News versus Sentiment: Comparing Textual Processing Approaches for Predicting Stock Returns
December 4, 2013 Leave a comment
News versus Sentiment: Comparing Textual Processing Approaches for Predicting Stock Returns
Steven L. Heston University of Maryland – Department of Finance
Nitish Ranjan Sinha Board of Governers of the Federal Reserve System
September 4, 2013
Robert H. Smith School Research Paper
Abstract:
This paper uses a dataset of over 900,000 news stories to test whether news can predict stock returns. It finds that firms with no news have distinctly different average future returns than firms with news. We measure sentiment with the Harvard psychosocial dictionary used by Tetlock, Saar-Tsechansky, and Macskassy (2008), the financial dictionary of Loughran and McDonald (2011), and a proprietary Thomson-Reuters neural network. Simpler processing techniques predict short-term returns that are quickly reversed, while more sophisticated techniques predict larger and more persistent returns. Conforming previous research, daily news predicts stock returns for only 1-2 days. But weekly news predicts stock returns for a quarter year. Positive news stories increase stock returns quickly, but negative stories have a long-delayed reaction.