Bankruptcy Prediction from 10-K Narratives: Evidence from Interpretable Text Scores and Accounting Baselines

This study examines whether context-validated acute distress disclosures in annual Form 10-K filings improve bankruptcy-risk ranking beyond accounting variables. The analysis links U.S. Securities and Exchange Commission filing data, Item 7 Management’s Discussion and Analysis text, and bankruptcy events from the Florida–UCLA–LoPucki Bankruptcy Research Database for fiscal years 2010–2021. The primary sample contains 21,239 nonfinancial firm-year observations and 159 one-year bankruptcy events. The paper develops a Validated Distress Disclosure (VDD) Dictionary that flags going-concern uncertainty, covenant noncompliance, and lender forbearance or waiver after sentence-level context filtering. In the 2019–2021 holdout test, adding VDD indicators to a six-variable Ohlson-related accounting baseline increases the area under the receiver operating characteristic curve (AUC) from 0.8682 to 0.8827 (Δ=0.0146; 95% confidence interval (CI) [0.0040, 0.0275]) and precision–recall AUC (PR-AUC) from 0.0924 to 0.2060 (Δ=0.1136; 95% CI [0.0458, 0.2132]). Benchmark and decomposition tests indicate that the main signal is concentrated in going-concern disclosures and that VDD is best interpreted as an auditable ranking supplement to accounting variables and broader text models, not as a calibrated probability-of-default model.