{"id":"0efd4a81-39cf-4ea3-982e-6c5aa1386c1f","arxiv_id":"2508.18653","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Emotional changes in executives' voices and words during earnings calls explain 43.8% of out-of-sample 30-day volatility even though they do not predict stock return direction.","lead":"Using audio and text from 1,795 earnings calls, this paper reports that executives' emotional shifts predict how volatile a company's stock will be over the next 30 days, while not predicting the direction of returns. The claim matters because it suggests voice patterns can expose hidden corporate uncertainty that scripted statements are designed to hide.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Bootstrap validation lacks temporal/firm blocking; 43.8% R² may be inflated by data leakage.","rationale":"The reader's weakest-assumption identified the same load-bearing concern: the bootstrap validation in Section 4.3 does not specify any temporal or firm-level split, leaving the 43.8% out-of-sample R² vulnerable to leakage. This is the single most critical issue because the paper's central empirical claim—that multimodal emotional features predict future volatility—depends entirely on the evaluation being genuinely out-of-sample. If the result is inflated by leakage, the paper's main contribution evaporates. The paper's own limitations section (Section 6.2) does not mention this validation gap, making it an unacknowledged weakness. The proposed concrete test directly settles whether the concern lands by re-evaluating under strict temporal and firm-blocked splits. If the R² collapses, the claim is not supported; if it survives, the result is credible. I agree with the reader's assessment, so the verdict remains CONDITIONAL pending the requested validation details.","tokens_in":10561,"tokens_out":5000,"duration_ms":59941,"concrete_test":"Rerun the 30-day volatility prediction from Section 5.3 under two strict validation schemes: (1) temporal split—order all 1,795 calls by earnings date; train on the earliest 70% and test on the latest 30%, ensuring no training call's 30-day label window overlaps the test period (discard training calls with dates within 30 days of the cutoff); (2) firm-blocked 5-fold CV—randomly split the 283 firms into 5 folds, train on 4 folds, test on the held-out firms. Report out-of-sample R² and compare with 0.438. If either R² falls sharply (e.g., below 0.2), the original result is not robust to leakage; if both remain above ~0.3, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline result (out-of-sample R²=0.438 for 30-day realized volatility, Section 5.3) rests entirely on the bootstrap validation described in Section 4.3 as 'random resampling of the training data in each iteration.' No temporal or firm-level blocking is mentioned. Because the dataset spans 2018–2023 with quarterly calls from 283 firms, random resampling can place the same firm's calls in both the training and evaluation folds, and more importantly, the 30-day realized-volatility labels for neighboring calls share overlapping calendar-time windows. Common market-wide volatility shocks during those windows are then observable in the training set, allowing the model to fit a latent time factor that artificially inflates the out-of-sample R². The paper never demonstrates that the reported R² survives a split that is strict in time or by firm. This is load-bearing because the central claim—that emotional dynamics carry incremental volatility information—is only meaningful if the evaluation is genuinely out-of-sample.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a multimodal pipeline for forecasting market volatility from earnings calls. Raw teleconference audio is processed by PIAM, a wav2vec 2.0-based multi-task model regularized by a Westervelt-equation physics residual, to produce transcripts and acoustic emotion labels; an LLM (DeepSeek-R1) produces textual emotion labels. Both modalities are mapped onto a three-dimensional Affective State Label (ASL) space with dimensions Tension, Stability, and Arousal, and aggregate/delta features are fed to an XGBoost regressor to predict 1-, 7-, and 30-day CAR and realized volatility. Using 1,795 calls from 283 NASDAQ firms (2018-2023), the paper reports that multimodal features have no predictive power for CAR but explain 43.8% of out-of-sample variance in 30-day realized volatility, versus 25.1% for a historical-volatility baseline. An ablation study attributes the gain to complementary acoustic and textual signals. The central regression is not circular: the ASL coordinates are not fitted to volatility outcomes. However, the headline result rests on a bootstrap validation procedure that may not be genuinely out-of-sample, and several supporting evaluations are missing or under-reported.","tokens_in":10823,"tokens_out":4229,"duration_ms":58035,"significance":"If the empirical claim holds, it would be a notable contribution: paralinguistic and textual emotion dynamics from earnings calls would carry incremental information about future uncertainty beyond historical volatility, and the asymmetry between return prediction (null) and volatility prediction (strong) would be an interesting and falsifiable result. The paper also contains useful engineering pieces: a multi-task acoustic model with a nonlinear-acoustics regularizer, a human-grounded ASL mapping, an ablation design, and evaluations on the public DNS challenge for noise suppression. The authors provide detailed hyperparameters and a reproducibility promise. The main value is conditional on the out-of-sample claim being credible; currently the validation protocol does not establish that. The paper would be substantially strengthened by block-resampled or strict temporal/firm-holdout validation, confidence intervals for R2, and a direct evaluation of PIAM's emotion-classification accuracy.","major_comments":[{"comment":"The headline R2=0.438 is produced by a bootstrap validation described only as \"random resampling of the training data in each iteration.\" With 2018-2023 quarterly calls per firm and 30-day realized-volatility labels that overlap in calendar time across calls, random resampling can place the same firm or overlapping time windows in both training and evaluation. This can inflate out-of-sample R2 through a latent time factor and is a genuine leakage risk. Please report results under strict temporal splits (e.g., train on calls before date T, test after) and firm-holdout splits, with the exact split boundaries and number of folds/iterations. Block bootstrap or Purged K-Fold CV would address this directly.","section":"§4.3, §5.3"},{"comment":"PIAM is the source of the acoustic ASL features, but its standalone evaluation in Table 3 is limited to DNS noise-suppression metrics (MOS-LQO/PESQ) and Table 4 reports only RTF. There is no evaluation of emotion-classification accuracy on earnings-call audio or on any labeled emotion corpus. If the discrete acoustic emotion labels are inaccurate, the acoustic ASL features are noisy in an unquantified way, which weakens the multimodal claim. Please report emotion classification accuracy/F1 (e.g., on IEMOCAP or a labeled earnings-call subset) and, ideally, transcription WER, so the quality of the acoustic channel can be assessed.","section":"§5.1, §5.3"},{"comment":"The paper states that CAR R2 values are \"statistically indistinguishable from zero\" and that 43.8% is \"remarkable,\" but no confidence intervals or significance tests are reported for any R2. Given 1,795 observations and roughly 150 features, the difference between Multimodal (0.438) and Factors-Only (0.251) needs uncertainty quantification. Please provide bootstrap or block-bootstrap CIs for all R2 values and a paired comparison (e.g., corrected resampled t-test or Diebold-Mariano) between Multimodal and Factors-Only/Text-Only. Without this, the ablation \"decisively quantifies\" claim is not supported.","section":"§5.3, Table 5"},{"comment":"The ASL mapping in Table 1 is a fixed input to the features, but Appendix A gives no participant count, inter-rater reliability, or distribution of annotations for the coordinates. Since the coordinates are explicitly adapted for financial risk (e.g., happiness has negative Tension), the reliability of this mapping matters for the volatility result. Please report the number of annotators, agreement metrics (ICC or Krippendorff's alpha), and sensitivity of the main result to the ASL coordinate values.","section":"§3.4, Appendix A"}],"minor_comments":[{"comment":"The text says \"substantially outperforms\" and \"decisively quantifies,\" but the table lacks error bars and sample sizes per cell. Adding CIs would also clarify the surprising pattern that Acoustic-Only and Text-Only underperform Factors-Only while the combination far exceeds it.","section":"§5.6, Table 5"},{"comment":"The physics regularization weight λ=0.01 is stated in Appendix B, but no sensitivity analysis or ablation with λ=0 is provided. A reader cannot tell whether the Westervelt residual, as opposed to the multi-task architecture, contributes to noise robustness.","section":"§3.2, Eq. (3)"},{"comment":"The definition of realized volatility is not given (e.g., sum of squared returns? annualization?), and the CAR computation is not described (market model? estimation window?). This makes the financial targets hard to reproduce.","section":"§4.2, Feature Engineering"},{"comment":"The claim that low concordance between modalities is \"a strength\" is plausible, but it would be more convincing with a quantitative concordance statistic (e.g., Cohen's kappa or mutual information) rather than only a histogram.","section":"§5.2, Figure 2"},{"comment":"The reproducibility statement promises release only \"upon acceptance.\" For an arXiv preprint, providing the emotion-classification evaluation and split definitions in the supplement would greatly help reviewers and readers assess the claims now.","section":"Appendix C"}],"recommendation":"major_revision","confidential_remarks":"The central empirical claim is promising but the validation protocol is not yet convincing. The paper's own text in §4.3 never states a temporal or firm-level split, which is precisely the load-bearing issue. I would advise the editor to invite a revision that reruns the predictive experiments under strict time-based and firm-holdout validation and reports CIs. If the 43.8% R2 collapses under a proper split, the paper should be reconsidered. Also, the paper appears to be formatted as an AAAI submission (copyright notice) but is submitted to cs.LG; this is not a technical problem but worth noting for scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Hi,\n\nThe headline result is exactly what the abstract says: earnings-call emotion dynamics explain 43.8% of out-of-sample variance in 30-day realized volatility and nothing for returns. If that holds, it's a real alternative-data signal. The paper is clear about what it did, and the scripted-to-Q&A delta features are a sensible way to capture stress. The ablation design is standard and the limitations section is honest about correlation vs causation and demographic bias.\n\nBut I don't trust that 43.8% yet. The validation in Section 4.3 is a bootstrap with random resampling of the training data, and there is no mention of a temporal or firm-level split. That is a problem. The 30-day volatility labels for overlapping calls share calendar time, so common market shocks can be picked up from the training set and appear as \"prediction\" in the evaluation. Same-firm calls can also be in both train and test. This is exactly the kind of leakage that inflates out-of-sample R2. The stress-test note is right: the paper never shows the result survives a strict time or firm split.\n\nThere is also a mismatch in the physics part. The authors say they regularize with the Westervelt residual, but Equation 2 defines Lphys as the mean squared norm of an MLP output, not the PDE residual from Equation 1. And they say they compute temporal derivatives, but the stated residual involves a Laplacian. That doesn't invalidate the financial result, but it means the \"physics-informed\" claim isn't backed by the equations as written.\n\nSome smaller issues: no confidence intervals for the headline R2, no benchmark of the emotion classifier itself (only noise suppression on DNS), and the reproducibility statement promises code/data only upon acceptance. These are all addressable.\n\nMy verdict: the core empirical claim is new and worth taking seriously, but the validation is load-bearing and not yet established. This paper deserves peer review, not as a rejection but as a request to re-run the evaluation with strict time and firm blocking, report the split, and clean up the regularizer definition. I would bring it to a reading group to discuss leakage in ML-for-finance.\n\nBest","headline":"Earnings-call emotion predicts volatility but not returns — interesting if real, but the bootstrap validation needs a temporal split before the headline R2 is credible.","tokens_in":11298,"tokens_out":3949,"would_cite":false,"duration_ms":42059,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Executives' emotional dynamics in earnings calls predict future market volatility, not stock returns, with a 43.8% out-of-sample R2 for 30-day volatility.","keywords":["earnings calls","market volatility forecasting","computational paralinguistics","affective state label","physics-informed neural networks","multimodal finance","vocal stress","realized volatility"],"falsifier":"Run the identical feature pipeline with a strict temporal split (all training calls before a cutoff date, all test calls after) and with all calls from a given firm confined to one fold. If the 30-day volatility R-squared falls to roughly the 0.251 financials-only baseline, the multimodal outperformance is an artifact of leakage rather than emotional signal.","tokens_in":10481,"feed_emoji":"🎙️","tokens_out":6396,"duration_ms":69787,"temperature":0.7,"pith_summary":"This paper tries to show that how executives sound during earnings calls carries information about future market risk that traditional text analysis and historical prices miss. Using a physics-informed acoustic model that extracts vocal emotion from raw, distorted teleconference audio, plus a large language model that labels the emotional tone of transcripts, the authors map both channels onto a common three-dimensional affect space. The central finding is a split: these multimodal emotional features do not predict future stock returns, but they explain up to 43.8% of out-of-sample variance in 30-day realized volatility. The strongest signals come from the shift between scripted presentation and spontaneous Q&A, especially a drop in the CFO's textual stability, acoustic instability from CFOs, and arousal variability from CEOs. If the result holds, monitoring executive affect in earnings calls could give investors and regulators a direct, hard-to-manipulate gauge of hidden corporate uncertainty.","feed_headline":"Executive emotion in earnings calls predicts 43.8% of volatility","feed_subtitle":"Voice and text cues from the shift to Q&A reveal hidden corporate uncertainty that prices and transcripts alone miss.","key_machinery":"PIAM, the Physics-Informed Acoustic Model, is a multi-task neural network that takes raw earnings-call audio and simultaneously produces a transcript, a discrete acoustic emotion label, and acoustic event labels. It is regularized by the Westervelt equation, a nonlinear-acoustics PDE that penalizes latent-space trajectories inconsistent with nonlinear sound propagation, which the authors argue makes it robust to teleconference clipping and compression. The discrete labels, acoustic and textual, are mapped onto a three-dimensional Affective State Label (ASL) space—Tension, Stability, Arousal—anchored by a human annotation study. The predictive features are statistical moments of these ASL val","core_discovery":"The core claim is the predictive divergence between risk and return. On 1,795 earnings calls from 283 NASDAQ firms (2018-2023), the authors build around 150 features per call from executive affect, including moments of tension, stability, and arousal and the change in those states from prepared remarks to Q&A. An XGBoost model using these features achieves out-of-sample R2 of 0.438 for 30-day realized volatility, compared to 0.251 for a financials-only baseline using historical volatility, while out-of-sample R2 for cumulative abnormal returns is statistically indistinguishable from zero. The paper interprets this as evidence that executive emotion is not a directional signal but a barometer","pith_inferences":["A strict temporal or firm-level split is the natural check: if the 43.8% R2 survives, the signal is real; if it collapses to the baseline, the reported out-of-sample gain is leakage.","The presentation-to-Q&A delta design could be ported to other scripted-then-spontaneous disclosure settings, such as central-bank briefings or IPO roadshows, as a ready-made stress test.","Because returns are unpredictable but volatility is, the acoustic-textual signal may be pricing a volatility risk premium; the paper does not test tradeability, so that remains an open question.","The ASL emotion coordinates were validated on North American public-company speech; applying the pipeline to other dialects, languages, or non-corporate speakers would require re-anchoring the mapping."],"forward_implications":["The full multimodal model adds roughly 18 to 19 percentage points of out-of-sample R2 over the financials-only baseline for 30-day volatility, so acoustic and textual emotion carry information beyond historical price swings.","Emotional cues do not forecast cumulative abnormal returns, meaning they signal uncertainty about future performance rather than the direction of that performance.","The most stable predictors are role-specific and context-specific: a CFO's drop in textual stability and a CEO's arousal variability in Q&A dominate the top features, so who speaks and when matters.","PIAM's real-time factor below 1.0 means the acoustic channel could be monitored live as calls happen, not just in hindsight.","If markets already price textual disclosure but not vocal affect, regulators and investors could use this channel to detect hidden uncertainty earlier than price-based measures reveal it."],"supporting_citations":[{"why":"Supplies the computational-paralinguistics premise that vocal properties carry emotional and cognitive state.","marker":"Schuller and Batliner 2013"},{"why":"Provides prior evidence that vocal parameters can predict financial and labor-market outcomes, motivating the acoustic channel.","marker":"Mayew, Parsons, and Venkatachalam 2013"},{"why":"Provides the wav2vec 2.0 self-supervised encoder that forms PIAM's acoustic foundation.","marker":"Baevski et al. 2020"},{"why":"Supplies the physics-informed neural network technique used to compute the PDE residual regularizer.","marker":"Raissi, Perdikaris, and Karniadakis 2019"},{"why":"The Deep Noise Suppression benchmark data used to evaluate PIAM's speech-quality performance.","marker":"Reddy et al. 2020"},{"why":"Provides the circumplex model of affect that the ASL space adapts into Tension, Stability, and Arousal.","marker":"Russell 1980"},{"why":"Grounds the tension dimension in established financial textual analysis of risk tone.","marker":"Loughran and McDonald 2011"},{"why":"Marks the limitation of text-only LLM analysis in finance that the multimodal approach aims to overcome.","marker":"Li et al. 2023"}],"fun_headline_variants":["CEO and CFO voices predict 43.8% of market volatility","Earnings call emotion foretells volatility, not returns","Tone shifts in earnings calls explain 43.8% of risk","Executive affect on calls predicts volatility, not returns","How executives sound on earnings calls signals volatility"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The headline 43.8% out-of-sample R-squared depends on the bootstrap validation being a genuine holdout, but the paper never states a temporal or firm-level split; overlapping calls from the same firm could leak between training and evaluation.","fun_headline_variants_meta":{"raw":{"variants":["CEO and CFO voices predict 43.8% of market volatility","Earnings call emotion foretells volatility, not returns","Tone shifts in earnings calls explain 43.8% of risk","Executive affect on calls predicts volatility, not returns","How executives sound on earnings calls signals volatility"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000211,"raw_usage":{"total_tokens":1290,"prompt_tokens":820,"completion_tokens":470,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":564,"completion_tokens_details":{"reasoning_tokens":389}},"tokens_in":564,"tokens_out":470,"duration_ms":4855,"temperature":1.0,"reasoning_tokens":389,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T16:19:08.424938+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the identical feature pipeline with a strict temporal split (all training calls before a cutoff date, all test calls after) and with all calls from a given firm confined to one fold. If the 30-day volatility R-squared falls to roughly the 0.251 financials-only baseline, the multimodal outperformance is an artifact of leakage rather than emotional signal.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the computational-paralinguistics premise that vocal properties carry emotional and cognitive state."},{"cited_title":"J.; Parsons, C","cited_arxiv_id":null,"evidence_quote":"Provides prior evidence that vocal parameters can predict financial and labor-market outcomes, motivating the acoustic channel."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the wav2vec 2.0 self-supervised encoder that forms PIAM's acoustic foundation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the physics-informed neural network technique used to compute the PDE residual regularizer."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the circumplex model of affect that the ASL space adapts into Tension, Stability, and Arousal."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Grounds the tension dimension in established financial textual analysis of risk tone."}],"review_version":1}