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

Battle Count: 69

The paper is primarily relevant to credit risk and distress prediction rather than direct quantitative trading strategies. However, the PB Stress Score could inform credit spread trading, distressed debt investing, short-selling signals for vulnerable firms, and portfolio risk management. The interpretable, transparent nature of the score makes it suitable for risk-based portfolio construction and early-warning systems that could feed into trading decisions. The 0.32% event rate and one-year horizon make it more suited for strategic risk monitoring than high-frequency trading.

The methodology is relatively straightforward to implement: (1) extract text from 10-K Items 1, 1A, 7 via SEC EDGAR API; (2) apply a frozen dictionary with paragraph-level pattern matching and severity weighting; (3) compute pillar scores with exponential saturation; (4) fit L2-penalized logistic regression with class weighting. The main complexity lies in text extraction quality, handling multi-word phrases, and the bootstrap inference procedure. No deep learning or complex NLP pipeline is required, making it accessible for practitioners.

The paper provides a fully frozen dictionary (Appendix B), explicit scoring formulas, detailed data pipeline description (SEC EDGAR, Florida-UCLA-LoPucki BRD), and transparent model specifications. However, no GitHub repository or code link is provided. The dictionary-based approach is inherently reproducible given the published word lists and formulas. Bootstrap methodology is clearly specified (2,000 replicates, firm-year level resampling).