Which Voices Move Markets? Speaker Identity and the Cross-Section of Post-Earnings Returns

Battle Count: 123

Highly relevant to quantitative trading. The paper identifies a specific, implementable alpha source: buying stocks with positive section-weighted earnings call sentiment and selling those with negative sentiment, with weights tilted toward analyst commentary. The monthly long-short alpha of 2.03% (t=6.49) unexplained by Fama-French five factors is economically significant. The signal has a half-life of 6-7 trading days, making it actionable for short-horizon strategies. The out-of-sample IC of 0.142 and the analyst-only IC of 0.171 (XGBoost) are strong for a single signal. The paper provides concrete portfolio construction (quintile sorts), factor-adjusted returns, and decay analysis directly relevant to strategy implementation. The finding that the signal is orthogonal to SUE means it can complement existing earnings-surprise strategies.

Moderate to high complexity. Requires: (1) Large-scale NLP pipeline processing 6.5M sentences through FinBERT, (2) Speaker classification logic based on title/affiliation parsing, (3) Five different aggregation methods with IC-derived weight computation, (4) Event-window return construction accounting for BMO/AMC timing, (5) Fama-French five-factor and Fama-MacBeth regression frameworks, (6) Out-of-sample validation with frozen weights, (7) XGBoost-SHAP validation pipeline. The core FinBERT inference is straightforward with HuggingFace, but the full pipeline from raw transcripts to tradable signal requires significant engineering. Data acquisition from Alpha Vantage at scale is the primary practical bottleneck.

The methodology is described in detail with explicit formulas for all five aggregation methods, IC-derived weighting, Fama-MacBeth regressions, and Fama-French factor regressions. However, the paper relies on commercial data sources (Alpha Vantage API, Yahoo Finance) that may have access restrictions. The FinBERT model is publicly available. The Loughran-McDonald dictionary is available from Notre Dame. No code repository is mentioned. The temporal split (train: pre-2023, test: 2023+) is clearly defined. The main barrier to reproducibility is the large-scale data collection (6.5M sentences from 16,428 transcripts) and the specific Alpha Vantage API access.