The Effects of Sentiment Evolution in Financial Texts: A Word Embedding Approach

ABSTRACT

We examine the evolutionary effects of sentiment words in financial text and their implications for various business outcomes. We propose an algorithm called Word List Vector for Sentiment (WOLVES) that leverages both a human-defined sentiment word list and the word embedding approach to quantify text sentiment over time. We then apply WOLVES to investigate the evolutionary effects of the most popular financial word list, Loughran and McDonald (LM) dictionary, in annual reports, conference calls, and financial news. We find that LM negative words become less negative over time in annual reports compared to conference calls and financial news, while LM positive words remain qualitatively unchanged. This finding reconciles with existing evidence that negative words are more subject to managers’ strategic communication. We also provide practical implications of WOLVES by correlating the sentiment evolution of LM negative words in annual reports with market reaction, earnings performance, and accounting fraud.

KEYWORDS:

• financial texts
• market sentiment
• word embedding
• sentiment evolution
• textual analysis
• financial communication
• word corpora
• financial word lists

Acknowledgements

The authors thank the anonymous reviewers and the Editor-in-Chief for their constructive comments on this paper. The authors have benefited from discussions with Yanzhen Chen, Carlos Fernández-Loría, Kai-Lung Hui, Chao Jin, Ohchan Kwon, Jing Wang, Juanting Wang, Zhitao Yin, and Ruichao Zhu and comments and suggestions from participants in HKUST Business School Ph.D. Student Conference and PACIS 2023 Doctoral Consortium.

Disclosure statement

No potential conflicts of interest are reported by the author(s).

Supplemental data

Supplemental data for this article can be accessed online at https://doi.org/10.1080/07421222.2023.2301176.

Correction Statement

This article has been corrected with minor changes. These changes do not impact the academic content of the article.

Notes

^1. https://www.sarbanes-oxley-101.com/SOX-404.htm

^2. The sentiment word list became publicly available after Loughran and McDonald published their influential paper in 2011.

^3. Cosine similarity between vectors is defined as: $\mathit{cos}\left(w_1,w_2\right)=\frac{\left(w_1\cdot w_2\right)}{\parallel w_1\parallel \parallel w_2\parallel }=\frac{{\sum }_{i=1}^n{w_1}_{\mathit{i}}{w_2}_{\mathit{i}}}{\sqrt{{\sum }_{i=1}^n{w}_1{{ }_i}^2}\sqrt{{\sum }_{i=1}^n{w}_2{{ }_i}^2}}$ , where ${w_1}_{\mathit{i}},{w_2}_{\mathit{i}}$ are components of vector $w_1,w_2$ respectively.

^4. Our results are qualitatively consistent for a reasonable range of $\mathit{\eta }$ .

^5. We exclude 1994-1996 because there are too few 10-Ks in this period.

^6. https://sraf.nd.edu/. The last update was 01/01/2019 when we retrieve the data.

^7. The initial dataset includes transcripts from 2001 to 2018. We exclude 2001-2002 because there are too few transcripts in this period.

^8. We drop transcript files (1) without information about the occupations of the speakers or (2) without dialogue in Q&A sections or (3) if we fail to detect Q&A sections in the full transcript.

^9. https://developer.dowjones.com/site/global/home/index.gsp

^10. A complete list of Factiva fields used for financial news collection is provided here: https://developer.dowjones.com/site/docs/factiva_apis/factiva_analytics_apis/factiva_snapshots_api/index.gsp#product-overview-353.

^11. A complete list of top-level subject codes is provided here: http://www.factiva.com/en/cp/content/indexing/DJID_FeaturesBenefits_2017.pdf

^12. A word is removed if its embedding vector appears fewer than five times in a data folder.

^13. ${\bar{V}}_{\mathit{positive}}$ (${\bar{V}}_{\mathit{negative}}$) is the representative group vector of positive (negative) words, calculated as the average of the word embedding vectors of all positive (negative) words defined by the LM dictionary.

^14. We only include words with valid word embedding vectors across two domains each year in each sample set to ensure a fair comparison. We thus end up with 260 positive words and 740 negative words in the annual report – conference call sample set and 269 positive words and 1,163 negative words in the annual report – financial news sample set.

^15. Loughran and McDonald [41] leverage the abnormal stock return during the trading window [0, +3] to validate the informativeness of their proposed LM dictionary. We thus follow their setting to use the trading window [0, +3]. Our results are qualitatively consistent if using another commonly used trading window [0, +1] [17].

^16. We accessed the database in October 2021.

^17. The results remain qualitatively consistent when we follow the regression model of Fama and MacBeth [14], in which a standard error with one lag is applied.

^18. Li [35] and Henry and Leone [24] only focus on forward-looking statements in MD&A. In contrast, we analyze the whole MD&A section.

Additional information

Funding

This project was funded by the Theme-based Research Grant on Fintech (T31-604/18-N) of the Research Grant Council of Hong Kong.

Notes on contributors

Jiexin Zheng

Jiexin Zheng is a Ph.D. student in the Department of Information Systems, Business Statistics, and Operations Management at the Hong Kong University of Science and Technology (HKUST). His research interests include financial text analysis and the economics of AI. His research has been presented at several international conferences and workshops, including International Conference on Information Systems and Workshop on Information Technologies and Systems.

Ka Chung Ng

Ka Chung Ng is an Assistant Professor and Presidential Young Scholar in the Department of Management and Marketing, Faculty of Business, Hong Kong Polytechnic University. He received his Ph.D. in Information Systems from HKUST. Dr. Ng’s research interests lie in fake news, business analytics, and fintech. He has published in Journal of Management Information Systems, Production and Operations Management, ACM Transactions on MIS, and the proceedings of the leading Information Systems conference.

Rong Zheng

Rong Zheng is an Associate Professor of Information Systems at the Business School. He earned his doctoral degree in Information Systems from the Stern School of Business at New York University. Dr. Zheng’s general research interest is in realizing business value with AI. More recently, his research examines how the use of AI methods in information processing can change the information environment of financial market. His work has been published in such journals as Information Systems Research, Management Science, The Accounting Review, and Communications of the ACM. He is an associate editor of Business & Information Systems Engineering.

Kar Yan Tam

Kar Yan Tam is Dean of the Business School and Chair Professor of Information Systems, Business Statistics, and Operations Management at HKUST. He received his Ph.D. from Purdue University and is a founding member of the HKUST Business School. His current research interests lie in fintech, business analytics, and sustainable and green finance. Dr. Tam has published in Journal of Management Information Systems, MIS Quarterly, Information Systems Research, Management Science, and other journals. He is serving on the Board of AACSB and EFMD and on the editorial boards of a number of Information Systems journals.