Text Sentiment’s Ability to Capture Information: Evidence from Earnings Calls

Abstract

We investigate the type of information text sentiment uncovers using earnings conference call transcripts and find that text sentiment fails to explain returns during intraday calls, while average trading volume and return volatility are higher during the call. This finding indicates that intraday calls do convey value-relevant information to the market, but text sentiment cannot capture it. However, text sentiment explains overnight returns extremely well. Since overnight periods are dominated by fundamental news from earnings releases, this finding suggests that text sentiment more forcefully captures firms' fundamental information. In addition, we show that a Lasso-based sentiment measure explains returns significantly better than a dictionary-based sentiment, suggesting that Lasso approach has superior ability to capture information in textual analysis.

Keywords: text sentiment, earnings calls, supervised-learning methods, Lasso, fundamental information

JEL Classification: G12, G14, M41