersage von Wirtschaftsindikator mittels Sentimentenanalyse von Nachrichtenartikeln und maschinellem Lernen

Abstract

The analysis of textual data and their predictive quality has gained the interest of many researchers, especially in the financial domain. This thesis investigates whether newspaper articles contain information to describe the changes of the Austrian Traded Index (ATX). We apply state of the art methods to extract newspaper articles from the online platform of an Austrian newspaper, to perform sentiment analysis of the articles and to build machine learning models in order to predict price and volatility developments of the ATX. As the newspaper articles are written in German, we create a new sentiment lexicon, called German Financial Sentiment Lexicon (GFSL), by extracting sentiments from the SentiWS, a general German sentiment lexicon, and adding financial sentiment words to the lexicon. Our findings show the newspaper articles contain information which allow predictions of price and volatility movements. The GFSL does not clearly outperform the SentiWS lexicon, although in some scenarios it clearly has an advantage over the general lexicon. The results confirm the findings of previous studies such that negative sentiments highly influence the outcome of the model while positive sentiments are hardly relatable to positive development of the index.