earnings call word count anomaly
What it checks
When an earnings call runs unusually long OR the Q&A balloons relative to the prepared remarks (z >= 1.5 vs the ticker's own history), management is hedging — short.
Mechanism
Bushee-Matsumoto-Miller 2003 found longer earnings-call transcripts correlate with greater management uncertainty (saying more = hedging). Cross-source composite: leg 1 is transcript content_length z vs trailing-8-call own-baseline; leg 2 is Q&A share z (qa_n_sentences / total).
Signal rule
content_length z >= +1.5 OR Q&A-share z >= +1.5 (vs ticker's 8-call baseline) → SHORT on T+1; hold 30/60d.
Data dependencies
daily_pricesAdjusted-close OHLCV for every US-listed ticker; primary price feed.
earnings_call_transcriptsFull earnings-call transcripts (prepared + Q&A), tokenised.
transcript_finbert_scoresWorker data table — see services/worker schema.
Expected edge
- Paper alpha
- 75-200 bps
- Paper window
- T+1 to T+60d
75-200 bps over 30-60d.
Example tickers where this is likely to fire
Illustrative only — the signal fires based on the live data, not a fixed list.
Related families
transcript obfuscation bloomfieldText-NLPBloomfield (2002) Incomplete Revelation Hypothesis: managers obfuscate bad news through reduced readability. Bochkay-Chychyla-Nanda (2020 JAR) formalized the test on conference calls: the Gunning Fog Index Q/Q delta predicts 12-month negative drift on the readability-decay decile (~4% underperformance annualised). Own-history z of the call's Fog index: short when fog_z ≥ +1 (sudden complexity spike → obfuscation), long when fog_z ≤ −1 (sudden simplification, often pre-positive guidance).
transcript finbert qa dispersionText-NLPHuang-Wang-Yang (2023 CAR) "FinBERT: A Large Language Model for Extracting Information from Financial Text." FinBERT-tone outperforms Loughran-McDonald wordcount on every post-call drift measure tested (3-day CAR R² roughly doubles). Follow-up Chen et al. (2024 JFE) reports ~5% annualised L/S on the Q&A-dispersion sub-signal. Own-history z of per-call Q&A sentiment dispersion (stdev of per-sentence pos-neg score): short when z ≥ +1 (uncertainty), long when z ≤ −1 (confident answers).
earnings announcement premiumEventStocks earn an abnormally positive return in the days surrounding scheduled earnings announcements (T−2 to T+1). The original 2007 paper documented ~9bp/day average premium during the window vs the non-announcement baseline; later replications show the effect has compressed to ~5-6bp/day but is still robust. The mechanism is asymmetric resolution of uncertainty plus retail inattention immediately around announcement. Signal: long the (T−2 to T+1) window per announcement, flat otherwise.
Explore earnings call word count anomaly on alphactor.ai
See which tickers this family is currently firing on, with live signals and rankings.