Using textual analysis to assess the materiality of risk factor disclosures in company's financial reports

The increasing amount of corporate failures in recent years has spurred organisations to manage their risks more effectively. One approach that has gained prominence during this period is enterprise risk management (ERM), a process designed to identify and control business risks and ensure they are consistent with corporate strategies and risk appetite. Several surveys on ERM practices highlight the CFO’s primary role and ownership responsibility in the company-wide implementation of ERM as it facilitates operational and strategic decision-making. Furthermore, recent research from Cohen et al. (2017) has shed a new light on the link between ERM and the quality of the financial reporting process, suggesting that the latter should reflect the company’s financial status (e.g., valuations, estimates) along with its risk assessment activities.

This PhD aims to explore how the extent to which companies depict their ERM activities in their financial reporting process affects their subsequent firm performance and value by applying a variety of more refined NLP analytic techniques. More specifically, by analysing numerous financial reports of companies with a text algorithm/regression model, this PhD seeks to measure the impact of the CFO’s investment in ERM on the company’s corporate (financial, market and shareholder) performance.

While natural language processing (NLP) methods developed in computer science and corpus linguistics have been widely applied in areas such as medicine and education, uptake in the financial reporting domain has been surprisingly slow. Literature in accounting and finance has only scratched the surface of textual analysis capabilities, relying on basic NLP techniques such as word-level analysis through bag-of-words approaches (e.g. dictionary methods, readability and complexity methods, text similarity). As already mentioned, in order to analyse corpora and categorise content that might indicate the materiality of ERM in financial reports and the subsequent impact on corporate health, this study is determined to implement more sophisticated NLP methods such as Topic Identification through Latent Dirichlet Allocation (LDA) and Latent Semantic Analysis (LSA) and drawing on data-driven and statistical text classification through neural networks and machine learning. Furthermore, alternative classification methods ranging from Random Forest and Fisher’s Linear Discriminant to support vector machines (SVM) could be used instead of exclusively focusing on supervised generative classification via naïve Bayes.