Essays on Textual Analysis and Causal Inference in Finance

This dissertation draws on recent advancements of textual analysis and causal inference, applying them to current topics in finance. Specifically, the first research paper (Chapter I) investigates whether financial news media coverage causally affects managers’ propensity to obfuscate their communication in corporate disclosures. The paper first applies a simple transformation to the text data of the transcripts of earnings conference calls to measure the degree of manager obfuscation based on the linguistic complexity of manager speech. Finding that increased media coverage overall reduces managers’ incentives to obfuscate, the paper then measures the degree of text similarity between the news articles and prior firm disclosures. The paper finds that articles that are dissimilar to prior firm disclosures (i.e., articles with high editorial content) lead to a significant reduction in subsequent manager obfuscation. Conversely, articles that merely repeat the information provided in prior firm disclosures have no significant effect on manager obfuscation. To mitigate concerns of endogeneity (Roberts and Whited, 2013) and facilitate a causal interpretation of the effects, the paper applies a stacked DID design based on restructuring events at the Wall Street Journal to induce plausibly exogenous variation in news coverage, using the estimator recently proposed by Wing et al. (2024), which is robust to heterogenous treatment effects. The second research paper (Chapter II) similarly investigates whether firms strategically obfuscate in corporate disclosures. Specifically, it studies whether financial institutions lower the comprehensibility of investment prospectuses for complex financial instruments, namely securitizations, to obfuscate low security quality. Measuring the (in-)comprehensibility of investment prospectuses as a combination of prospectus volume and linguistic complexity, the paper applies a simple transformation to the text data of investment prospectuses. Further, various text information is extracted from the prospectuses, proxying for the complexity of the underlying securitization structure and securitized loans. The paper finds that low prospectus comprehensibility is associated with greater defaults and lower returns on the European mortgage-backed security (MBS) market. As large parts of the underlying datasets lack a time dimension, the paper cannot apply econometric methods that would allow for the worse performance of securities with less comprehensible prospectuses to be interpreted as a strategic (causal) decision by the financial institution. However, the paper makes extensive efforts to rule out alternative explanations of the main results, showing that low prospectus comprehensibility is unrelated to the complexity of the securitization structure, loan complexity, or the ex-ante riskiness of the underlying loans. The third research paper (Chapter III), in contrast, uses exogenous shocks by studying the impact of natural disasters on reconstruction labor wages. Specifically, the paper examines under which economic conditions catastrophes lead to a surge in the construction wages, thereby amplifying insured and uninsured losses. Leveraging a large panel dataset covering 9,009 catastrophe regions in the United States, the paper employs a DID design to isolate the wage effects of natural disasters from general business cycle dynamics, thereby exploiting the exogenous nature of disaster events as shocks to local economies. This identification strategy allows for a causal interpretation of post-disaster wage changes. The analysis reveals substantial wage surges of up to 50%, with the magnitude strongly dependent on local labor market conditions such as pre-existing wage differentials, economic growth stage, and workload per employee. The fourth research paper (Chapter IV) is of descriptive nature, thereby drawing minimal causal conclusions. It investigates the perception of economic uncertainty in China by constructing a news-based Uncertainty Perception Indicator (UPI). Using the traditional machine learning method of topic modeling based on Latent Dirichlet Allocation (LDA), the paper analyzes 5,600 uncertainty-related news articles published in the South China Morning Post between 2000 and 2020. This methodological approach allows for the decomposition of overall perceived economic uncertainty into three components—real economy, financial markets, and politics—enabling a more granular understanding of the sources of overall perceived uncertainty. The paper further combines the text-based indicator with stock market data and applies Fama–MacBeth regressions to assess which type of perceived uncertainty has the strongest association with stock returns. The analysis shows that uncertainty related to the real economy is the dominant component of overall perceived uncertainty and that stock prices are most exposed to changes in this dimension.