Towards a Holistic Understanding of the Disclosure Process

This study applies advanced natural language processing and machine learning techniques to empirically model an important aspect of the disclosure process; how managers translate economic activity into narrative disclosures. Analyzing a large sample of 10-Ks using plagiarism detection software, we categorize MD&A disclosures into four types—new, unchanged, edited, and deleted text. We use an LDA model to measure content of these disclosures and develop 400 measures of managers MD&A disclosure activity. Using XGBoost, a machine learning algorithm, we link firm’s financial statement data to the managers disclosure activity. Findings indicate that over 80% of MD&A content remains the same or is edited annually, with most disclosure growth stemming from sentence edits. We identify patterns in disclosure evolution. Leveraging explainable AI techniques, we find significant predictive power of financial ratios, and that nearly all financial statement variables in our feature set are important in at least one model. This research contributes to the disclosure literature by connecting financial data with textual analysis, offers insights into the potential for AI in accounting, and provides models for stakeholders to evaluate disclosure quality, thus facilitating a smoother transition to AI-assisted disclosure preparation and enhancing the comparability of financial reports.