The Hype Index: an NLP-driven Measure of Market News Attention

Helyette Geman; Wanchaloem Wunkaew; Zheng Cao

REVIEW 5 major objections 7 minor 8 references

The Hype Index: an NLP-driven Measure of Market News Attention

T0 review · 5 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims a stock's share of financial headlines, divided by its share of market capitalization, reveals when media attention is out of proportion to economic size.

desk verdict The Hype Index is a clean descriptive metric, but the paper's advertised predictive claims are explicitly disclaimed in its own Section 5.2. read the letter →

arxiv 2506.06329 v1 pith:IAP3SW7N submitted 2025-05-30 q-fin.ST cs.CEcs.CL

Zheng Cao , Wanchaloem Wunkaew , Helyette Geman This is my paper

classification q-fin.STcs.CEcs.CL MSC 91G8062P0591B84

keywords HypeIndexmediaattentionnaturallanguageprocessingmarketsignalingstockvolatilityS&P100capitalizationadjustmentinvestorsentiment

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper sets out to show that media 'hype' can be measured directly as the share of financial headlines mentioning a stock or sector, without reading tone. Because large companies naturally appear in the news more often, it constructs a second measure: that headline share divided by the firm's share of S&P 100 market capitalization, so that a value above 1 flags attention out of proportion to economic size. Both indices are built for the S&P 100 over 326 trading days and examined through sector clusters, correlations with returns and volatility, and spikes around events such as the August 2024 market selloff. The payoffs would be a simple, interpretable attention metric for volatility analysis and market signaling.

What carries the argument

The load-bearing object is the ratio $CapHypeIndex_{i,t} = (N_{i,t}/\sum_j N_{j,t}) / (MC_{i,t}/\sum_j MC_{j,t})$. The numerator is a stock's share of daily news mentions; the denominator is its market-capitalization weight within the S&P 100 universe. The ratio converts raw attention into attention per unit of economic size, with 1 as the neutrality benchmark, and the paper reads sustained deviations from 1 as hype or neglect. Sector-level versions are computed by summing constituent ticker hype indices within each GICS sector, and cluster labels are assigned from the resulting trajectories.

What would settle it

Take a random sample of roughly 500 headlines that the LSEG pipeline tagged to S&P 100 tickers and have independent annotators judge whether the tag matches the company actually discussed. If the false-tag rate is high or systematically concentrated in certain sectors, then correcting the tags would move the capitalization-adjusted cluster rankings and the index would fail as an unbiased attention measure.

Watch

Extended reading notes

Core claim

The central claim is that the Hype Index family quantifies attention distortions: $HypeIndex_{i,t} = N_{i,t}/\sum_j N_{j,t}$ gives the fraction of all S&P 100 news mentions going to stock $i$, and $CapHypeIndex_{i,t}$ divides that fraction by the stock's market-capitalization weight. The paper argues that this ratio is a valid signal of over- or under-hyping and supports it with three empirical observations: the raw index puts Financials and Information Technology at three to four times the average coverage; the capitalization-adjusted version moves Information Technology into the 'less prominent' cluster and pushes Real Estate, Industrials, and Utilities into the 'relatively hyped' cluster; and the two indices are strongly correlated, with sector-level correlations from 0.82 to 0.98. The paper also reports that adjusted hype moves with VIX and GPR changes around stress events and that normality is rejected for the adjusted index and its percent changes.

Load-bearing premise

The index inherits the accuracy of LSEG/Refinitiv's proprietary entity-recognition tags: if a headline is mapped to the wrong ticker, or if a bare article count assigns equal weight to a one-word mention and a full analysis, every hype value, cluster, and correlation inherits that error.

Editorial extensions

If this is right

Financials and Information Technology receive three to four times the market-average news share over the sample, while Utilities, Real Estate, and Materials receive less than half the average.
Adjusting for market capitalization reverses the picture: Information Technology becomes less prominent relative to its size, while Real Estate, Industrials, and Utilities look relatively hyped.
Because sector-level correlations between the raw and capitalization-adjusted indices run from 0.82 to 0.98, raw news share can proxy for the adjusted measure whenever market-capitalization weights are slow-moving.
Spikes and troughs in capitalization-adjusted hype concentrate around identified market events, including the August 2024 selloff, the November 2024 election rally, and the April 2025 tariff shock.
The normality tests reject a normal model for the adjusted index and its percent changes, which matters for any later statistical use of the index.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The descriptive event analysis does not by itself prove forecasting power; a strict out-of-sample test asking whether a surprise jump in adjusted hype predicts next-day or next-week volatility would settle that question.
Because the adjusted index is a ratio of two weights, its daily variation is dominated by the news numerator when market caps are stable; the distinct information in the capitalization adjustment is likely concentrated in stress episodes rather than in the daily series.
Counting headlines treats every mention equally; weighting by source reach or combining with sentiment would separate 'hype' from 'information,' a natural extension of the same data pipeline.
The strong raw-adjusted correlation suggests the index would be most useful as an attention-risk factor for volatility modeling or event detection, not as a standalone directional return signal.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

This paper is a descriptive attention index that reads like a validated forecasting tool, and that mismatch is the whole story. The raw Hype Index is a news-count share; the adjusted version divides it by market-cap weight. Both are clean definitions, and the data-collection section is honest about the LSEG pipeline and its entity-recognition limitations. But the abstract and conclusion promise associations with returns, volatility, and VIX, and Section 5.2 says outright: 'We do not directly compute the relationship between changes in the Hype Index and subsequent market dynamics.' So the central claim is not tested in the manuscript.

What is actually new? Not much. Media-coverage share is a known attention measure—the paper doesn't cite Da, Engelberg, and Gao or the broader investor-attention literature—and the cap-adjusted variant is a mechanical ratio. The high correlations in Table 4 are mostly forced because the denominator (market-cap weight) is stable for mega-caps; the authors even acknowledge this structural relationship. The sector clustering and normality tests are competent descriptive work, but they don't establish signaling power.

The soft spots are proportionate to this: the conclusion contradicts the methods section, the forecasting claims lean on a self-citation to Cao and Geman (2025), the manuscript contains editorial artifacts ('Mention what tickers are removed,' 101 firms for the S&P 100), and there is no code or data. The LSEG ticker-mapping assumption is a real secondary concern, but even with perfect mapping, the paper still doesn't test its advertised claim.

Who gets value from this? Someone who wants a transparent, count-based attention dashboard for descriptive purposes. As a scientific contribution, it's not ready. I would not bring it to a reading group, I would not cite it, and if I were the editor I would desk-reject or request major revision that actually estimates the lagged hype-market relationships advertised in the abstract.

Referee Report

5 major / 7 minor

Summary. The paper defines a News Count-Based Hype Index (the share of daily news mentions for each S&P 100 stock or sector) and a Capitalization Adjusted Hype Index (that share divided by the stock's or sector's market-capitalization weight). Using LSEG/Refinitiv news headlines for roughly 101 S&P 100 constituents over 2023–2025, it reports sector-level time series, clusterings into hype groups, normality tests, correlations between the two index variants, and visual comparisons with VIX and GPR. The abstract and conclusion assert that the index family has associations with returns, volatility, and the VIX at various lags, and that it has signaling power for short-term market movements.

Significance. The construction is transparent and the definitions are explicit, with a clear link to a simple, interpretable attention measure. If the advertised lagged associations and signaling power were actually demonstrated, the contribution would be of practical interest for volatility analysis and market monitoring. However, the manuscript as submitted does not compute those associations: Section 5.2 explicitly disclaims the direct test, Section 4.4 is a visual comparison, and Section 4.3 delegates forecasting evidence to a separate paper. The significance of the current version therefore rests on the descriptive index construction rather than on the predictive claims that motivate the abstract and conclusion.

major comments (5)

[Section 5.2, Abstract, Conclusion] The abstract advertises 'associations with returns, volatility, and VIX index at various lags' and 'signaling power for short-term market movements,' and the Conclusion repeats the association claim. Section 5.2 states, 'We do not directly compute the relationship between changes in the Hype Index and subsequent market dynamics.' No lagged regression, VAR, panel model, or predictive accuracy statistic appears in the empirical sections. This is a load-bearing mismatch: the paper's central advertised result is absent from its own evidence. The authors should either add formal tests (for example, panel regressions of future returns or realized volatility on lagged hype changes with appropriate controls and clustered standard errors) or rewrite the abstract and conclusion to report only the descriptive findings actually presented.
[Section 4.4] The claimed association with the VIX is supported only by a visual inspection: the text says the indices 'appear to move in tandem' and the figure shows 7-day rolling means. No correlation coefficient, co-movement statistic, or test of statistical significance is reported for the VIX relationship. As written, the Conclusion's statement that the index 'exhibits meaningful associations with ... market sentiment indicators such as the VIX' is not supported by any quantitative result.
[Section 5.1] Hype Momentum is defined in Definition 5.2 but never operationalized or estimated. The text says its empirical roles are 'further explored in Section 5,' yet Section 5.2 explicitly disclaims a direct computation of hype-to-market relationships. The Conclusion sentence that 'persistent deviations from neutrality often precede significant movements in price or volatility' is therefore unsupported. The authors need to provide an estimator for Hype Momentum and report its empirical performance, or remove the claim.
[Section 4.3] The sentiment forecasting evidence is delegated to a separate paper by the same authors; the current manuscript states that 'the authors of this paper ... have investigated' the relationship in 2025. This means one of the four advertised evaluation lenses, namely signaling power, is not actually evaluated here. If this is intended as background literature, it should be presented as such and removed from the list of evaluations claimed for this paper.
[Section 4.1, Table 4] The high correlations in Table 4 are largely structural: by Definition 3.1, CapHypeIndex_{i,t} = HypeIndex_{i,t} / MarketCapWeight_{i,t}, and the paper itself notes in Section 4.1 that market-cap weights are relatively stable for large caps. The claim that one index 'can serve as an effective proxy for the other' therefore needs a stronger test of incremental information, for example whether the capitalization-adjusted index explains any variation in future volatility or returns beyond the raw index and a market-cap control.

minor comments (7)

[Section 1.2] The text contains an unresolved editorial instruction: 'Mention what tickers are removed We remove X.TSLA...' The final sample is described as 101 companies, which does not match the stated S&P 100 universe; please clarify the exact list of included and removed tickers.
[Section 4.2] The paper states that around August 5, 2024, the S&P 500 experienced a decline of 'over 10% over a single weekend.' The actual S&P 500 move on that Monday was roughly 3%; please correct the figure or provide a source for the 10% claim.
[Figure 9] The figure caption contains a typo: 'Hyp' should read 'Hype'.
[Section 3.4] The normality tests are reported without the sample size or a clear connection to any downstream modeling choice. If the normality assumption is not used later, the subsection should be shortened or explicitly framed as descriptive.
[Section 2.2] The normalization description is inconsistent: the caption says 'Scaled by Overall Avg = 1,' while the text says 'scaling by daily average enforces...' Please clarify whether the scaling uses the overall sample average or a daily average.
[Section 2.1] The sector Hype Index counts a multi-ticker news item once for each mentioned ticker. This is a legitimate modeling choice, but its effect on sector shares and cross-sector comparisons should be discussed explicitly.
[Section 5.2, Figure 10] The regression in Figure 10 reports p-values of 0.0000 for firm-day observations, but the standard errors are not clustered by firm or date. Given the repeated-observation structure, those p-values are likely overstated.

Circularity Check

2 steps flagged · score 6.0 of 10

The high-correlation 'finding' between the two Hype Index variants is mechanical by definition, and the advertised forecasting evidence is imported from the authors' own prior paper after Section 5.2 disclaims direct testing.

self definitional [Section 4.1, Table 4, and Definition 3.1]
"There exist high empirical correlations between the News Weighted Hype Index and the Capitalization Adjusted Hype Index across sectors, suggesting that one can serve as an effective proxy for the other."

The Capitalization Adjusted Hype Index is defined as CapHypeIndex_{i,t} = HypeIndex_{i,t} / MarketCapWeight_{i,t}. When the market-cap weight is stable over time, the correlation between a variable and itself divided by a nearly constant positive denominator is forced to be near 1 by construction. The paper even acknowledges this: 'When the market capitalization weight ... remains relatively stable over time ... the variation in the Capitalization Adjusted Hype Index is largely driven by the numerator.' Presenting this mechanical consequence as an empirical validation of one index as a 'proxy' for the other is a finding that reduces to the definition of the adjusted index, not independent evidence.
self citation load bearing [Section 4.3, with Abstract and Conclusion claims]
"The authors of this paper, Cao and Geman, have investigated the relationship between hype levels and sentiment scores derived from news text in 2025 [2]. They showed that sentiment scores can forecast the direction of stock price and volatility directions with a precision of over 75%."

The paper's advertised central claim is that the Hype Index family has 'signaling power for short-term market movements' and meaningful 'associations with returns, volatility, and market sentiment indicators such as the VIX.' Yet Section 5.2 states: 'We do not directly compute the relationship between changes in the Hype Index and subsequent market dynamics.' The only quantitative support offered for the forecasting claim is a citation to the authors' own prior paper [2], whose predictive results are not re-derived or independently tested in this manuscript. Thus the load-bearing evidence for the central market-signaling claim is a self-citation, and the current paper's own empirical sections are descriptive rather than predictive.

full rationale

The construction of the Hype Index itself is self-contained: it is a news-count share, and the Capitalization Adjusted Hype Index is its ratio to market-cap weight. Those definitions are not circular by themselves. However, two steps weaken the derivation chain materially. First, the high correlations in Table 4 are presented as an empirical finding that the two indices 'can serve as effective proxies' for each other, but since the adjusted index is defined as the raw index divided by a relatively stable market-cap weight, the near-unity correlation is arithmetically forced; this is a self-definitional result rather than independent validation. Second, the paper's headline promise of predictive/signaling power is not tested in the manuscript; Section 5.2 explicitly disclaims direct computation of the relationship between Hype Index changes and subsequent market dynamics. The only cited evidence for forecasting ability is the authors' own prior paper [2], making the central advertised claim load-bearing on self-citation. These issues are real circularity-related defects, but the descriptive index construction and the sector-clustering analysis do contain independent content, so the score is set at 6 rather than higher. The unsupported nature of the forecasting claim is primarily a correctness/evidentiary concern, but to the extent the paper relies on its own prior work for that claim, it is also a circularity concern.

Assumptions & free parameters 5 free parameters · 6 assumptions · 4 invented entities

The index definitions themselves introduce no free parameters, but the empirical section contains fitted regression coefficients and hand-assigned cluster labels. The main axioms are about data quality: entity recognition, the market-cap benchmark, and the validity of article counts as an attention proxy. The invented entities are the two indices and two interpretive concepts; none is validated against an external benchmark or used to generate a falsifiable prediction.

free parameters (5)

Linear regression slope (news weight vs market weight) = 0.2166
Fitted in Section 5.2 on firm-day observations to describe the news-weight/market-weight relationship; not an input to the Hype Index itself.
Linear regression intercept = 0.0078
Same fit as above, reported in Section 5.2.
Power-law coefficient = 2.28
Fitted in Section 5.2 at sector level to characterize the market-cap fraction versus news fraction relationship.
Power-law exponent = 1.41
Same sector-level fit, reported in Section 5.2.
Sector cluster cutoffs (over, neutral, under-hyped) = not specified
Assigned by visual inspection of normalized Hype Index trajectories in Tables 2 and 3, not by a formal algorithm.

assumptions (6)

domain assumption LSEG entity recognition correctly maps news headlines to the intended S&P 100 ticker.
Invoked in Section 1.2 and Section 3.1; the entire Hype Index is built on these ticker-mention counts, and the authors state they do not modify the mapping logic.
domain assumption A news item that mentions multiple tickers can be counted once per ticker without distorting relative attention.
Stated in Section 2.1; the sector Hype Index sums per-ticker counts, so multi-ticker articles receive multiple weight.
domain assumption Market capitalization weight is the appropriate benchmark for expected media attention.
Definition 3.1 divides news weight by market-cap weight; if this benchmark is wrong, the adjusted index's over-hyped and under-hyped labels are not meaningful.
domain assumption The sample of 101 tickers on 326 trading days from December 2023 to April 2025 is representative of S&P 100 attention dynamics.
Collection period and exclusions are described in Section 1.2; no robustness checks across time periods or constituent changes are provided.
domain assumption The maximum retrieval cap of 100,000 headlines per stock-week is never binding, so no relevant news is lost.
Stated in Section 1.2; the authors say average AAPL headlines are about 400 per day, but they do not verify this for all stocks or event days.
domain assumption News article count, independent of tone or content, is a valid measure of investor attention.
The paper explicitly isolates media intensity from sentiment polarity in the introduction, making this a load-bearing modeling choice.

invented entities (4)

Hype Index (News Count-Based)
purpose: Quantify the share of daily news coverage received by a stock or sector.
Defined in Definition 2.1; no external validation or falsifiable prediction is provided beyond descriptive correlation.
Capitalization Adjusted Hype Index
purpose: Measure media attention relative to market capitalization weight.
Defined in Definition 3.1; its interpretation as over-hyped or under-hyped depends on the assumed benchmark and is not independently validated.
Hype Neutrality
purpose: Baseline state where the adjusted Hype Index equals 1.
Defined in Definition 5.1; approximated by the sample mean, but no statistical test shows that deviations from 1 are meaningful.
Hype Momentum
purpose: Intended to measure the speed of price convergence to Hype Neutrality.
Defined in Definition 5.2 but never estimated or tested anywhere in the paper.

how reviews work

0 comments

Cite this review

Pith. "Pith review of The Hype Index: an NLP-driven Measure of Market News Attention." pith.science (2026). https://pith.science/paper/IAP3SW7N

@misc{pith2026250606329,
  author       = {Pith},
  title        = {Pith review of: The Hype Index: an NLP-driven Measure of Market News Attention},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IAP3SW7N}},
  note         = {Machine review of arXiv:2506.06329}
}

read the original abstract

This paper introduces the Hype Index as a novel metric to quantify media attention toward large-cap equities, leveraging advances in Natural Language Processing (NLP) for extracting predictive signals from financial news. Using the S&P 100 as the focus universe, we first construct a News Count-Based Hype Index, which measures relative media exposure by computing the share of news articles referencing each stock or sector. We then extend it to the Capitalization Adjusted Hype Index, adjusts for economic size by taking the ratio of a stock's or sector's media weight to its market capitalization weight within its industry or sector. We compute both versions of the Hype Index at the stock and sector levels, and evaluate them through multiple lenses: (1) their classification into different hype groups, (2) their associations with returns, volatility, and VIX index at various lags, (3) their signaling power for short-term market movements, and (4) their empirical properties including correlations, samplings, and trends. Our findings suggest that the Hype Index family provides a valuable set of tools for stock volatility analysis, market signaling, and NLP extensions in Finance.

Figures

Figures reproduced from arXiv: 2506.06329 by the authors.

**Figure 1.** Raw Sector Hype Index (Unscaled) The News Coverage Share level directly indicates the portion of news attention of each sector with respect to the complete S&P 100. To interpret the patterns observed in the plots, we categorize sectors into three relative clusters based on their long-term positioning in the normalized Hype Index trajectories: overhyped, neutral-hyped, and under-hyped. These classifications are summ… view at source ↗

**Figure 2.** Normalized Sector Hype Index (Scaled by Overall Avg = 1) [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

**Figure 3.** Smoothed News-Weighted Ticker Hype Index for the Information Technology Sector without [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Figures from the paper (7 more)

**Figure 4.** Figure 4: Capitalization Adjusted Hype Index by Sector (Unscaled, New Clustering) [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

**Figure 5.** Figure 5: Smoothed Capitalization-Weighted Ticker Hype Index for Information Technology Tickers [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]

**Figure 6.** Figure 6: Histogram of Capitalization Adjusted Hype Index for Information Technology Sector [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]

**Figure 9.** Figure 9: Hyp, VIX, and GPR Index and Their Changes (7-Day Rolling Mean) [PITH_FULL_IMAGE:figures/full_fig_p014_9.png]

**Figure 10.** Figure 10: News Weight vs. Market Weight across all 11 sectors. Each dot represents a firm-day [PITH_FULL_IMAGE:figures/full_fig_p015_10.png]

**Figure 11.** Figure 11: Market Cap Fraction vs News Fraction by Sector Clusters with Overall Line of Best Fit [PITH_FULL_IMAGE:figures/full_fig_p016_11.png]

**Figure 22.** Figure 22: Capitalization Adjusted Hype Index vs 5-Day Rolling Log Return Std by Sector [PITH_FULL_IMAGE:figures/full_fig_p023_22.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

8 extracted references · 7 canonical work pages

[1]

Measuring Geopolitical Risk

Dario Caldara and Matteo Iacoviello. “Measuring Geopolitical Risk”. In:American Eco- nomic Review112.4 (2022), pp. 1194–1225.doi:10.1257/aer.20191823.url:https: //doi.org/10.1257/aer.20191823

work page doi:10.1257/aer.20191823.url:https: 2022
[2]

A Hype-Adjusted Probability Measure for NLP Stock Return Forecasting

Zheng Cao and Helyette Geman. “A Hype-Adjusted Probability Measure for NLP Stock Return Forecasting”. In:Frontiers in Artificial Intelligence8 (2025), p. 1527180.doi: 10.3389/frai.2025.1527180

work page arXiv 2025
[3]

Does the Stock Market Overreact?

Werner F. M. De Bondt and Richard Thaler. “Does the Stock Market Overreact?” In: The Journal of Finance40.3 (1985), pp. 793–805.doi:10.2307/2327804.url:https: //doi.org/10.2307/2327804

work page doi:10.2307/2327804.url:https: 1985
[4]

A sentiment analysis approach to the prediction of market volatil- ity

Justina Deveikyte et al. “A sentiment analysis approach to the prediction of market volatil- ity”. In:Front. Artif. Intell.5 (2022), p. 836809.doi:10.3389/frai.2022.836809

work page arXiv 2022
[5]

Does Unusual News Forecast Market Stress?

Paul Glasserman and Harry Mamaysky. “Does Unusual News Forecast Market Stress?” In:The Journal of Financial and Quantitative Analysis54.5 (2019). Accessed: 2025-05- 27, pp. 1937–1974.doi:10.1017/S0022109018001434.url:https://www.jstor.org/ stable/26782117

work page arXiv 2019
[6]

VADER: A Parsimonious Rule-based Model for Senti- ment Analysis of Social Media Text

Clayton J. Hutto and Eric Gilbert. “VADER: A Parsimonious Rule-based Model for Senti- ment Analysis of Social Media Text”. In:Proceedings of the International AAAI Conference on Web and Social Media8.1 (2014), pp. 216–225.url:https://ojs.aaai.org/index. php/ICWSM/article/view/14550

work page 2014
[7]

Martin.Speech and Language Processing: An Introduction to Natural Language Processing, Computational Linguistics, and Speech Recognition

Daniel Jurafsky and James H. Martin.Speech and Language Processing: An Introduction to Natural Language Processing, Computational Linguistics, and Speech Recognition. 2nd. Prentice Hall, 2000

work page 2000
[8]

Giving Content to Investor Sentiment: The Role of Media in the Stock Market

Paul C. Tetlock. “Giving Content to Investor Sentiment: The Role of Media in the Stock Market”. In:The Journal of Finance62.3 (2007), pp. 1139–1168.doi:10.1111/j.1540- 6261.2007.01232.x. [9]Volatility Index Methodology: Cboe Volatility Index (VIX). Tech. rep. Available at:https: //cdn.cboe.com/api/global/us_indices/governance/Volatility_Index_Methodology_...

arXiv 2007

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Measuring Geopolitical Risk

Dario Caldara and Matteo Iacoviello. “Measuring Geopolitical Risk”. In:American Eco- nomic Review112.4 (2022), pp. 1194–1225.doi:10.1257/aer.20191823.url:https: //doi.org/10.1257/aer.20191823

work page doi:10.1257/aer.20191823.url:https: 2022

[2] [2]

A Hype-Adjusted Probability Measure for NLP Stock Return Forecasting

Zheng Cao and Helyette Geman. “A Hype-Adjusted Probability Measure for NLP Stock Return Forecasting”. In:Frontiers in Artificial Intelligence8 (2025), p. 1527180.doi: 10.3389/frai.2025.1527180

work page arXiv 2025

[3] [3]

Does the Stock Market Overreact?

Werner F. M. De Bondt and Richard Thaler. “Does the Stock Market Overreact?” In: The Journal of Finance40.3 (1985), pp. 793–805.doi:10.2307/2327804.url:https: //doi.org/10.2307/2327804

work page doi:10.2307/2327804.url:https: 1985

[4] [4]

A sentiment analysis approach to the prediction of market volatil- ity

Justina Deveikyte et al. “A sentiment analysis approach to the prediction of market volatil- ity”. In:Front. Artif. Intell.5 (2022), p. 836809.doi:10.3389/frai.2022.836809

work page arXiv 2022

[5] [5]

Does Unusual News Forecast Market Stress?

Paul Glasserman and Harry Mamaysky. “Does Unusual News Forecast Market Stress?” In:The Journal of Financial and Quantitative Analysis54.5 (2019). Accessed: 2025-05- 27, pp. 1937–1974.doi:10.1017/S0022109018001434.url:https://www.jstor.org/ stable/26782117

work page arXiv 2019

[6] [6]

VADER: A Parsimonious Rule-based Model for Senti- ment Analysis of Social Media Text

Clayton J. Hutto and Eric Gilbert. “VADER: A Parsimonious Rule-based Model for Senti- ment Analysis of Social Media Text”. In:Proceedings of the International AAAI Conference on Web and Social Media8.1 (2014), pp. 216–225.url:https://ojs.aaai.org/index. php/ICWSM/article/view/14550

work page 2014

[7] [7]

Martin.Speech and Language Processing: An Introduction to Natural Language Processing, Computational Linguistics, and Speech Recognition

Daniel Jurafsky and James H. Martin.Speech and Language Processing: An Introduction to Natural Language Processing, Computational Linguistics, and Speech Recognition. 2nd. Prentice Hall, 2000

work page 2000

[8] [8]

Giving Content to Investor Sentiment: The Role of Media in the Stock Market

Paul C. Tetlock. “Giving Content to Investor Sentiment: The Role of Media in the Stock Market”. In:The Journal of Finance62.3 (2007), pp. 1139–1168.doi:10.1111/j.1540- 6261.2007.01232.x. [9]Volatility Index Methodology: Cboe Volatility Index (VIX). Tech. rep. Available at:https: //cdn.cboe.com/api/global/us_indices/governance/Volatility_Index_Methodology_...

arXiv 2007