| Abstract: |
We examine whether empirical results using text-based sentiment of financial reports depend on the underlying context (within documents) from which sentiment is measured and the impact of a common methodological assumption: document-level aggregation. We measure context-level sentiment by constructing a clause-level measure of context and then applying commonly used sentiment dictionaries and a neural network sentiment classifier at the clause level. Examining stock return prediction, we find that context-level sentiment exhibits coefficient signs that are both consistent with and opposite to document-level sentiment results and that context helps to improve the predictive power of the regressions. Similar results hold across three other prediction problems (material weaknesses, volatility, and volume). This suggests that context is important for interpreting sentiment, and that removing context can lead to inconsistent inferences. To examine document-level aggregation as the likely source of inconsistency, we use a simulation to directly demonstrate the masking effect aggregation has for regressions on sentiment. In sum, context is essential for empirically linking text-based sentiment with various outcome variables, and that document-level aggregation masks this empirical nuance. Our results suggest that sentiment is best applied at the level of specific contexts rather than across whole documents or sections thereof.
Keywords: Sentiment analysis, context, machine learning, aggregation, lasso regression
JEL Classification: C18, C45, D83, M40, M41
|
|
Biography:
|
Dr. Richard M. Crowley is an Assistant Professor of Accounting at SMU. His research sits at the intersection between accounting and machine learning, with a focus on applications of natural language processing. This work aims to help companies, investors, and regulators to understand how textual information impacts financial markets. His recent work focuses on two areas: 1) social media disclosure by corporations and 2) text analytics methodologies for understanding regulatory filings. His research is supported by national-level grants in Singapore, Hong Kong, and Canada and has been published in both Journal of Accounting Research and Contemporary Accounting Research. He holds a PhD in Accountancy from the University of Illinois at Urbana-Champaign.
|
