Disclosed Human-Capital Disruption and Firm-Specific Risk

Battle Count: 239

The paper provides a firm-specific risk signal (idiosyncratic volatility, downside deviation, worst-month return) that is distinct from market beta and labor-shortage exposure. The call-timed evidence shows predictive content for 42-day forward idiosyncratic volatility (~0.50% per SD), which could inform risk-adjusted position sizing, volatility forecasting, or tail-risk hedging. However, the signal is quarterly-frequency (earnings calls), contemporaneous in annual tests, and the effect sizes are modest. The organizational outcomes (CEO turnover) could inform event-driven strategies. The measure is not directly a return predictor but a risk-state indicator, making it more relevant for risk management than alpha generation.

The full pipeline requires: (1) transcript collection and cleaning from multiple sources, (2) vocabulary-based sentence screening, (3) sentence embedding with all-mpnet-base-v2, (4) SVM training for broad workforce classification, (5) fine-tuning three DeBERTa models on labeled data, (6) ensemble scoring and threshold selection, (7) aggregation to fiscal years, (8) matching to CRSP/Compustat/Execucomp, (9) panel regression with fixed effects and clustering, (10) call-timed event-study construction. The NLP pipeline is moderately complex but uses standard tools. The economic analysis requires careful sample construction, fiscal-year alignment, and multiple robustness checks. Reproducing the exact measure requires access to the author's training labels and Opus prompts.

The paper uses publicly available transcript archives (Motley Fool, Kurry dataset), CRSP, Compustat, Fama-French factors, Execucomp, and the released Harford et al. (2026) HHQ measure. The DeBERTa models are fine-tuned on Opus-labeled data, but the specific training labels and Opus prompts are not fully released. The coding criteria are documented in the appendix. No GitHub repository is mentioned. The measurement pipeline involves multiple judgment calls (thresholds, vocabulary, context windows) that are described but not fully automated.