7.4.6 Advanced Extensions

Basic Prerequisites: 03 · Sentiment & Tone and 04 · Embeddings & Transformers.

---

1. Intuition & Practical Objective

Earnings calls are the highest-signal financial text because they are two-staged: a scripted opening (prepared remarks) and a spontaneous Q&A that management cannot fully script. The discretionary analyst’s edge is reading the Q&A - the hedging, the evasions, the unquantified caveats. Earnings-call analysis quantifies that edge by:

1. Separating the stages - prepared remarks carry boilerplate and managed tone; the Q&A carries the incremental, less-filtered information. Measuring them separately is itself the feature.
2. Weighting by speaker and content - an unquantified answer from the CFO about a charge is worth more than a scripted “we remain confident.”
3. Running the full pipeline - ingest (transcript segments) → clean (tokenize) → features (per-segment tone, linguistic change vs last quarter) → model (regress post-call drift on executive Q&A tone).

This is where the folder’s pieces assemble: the LM-calibrated tone (02/03) is the feature; the Q&A-vs-prepared split is the textual analog of the earnings-quality idea that spontaneous disclosures are more informative than scripted ones (Accounting Quality & Red Flags); and a FinBERT (04) contextual head replaces the dictionary when labeled training data justify it.

The advanced theme, one sentence: “the information in an earnings call is not uniform - it concentrates in the least-scripted, least-boilerplate text, so the pipeline must separate, weight, and timestamp that text before any model sees it.”

---

2. Mathematical Ground Truth & Derivations

The weighted executive-tone feature. For a call with Q&A segments $s$, each with word count $w_s$ and net tone $\text{tone}(s)$, the executive-weighted score is

$$\text{ExecTone}=\frac{{\sum }_{s\in \text{Q\&A},s\in \{\text{CEO,CFO}\}}{w}_s\cdot \text{tone}(s)}{{\sum }_{s\in \text{Q\&A},s\in \{\text{CEO,CFO}\}}{w}_s},$$

a word-weighted average so longer, more substantive answers dominate. Contrast the two summary scalars you would otherwise conflate:

$${\overline{\text{tone}}}_{\text{prepared}}=\frac{1}{|P|}\sum _{s\in P}\text{tone}(s),{\overline{\text{tone}}}_{\text{Q\&A}}=\frac{1}{|Q|}\sum _{s\in Q}\text{tone}(s).$$

A systematic gap - scripted optimism vs. guarded Q&A - is itself a signal (the “tone gap” used in the earnings-call literature to flag management holding back).

The cross-sectional model. Across $n$ firms, regress post-call drift on executive Q&A tone:

$$y_i={\beta }_0+{\beta }_1{\text{ExecTone}}_i+{\epsilon }_i,$$

estimated by OLS (normal equations as in 03). A stable positive ${\beta }_1$ is the claim that Q&A tone is priced.

Linguistic change as a separate signal. Using TF–IDF document vectors (02):

$$\Delta {\text{Text}}_t=1-\cos ({\mathbf{v}}_t,{\mathbf{v}}_{t-1}),$$

the “Lazy Prices” construction (Cohen–Malloy–Nguyen 2020): silent wording changes in a firm’s filings - no number changed - forecast underperformance. This is a structural text feature orthogonal to sentiment, and it generalizes directly to year-over-year transcript comparison.

---

3. Computational Implementation - the full pipeline on a mini call + cross-section

Ingests a transcript, separates prepared vs Q&A, computes the word-weighted executive Q&A tone, and regresses post-call drift across five firms.

The numbers carry the whole story. The prepared remarks are optimistic (+0.156) but the Q&A is guarded (−0.075): management leads with confidence and then hedges under questioning - the tone gap a purely scripted analysis would miss. Weighted over CEO/CFO Q&A the executive tone is −0.032, and across the five-firm cross-section drift tracks it with slope +10.797 and $R^2=0.99$. On a real panel $R^2$ will fall far below this clean sample, but the direction - spontaneous Q&A tone is priced, and it differs from scripted tone - is exactly the mechanism the earnings-call literature reports.

---

4. Failure Modes & First-Principles Breakdowns

1. Reading the whole call as one blob. Averaging sentiment over scripted + spontaneous text dilutes the Q&A signal with boilerplate (the prepared/Q&A gap measured above is invisible to a single score). Fix: separate and weight the stages.
2. Transcript timing. A call is public only after it ends - usually after the close. Computing features from the full transcript and backdating them to the pre-call open is look-ahead by construction (05).
3. Overfitting the segments. Each call yields many per-segment features but few independent quarters; a model with more features than firms is the $p\gg N$ trap again. Fix: keep features few and principal, screen in-fold.
4. Speaker misassignment / transcript errors. ASR transcripts misattribute and garble; a CEO’s answer tagged to an analyst inverts the weighting. Fix: validate speaker tags and weight conservatively.
5. Earnings-quality confound. Tone correlates with the reported numbers; without controlling for the accrual content you may be pricing the accruals, not the words (Accounting Quality & Red Flags).

---

5. References

• Cohen, Lauren, Malloy, Christopher & Nguyen, Quoc, “Lazy Prices,” Journal of Finance 75(3), 2020
• Araci, Dogu, “FinBERT,” arXiv:1908.10063, 2019; Yang, Uy & Huang, “FinBERT,” arXiv:2006.08097, 2020
• Dechow, Patricia, Ge, Weili & Schrand, Catherine, “Understanding Earnings Quality,” 2010
• Healy, Paul & Wahlen, James, “A Review of the Earnings Management Literature,” Journal of Accounting & Economics 27(2), 1999
• Loughran & McDonald, “When Is a Liability Not a Liability?” Journal of Finance 66(1), 2011

---

6. Connected Graph Bridges

• Back: 03 · Sentiment & Tone · 04 · Embeddings & Transformers · 05 · Failure Modes · Index Hub
• Earnings quality: Accounting Quality & Red Flags (the narrative of managed numbers)
• Data & hygiene: Data Sources & Corporate Data (PIT text) · Alternative Data Pipelines & Evaluation
• Models: Deep Learning for Sequences