SACkED: Sentiment Analysis of Corporate Earnings Documents

Abstract

In this paper, we present SACkED, Sentiment Analysis of Corporate Earnings Documents, with the goal of finding mispriced stocks by extracting sentiments (or opinions) about a company in the public statements that it issues. In particular, we will focus our attention on the publicly-available transcripts of earnings calls, where companies discuss financial performance after a reporting period. In developing the SACkED approach, we will use dependency trees to analyze sentiments towards key aspects of business operations, as opposed to existing methods that determines sentiments for the documents as a whole.

In particular, this project will focus on learning the language of business as a means of improving on existing Sentiment Analysis techniques. First, we will discuss the development of a new corpus. Next, we focus on learning the language of business using two different text-mining techniques that measure a word’s importance relative to a corpus: term frequency-inverse document frequency, and log-likelihood ratio. Finally, we will analyze whether identifying the important language of business will aid in sentiment analysis.