News Sentiment

Abstract

We introduce a novel method for training computer algorithms to measure news sentiment. Our approach leverages human-coded sentiment scores from over 200,000 newspaper articles to teach the computer to select words, word combinations, and their linear weights. In an out-of-sample test, examining newspaper articles about US companies, we show that: (i) our news sentiment metric displays a bimodal distribution similar to that observed in the human-coded sentiment scores, (ii) our news metric outperforms the widely-used bag-of-words approach and recent machine learning models in explaining human-coded news sentiment, and (iii) our news sentiment metric serves as a robust predictor for daily stock returns.

Keywords: Textual analysis, Machine Learning, News sentiment, Stock returns

JEL Classification: C53, C55, G11, G12, G14, G17, G41