Disclosure Sentiment: Machine Learning vs. Dictionary Methods

We compare the ability of dictionary-based and machine-learning methods to capture disclosure sentiment at 10-K filing and conference-call dates. Like Loughran and McDonald [Loughran T, McDonald B (2011) When is a liability not a liability? Textual analysis, dictionaries, and 10-Ks. J. Finance 66(1):35–65.], we use returns to assess sentiment. We find that measures based on machine learning offer a significant improvement in explanatory power over dictionary-based measures. Specifically, machine-learning measures explain returns at 10-K filing dates, whereas measures based on the Loughran and McDonald dictionary only explain returns at 10-K filing dates during the time period of their study. Moreover, at conference-call dates, machine-learning methods offer an improvement over the Loughran and McDonald dictionary method of a greater magnitude than the improvement of the Loughran and McDonald dictionary over the Harvard Psychosociological Dictionary. We further find that the random-forest-regression-tree method better captures disclosure sentiment than alternative algorithms, simplifying the application of the machine-learning approach. Overall, our results suggest that machine-learning methods offer an easily implementable, more powerful, and reliable measure of disclosure sentiment than dictionary-based methods.