Sentiment analysis of earnings calls via natural language processing in the Nordics: Uncovering abnormal returns and stock reversals through FinBERT and FinVADER

Abstract

In this thesis, we examine the relationship between the sentiment of earnings calls and subsequent stock performance, analyzing around 3,500 earnings call transcripts from public companies in the Nordics from January 2018 to December 2023. Specifically, we analyze the sentiment conveyed in the earnings calls via two separate models, a dictionary -based natural language processing classifier, FinVADER, and a large language model trained on financial text, FinBERT, and their combination. Our findings contribute to the existing literature in three important ways, namely through methodology, a unique dataset, and the uncovered financial results from their interplay. We bolster the still nascent literature on the performance of LLMs compared to dictionary -based algorithms in classifying sentiment in the financial context, showing that both have predictive power for abnormal returns, but that LLMs are superior. Distinctively, we demonstrate that combining a lexicon -based NLP with an LLM can yield improved predictability of abnormal returns over individual models. This finding is essential as sentiment is interpreted differently by different types of models. To the best of our knowledge, we are the first to analyze earnings call transcripts with both FinVADER and FinBERT, highlighting the value of integrating different NLP approaches for improved financial forecasting.

Furthermore, our focus on the Nordics allows us to examine the existence of post announcement earnings drift (PEAD) based on sentiment with a unique dataset, a phenomenon that studies show to be non-existent in the many developed markets such as the US, but still present in the Nordics. Contrary to other research, our findings suggest that there is no PEAD in the Nordics based on sentiment. Moreover, we find that while higher sentiment correlates with higher cumulative abnormal returns (CARs) around the earnings announcement, we see a reversal effect in the medium-term or a negative PEAD, suggesting a negative correlation between sentiment and CARs in the longer time window. This reaction is further highlighted by a unique approach of regressing only extreme observations of sentiment that indicate that extreme sentiment leads to market overreaction in the initial period and a higher reversal in the extended period, magnifying the result.