Empirical note · 14 October 2026 · cst-5cr5.10
We found plumbing.
Not alpha.
A 14-quarter test of earnings-call tone change, prepared-remarks reuse, and answer evasion produced no investable standalone signal. The evasion proxy failed its own calibration before the market result could matter.
The panel contains 17,641 liquid U.S. common-share call pairs from 1,844 tickers, 2023-01-05 through 2026-06-30. Returns begin after the call-day reaction and run to 20, 40, or 60 trading sessions versus SPY. U1
01 / Decisions
Four tests. Four stops.
“No-go” means the observed effect is too uncertain, too unstable, or not validly measured—not that the underlying economic mechanism is impossible.
Loughran–McDonald Q&A tone delta
60d Q5−Q1 −29.98 bp · 95% CI [−107.02, +46.26]
The 20- and 40-session signs are positive, the 60-session sign is negative, quintile returns are not monotone, and every robustness interval crosses zero. The baseline validates plumbing; it does not reproduce a usable drift trade here. T2
Explicit anticipatory coverage ratio
60d with-reference minus without −11.70 bp · 95% CI [−61.47, +41.61]
Prepared-remarks and earlier-point back-references appear in 29.9% of usable calls, but their return effect is small and remains null across horizons, splitter confidence, and question-toughness cuts. A3
Implicit anticipatory coverage ratio
60d Q5−Q1 −67.26 bp · 95% CI [−143.05, +7.31]
High answer-to-prepared similarity is directionally bearish and nearly excludes zero after length adjustment, but 14 quarterly spreads are unstable, HAC evidence is weak, every interval includes zero, and the sign was ambiguous ex ante. Do not promote a near miss discovered among many cuts. A4
Answer evasion and sector clustering
Calibration AUC 0.450 · expected direction > 0.500
The scorer separates a synthetic direct answer from a deflection, yet literal non-answer phrases score less evasive than unflagged exchanges. Its primary 60-session return interval also crosses zero. Neither alpha nor policy output is actionable. E2
02 / Design
The test was built to fail honestly.
Raw transcripts stayed on the authorized host. Every row preserves fiscal identity, call date, transcript IDs, and content hashes; ambiguous boundaries were excluded rather than guessed.
Section gate
Derived, then audited
Neither FMP nor StockQL supplies a prepared/Q&A boundary. A deterministic char-offset splitter reached 886/943 usable splits (94.0%) on the fixed validation sample; 3 legitimate no-Q&A calls remained a separate verdict and 54 unresolved calls stayed quarantined. G2
Outcome
Reaction skipped
Abnormal return is the stock’s cumulative close-to-close return divided by SPY’s, from the first market session strictly after the call date through sessions 20, 40, or 60. Call day and the first post-call close do not enter the return. U1
Inference
Time and issuer dependence
Signals are ranked into equal-weight quintiles within each call calendar quarter. The report shows observation-level effects, the mean of 14 quarterly spreads with Newey–West lag 1, and 2,000-draw ticker-cluster bootstrap intervals. U2
Universe
Liquid, not index-bound
U.S. normalized common shares; no current-active filter; pre-call median 60-session dollar volume at least $25 million; last pre-call close at least $5; complete return horizons; immediate fiscal-quarter pairs separated by 45–150 days. U1
Leakage
Frozen text, forward returns
Features use only call-time text. Every transcript hash must match the inherited panel before extraction; liquidity and price floors are pre-call; the embedding model is frozen and never receives returns. Completed-quarter ranks are research sorts, not live thresholds. A1
Privacy boundary
Aggregates leave the box
Committed artifacts contain aggregate statistics, IDs, and cryptographic hashes—not transcript text, participant text, raw panel rows, connection strings, or credentials. P1
03 / Results
The intervals do the deciding.
Effects are basis-point differences in abnormal return. Continuous features show within-quarter Q5 minus Q1; explicit ACR shows calls with any reference minus calls without one.
Baseline / tone
No stable horizon and no monotone payoff
The primary tone feature is current-minus-prior Q&A finance tone, where tone is (positive words − negative words) / all alphabetic tokens using the pinned March 2026 Loughran–McDonald dictionary. T1
| Horizon | Q5−Q1 | Quarterly NW mean | NW t | Ticker-cluster 95% CI |
|---|---|---|---|---|
| 20 sessions | +22.20 bp | +19.88 bp | +0.855 | [−17.56, +64.07] |
| 40 sessions | +12.03 bp | +10.84 bp | +0.415 | [−48.15, +71.78] |
| 60 sessions | −29.98 bp | −27.09 bp | −0.913 | [−107.02, +46.26] |
ACR / explicit
Cheap to compute. Too close to zero.
The explicit feature flags prepared/opening-remarks and earlier-point back-references, normalized by full Q&A length and also tested as any-reference incidence. There are 5,255 flagged calls among 17,570 usable features. A2
| Horizon | With−without | Quarterly NW mean | NW t | Ticker-cluster 95% CI |
|---|---|---|---|---|
| 20 sessions | −9.64 bp | −13.36 bp | −0.760 | [−36.10, +17.97] |
| 40 sessions | −20.56 bp | −21.68 bp | −0.994 | [−58.67, +19.19] |
| 60 sessions | −11.70 bp | −14.18 bp | −0.574 | [−61.47, +41.61] |
ACR / implicit
A suggestive negative direction is still a null
Each answer chunk takes its maximum cosine similarity to any prepared-remarks chunk; the call score is answer-token-weighted. Similarity correlates 0.490 with log prepared-chunk count, so length-residualization is mandatory—not optional. A5
| Variant | Q5−Q1 | Quarterly NW mean | NW t | Ticker-cluster 95% CI |
|---|---|---|---|---|
| 20 sessions | −33.71 bp | −31.11 bp | −1.152 | [−75.11, +9.20] |
| 40 sessions | −40.33 bp | −31.67 bp | −0.534 | [−97.97, +20.81] |
| 60 sessions | −67.26 bp | −56.51 bp | −0.604 | [−143.05, +7.31] |
| 60d, length-adjusted | −74.13 bp | −61.92 bp | −0.657 | [−145.06, +1.69] |
| 60d, tier-1 split only | −76.34 bp | −69.60 bp | −0.813 | [−158.64, +3.80] |
Evasion / alpha
The validity failure precedes the null return
Evasion is 1 − cosine(question, answer), averaged within fixed finance topics and differenced from the prior call only for topics present in both. It yields 17,285 events from 20,188 unique current-or-prior calls. E1
| Horizon | Q5−Q1 | Quarterly NW mean | NW t | Ticker-cluster 95% CI |
|---|---|---|---|---|
| 20 sessions | −7.22 bp | −9.05 bp | −0.613 | [−46.61, +32.87] |
| 40 sessions | −8.51 bp | −9.92 bp | −0.365 | [−68.54, +52.99] |
| 60 sessions | −28.43 bp | −28.47 bp | −1.183 | [−100.82, +45.29] |
04 / Policy mode
A watchlist-shaped output is not a warning system.
Alpha asks whether a signal predicts returns. Policy early warning asks whether coordinated disclosure behavior identifies a real-world event. The latter needs valid labels and event adjudication; this prototype has neither.
Bounded demonstration / 2026-Q2
Healthcare × capital allocation
The top descriptive cluster contains eight companies whose matched-topic evasion deltas are all positive against 56 earlier observations in the same current-sector/topic cell.
- Companies
- 8
- Mean historical z
- +1.160
- Positive share
- 100%
- Eligible groups
- 122
Non-actionable. Nine groups pass the descriptive screen, but the underlying scorer fails calibration; sector metadata is current rather than point-in-time; and 122 simultaneous groups are not multiplicity-adjusted. E5
| Sector × topic | Companies | Prior n | Mean z | Positive share |
|---|---|---|---|---|
| Healthcare × capital allocation | 8 | 56 | +1.160 | 100% |
| Basic Materials × balance sheet | 5 | 72 | +0.856 | 100% |
| Healthcare × inventory / supply chain | 5 | 89 | +0.825 | 80% |
The screen deduplicates to one observation per company within current-sector × fine-topic, requires at least three companies and 30 earlier sector-topic observations, and flags mean z ≥ 0.5 with at least two-thirds positive deltas. The thresholds are descriptive, not validated operating points.
05 / Evidence boundary
Alpha and policy fail for different reasons.
A market null does not invalidate a policy use. A policy-shaped cluster does not rescue a market null. Each track has its own promotion standard.
Investable-alpha standard
Not met
- A calibrated feature with a stable economic interpretation.
- Uncertainty excluding zero after issuer dependence and time variation.
- Consistent horizon, subperiod, section-confidence, and interaction behavior.
- A locked specification tested on a non-overlapping future period.
Policy early-warning standard
Not met
- Human-labeled answer-responsiveness data that validates the scorer in this corpus.
- Point-in-time sector and exposure mappings rather than present-day labels.
- Multiplicity control and an explicit false-alert budget across topic/sector cells.
- Independent adjudication against dated policy or operating events.
06 / Final call
Keep the rail. End the signal hunt.
The durable asset is the measured section splitter, identity-preserving panel, return construction, local embedding path, and failure-aware inference. None of the four tested features earns another alpha iteration on these results.
Reopen implicit ACR only if a separately motivated, pre-registered hypothesis fixes its sign and interaction before seeing new returns. Reopen evasion only after an independent labeled set validates the scorer. Those are entry conditions, not recommended follow-up work. The default decision is stop.
07 / Prior work
What this does—and does not—claim.
This is a bounded replication and prototype, not evidence against the broader literature. A 2023–2026 liquid-U.S. panel, one dictionary, one static embedding model, and 14 quarterly spreads cannot settle every specification.
- Loughran & McDonald (2011) motivates a finance-specific dictionary rather than a general-purpose negative-word list.
- Price, Doran, Peterson & Bliss (2012) documents incremental information in earnings-call textual tone; our bounded replication does not produce a usable standalone return spread.
- Cohen, Malloy & Nguyen (2020), “Lazy Prices” motivates change-based text features. It studies financial-report changes, not this Q&A specification.
- Allee, Do & Do, “Can Analysts Elicit Useful Information by Asking Unique Questions in Earnings Conference Calls?” is close prior art for question uniqueness relative to prepared remarks. ACR here is narrower: management answers reused against prepared remarks. The prepared/Q&A similarity space is not claimed as untouched.
- Gow, Larcker & Zakolyukina (2021) provides established conference-call non-answer measurement; our literal phrase family is only a directional control.
- Hynes, Chiang, Garvey & O’Brien (2026) reports human validation and return predictability for Q/A semantic similarity. Our faster static scorer does not reproduce their calibration and therefore cannot inherit their interpretation.
- minishlab/potion-base-8M model card documents the frozen MIT static embedding used for ACR and evasion. Exact snapshot and file hashes are in the evidence appendix.