The Four Whispers

Overview

This system monitors language patterns across quarterly and annual SEC filings for companies within defined peer groups. Rather than predicting outcomes, it detects divergence — moments when a company's filing language deviates significantly from its own historical baseline or from sector peers. The methodology operates in two stages: first, detect statistical deviation; second, investigate what the deviation means. The system is an early warning instrument, not a crystal ball.

Methodology

Filing text is extracted from SEC EDGAR using structured HTML parsing (no LLM extraction). Each filing's MD&A and Risk Factors sections are processed through four analytical lenses: Loughran-McDonald Sentiment Scoring, Term Frequency Deviation (Z-Score), Cosine Similarity Drift, and Conspicuous Silence detection.

Data Pipeline

Source: SEC EDGAR (XBRL financials + filing HTML). Extraction: edgartools structured API + BeautifulSoup HTML parsing. Storage: SQLite (la_deviations, la_flags, la_term_frequencies, la_lm_scores tables). Sectors live: Managed Care, Cable-Broadband, Athletic Retail, Chinese ADR. Refresh: After each quarterly filing season.

Validation

The system was validated by examining whether flagged language deviations preceded known material events. Gate 3 hypothesis testing across managed care and cable sectors confirmed deviation flags clustered 1-2 quarters before analyst downgrades.

Limitations

Filing language is backward-looking; material events may already be priced. Small peer groups (4-6 companies) limit statistical power. 8-K filings are excluded from baseline comparison. The system cannot distinguish between genuine concern and boilerplate risk factor updates.