Induction of a sentiment dictionary for financial analyst communication: a data-driven approach balancing machine learning and human intuition

Abstract

ABSTRACT While sentiment dictionaries are easy to apply and provide reproducible results, they often exhibit inferior classification performance compared to machine learning approaches trained for specific application domains. Nevertheless, both approaches typically require manual data analysis. This paper develops a domain-specific dictionary using regularised linear models drawing from textual reports of financial analysts. The first evaluation step demonstrates that the developed financial analyst dictionary can explain cumulative abnormal stock returns related to earnings events more accurately compared to other finance-related dictionaries and sentiment classifiers. In a second step, the approaches are compared using manually annotated sentiment. The financial analyst dictionary is more accurate than other dictionary-based approaches, although it cannot compete with a pre-trained deep learning sentiment classifier. While we show that the proposed approach is suited for texts of financial analysts, it can be applied to other use cases. The approach realises context specificity while reducing extensive manual data analysis.