Srinivasan, SurajDey, AiyeshaZhang, Siyu2026-07-0720262026-05-112026Zhang, Siyu. 2026. Corporate Disclosure as a Multi-Format Information System: Processing Frictions and the Prediction of Firm Performance and Risk. Doctoral Dissertation, Harvard University Graduate School of Arts and Sciences.32583531https://dash.harvard.edu/handle/1/42744011This dissertation examines the information content of corporate disclosures and their usefulness for assessing firm performance and risk. Across three essays, I study written narrative disclosures, accounting numbers, and managerial speech to evaluate whether information embedded in these disclosure formats improves the prediction of economically important outcomes. The dissertation is motivated by the idea that publicly available disclosures may not be fully incorporated into investors’ assessments when acquiring and integrating information is costly. The first solo-authored essay (Chapter 2) studies bilingual annual reports issued by firms listed in Hong Kong Stock Exchange and examines whether tone differences between the Chinese and English versions of the Management Discussion and Analysis section convey incremental information. I find that cross-language tone divergence predicts future losses and subsequent stock return mispricing, consistent with investors underreacting to differences in disclosure tone across audiences. The second solo-authored essay (Chapter 3) revisits the long-standing question of whether earnings or cash flows better predict future cash flows. Using Monte Carlo simulations and Compustat data, I show that coefficient-based inferences from conventional panel ordinary least squares (OLS) regressions can be misleading in the presence of persistence and endogeneity. Evaluating predictive performance out of sample, I find that earnings and cash flows contain complementary information, and that models combining both measures provide the strongest forecasts of future cash flows. The third essay (Chapter 4), a co-authored work with Suraj Srinivasan and Wilbur Chen, examines whether managerial speech in earnings conference calls provides incremental information for identifying ongoing latent misstatements, proxied by subsequent restatements. Using a pretrained GPT-2 model, we measure the uncertainty of managers’ responses in earnings-call Q&A and show that adding this behavioral signal to the Bertomeu et al. (2021) misstatement-detection framework improves the model’s ability to distinguish between misstatement and non-misstatement observations, increasing ROC-AUC from 0.7940 to 0.8335. Overall, the dissertation contributes to the accounting literature by applying the disclosure processing cost framework of Blankespoor, deHaan, and Marinovic (2020) to written narrative, numeric, and spoken corporate disclosures. Across the three essays, textual analysis and machine learning methods function as information extraction technologies that reduce awareness costs by uncovering signals investors may overlook, reduce acquisition costs by transforming complex and high-dimensional disclosures into usable measures, and reduce integration costs by allowing heterogeneous signals to be combined flexibly in predictive settings. In doing so, the dissertation shows how modern empirical methods can improve the use of corporate disclosures in assessing firm performance, downside risk, and reporting quality.application/pdfenCash Flow PredictionCorporate DisclosureDisclosure Processing CostsEarnings Call AnalysisMachine Learning in AccountingNarrative ToneBusiness administrationAccountingManagementThesis or Dissertation2026-07-070009-0008-1650-7777