Power of deep learning: Quantifying language to explain cross-sectional returns

Abstract

When quantifying information from unstructured textual data, the traditional bag-of-words approach only captures semantic features of single words or phrases. The context, the sequence of words, and the relationship between words are ignored. This paper introduces a novel approach to incorporate complex syntactical featuresin the textual analysis using machine learning (i.e., neural-network-based natural language parser and word embedding). We construct a new measure of sentiment that is specific to performance discussions and is adjusted for complex contextual negations. We find that this performance-specific sentiment explains cross-sectional returns and future operating performance better than the umbrella sentiment proxies used in the literature