Central Bank Sentiment Analysis and Asset Prices : Using Machine Learning and Natural Language Processing to Conduct Sentiment Analysis for Predicting Stock Prices in a Norwegian Financial Context

Abstract

This thesis seeks to investigate the relationship between text sentiment in central bank communication and stock market returns in a Norwegian context using a machine learning approach. We collect textual data from Norges Bank (the Norwegian Central Bank) consisting of monetary policy evaluations spanning the last 24 years and apply a multinomial inverse regression (MNIR) to create a positive and negative sentiment dictionary. For performance comparisons, we employ a set of naïve methods, one of which is developed by Kirkeby and Larsen (2021) and one developed by ourselves. Results indicate that there is no significant relationship between the sentiment of Norges Bank and stock returns at Oslo Børs (Oslo Stock Exchange) using test data. Out-of-sample, significant results were only found using a negative sentiment dictionary constructed by ourselves. Counterintuitively, these results indicate that a more negative sentiment leads to higher stock returns. We theorise that not finding significant results with the MNIR-dictionary can be contributed to a few factors. Loss of generality could explain parts of our results. Only using single terms to capture sentiment means that some significance might be lost in the process. We also discuss the possibility that our model mainly captures attributes in the financial landscape that leads to higher or lower stock returns, rather than capturing actual sentiment. Future research into this field using variations in data or methodology can be successful in further investigations of the relationship between central bank communication and asset prices.