Power of Deep Learning: Quantifying Language to Explain Cross-Sectional Returns

When quantifying qualitative information from unstructured textual data, traditional bag-of-words approaches capture only semantic features of single words/phrases. The context, the sequence of words, and the relations among words (i.e., higher-order interaction features) are ignored. We introduce deep neural networks (NNs) to encode and mimic human intelligence in processing natural language. Using the NN-based artificial intelligence, we construct a new sentiment measure 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 umbrella sentiment proxies used in the literature.