The reviewed record of science sign in
Pith

arxiv: 2607.06220 · v1 · pith:M6HCBZ5O · submitted 2026-07-07 · physics.soc-ph · cs.CY· physics.data-an· q-fin.ST· stat.AP

Stable Sentiment and Persistent Dynamics in U.S. Economic News over 45 Years

Luis Enrique Correa Rocha This is my paper

Reviewed by Pith T0 review T1 audit T2 compute T3 formal T4 kernel 2026-07-08 12:46 UTCglm-5.2pith:M6HCBZ5Orecord.jsonopen to challenge →

classification physics.soc-ph cs.CYphysics.data-anq-fin.STstat.AP
keywords sentimentnegativenewseconomiclongerpersistentpositivebursts
0
0 comments X

The pith

Economic news sentiment gets stickier over 45 years

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

This paper analyzes a daily index of U.S. economic news sentiment spanning 1980 to 2025, drawn from 24 newspapers, and argues that while the average balance of positive and negative coverage has remained broadly stable, the persistence of sentiment states has increased substantially. Using detrended fluctuation analysis (DFA), the author estimates Hurst exponents in rolling windows and finds that short-scale persistence (7-81 days) follows a U-shaped trajectory: weakening in the pre-web period (1980-1995), flat during the early web years (1996-2008), and rising significantly after 2009. Complementary rolling statistics show declining volatility, fewer sign reversals, and increasing bimodality, meaning sentiment spends more time in clearly positive or negative regimes and less time fluctuating around neutral. The author also documents an asymmetry in burst dynamics: negative sentiment bursts last longer and have heavier tails than positive ones. To explain these patterns, the paper proposes a minimal endogenous-memory model in which sentiment is the sum of a slow latent component (fractional Gaussian noise with period-specific memory) and a fast shock component (AR(1) with period-specific feedback). Calibration shows that strengthening the slow memory parameter while weakening short-range corrective feedback reproduces the observed drift in Hurst exponents across media eras. The central claim is that U.S. economic news sentiment has shifted from a reactive process that corrected shocks quickly to a state-dependent process in which current tone conditions future tone over weekly-to-quarterly horizons.

Core claim

The paper's central discovery is that the temporal memory of U.S. economic news sentiment has lengthened over the past 45 years, quantified by rising Hurst exponents at short scales (7-81 days), particularly after 2009. This means sentiment shocks leave longer traces than expected under short-memory exponential decay. The shift is accompanied by declining volatility, fewer reversals, and increasing bimodality, indicating that sentiment increasingly organizes into sustained positive or negative episodes rather than quickly reverting to a baseline. A minimal two-component model (slow latent memory plus fast corrective feedback) reproduces the trend by strengthening endogenous memory and weakin

What carries the argument

The primary analytical tool is detrended fluctuation analysis (DFA), which estimates Hurst exponents H to quantify how fluctuations grow with observation scale and thus how strongly a time series depends on its past. The paper uses DFA of order 2 (DFA2) on the sentiment index and DFA of order 1 (DFA1) on first differences, with rolling 1,095-day windows to track temporal evolution. Four null models (i.i.d. shuffle, moving-block bootstrap, AR(1), and IAAFT) benchmark whether observed exponents exceed what short-range dependence or spectral structure alone would produce. The mechanistic model decomposes sentiment into a slow fractional Gaussian noise component with period-specific memory d_k (

If this is right

  • If sentiment persistence has genuinely increased, forecasting models in finance, macroeconomic nowcasting, and consumer-confidence monitoring that treat sentiment as a short-lived reaction to events will systematically underestimate the carryover of past shocks into current sentiment states.
  • The documented asymmetry (negative bursts lasting longer than positive ones) implies that pessimistic economic narratives may be more structurally self-sustaining than optimistic ones, which could bias downward the expected recovery speed from negative economic shocks.
  • If the shift toward state-dependent sentiment continues, policy communications designed to correct negative sentiment may need to account for longer relaxation times, as corrective information may take longer to dislodge entrenched sentiment regimes.
  • Models that use news sentiment as an input variable should incorporate a time-varying memory parameter rather than assuming a fixed autocorrelation structure across the full sample period.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The rise in persistence after 2009 coincides with the social-media period, but the paper's media-era stratification is descriptive rather than causal. A direct test would compare sentiment persistence across outlets with different degrees of platform integration or audience-feedback exposure, holding macroeconomic conditions constant.
  • If the measurement instrument (lexicon-based scoring) itself drifts in sensitivity over time, the observed persistence increase could partly reflect changes in how newspapers use economic language rather than changes in the underlying sentiment process. A cross-validation using a different scoring method (e.g., contemporary LLM-based scoring applied retroactively to the same corpus) on a subsample
  • The two-component model captures the trend but is not uniquely identified. Alternative generative mechanisms, such as time-varying exogenous shock regimes or changing topic composition of economic news, could produce similar scaling patterns and would need to be ruled out to strengthen the endogenous-memory interpretation.

Load-bearing premise

The load-bearing premise is that the lexicon-based sentiment index is temporally comparable across four decades despite acknowledged changes in article volume, topic composition, journalistic style, semantic drift, and time-varying sentiment intensity. If these composition effects systematically alter the autocorrelation structure of the index, the observed increase in persistence could reflect changes in the measurement instrument rather than a genuine shift in sentiment.

What would settle it

If the persistence trend disappears or reverses when a different sentiment-scoring method (e.g., contextual language model) is applied to the same newspaper corpus, or if controlling for changes in topic composition and article volume eliminates the post-2009 rise in Hurst exponents, the central claim would be undermined.

Figures

Figures reproduced from arXiv: 2607.06220 by Luis Enrique Correa Rocha.

Figure 1
Figure 1. Figure 1: Statistical characterization of the daily news sentiment index from 1980 to 2025. ( [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Estimated Hurst exponents. Log–log fluctuation functions for ( [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Empirical and modeled rolling Hurst exponents. Panels show the rolling ( [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Rolling organization and burst statistics of daily news sentiment. Panels show 1,095- [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗
read the original abstract

Collective emotion is often inferred from the tone of mass media, but such emotion is not directly observed. One approximation is to extract sentiment from text and use sentiment indexes as proxies to study the temporal organization of news sentiment. Using a daily index of U.S. economic news sentiment from 24 newspapers (1980-2025), we examine whether the response time of this sentiment process has changed. Although the average balance of positive and negative coverage has remained broadly stable, the persistence of news sentiment states has increased substantially. In dynamical terms, this implies longer residence times in optimistic or pessimistic regimes and weaker short-run correction of sentiment shocks. Complementary statistics show declining sentiment volatility, fewer reversals, and increasing bimodality, i.e. a stronger separation between positive and negative sentiment states. We also find an asymmetry between bursts of negative and positive sentiment, with negative bursts tending to last longer. These patterns are consistent with a minimal endogenous-memory model in which a slowly evolving latent sentiment component becomes more persistent while short-range corrective feedback weakens. The findings indicate a change in the temporal response of the U.S. economic newspaper sentiment index over the last 45 years, with sentiment shocks leaving longer traces than expected under short-memory exponential decay. News-based sentiment is thus better modeled as persistent episodes rather than as daily reactions that reset after each event.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 7 minor

Summary. This manuscript analyzes a daily U.S. economic news sentiment index (1980–2025, 24 newspapers) using detrended fluctuation analysis (DFA) to characterize the temporal memory of news sentiment. The central empirical finding is that short-scale persistence (7–81 day Hurst exponents) has increased over the past 45 years, particularly after 2009, while long-scale organization remains stable. Complementary rolling statistics (volatility, zero-crossing rate, bimodality coefficient) and burst analysis support the interpretation of a shift from reactive to state-dependent sentiment dynamics. A minimal endogenous-memory model with period-specific memory and feedback parameters is calibrated to reproduce the observed rolling Hurst exponent trends. The DFA methodology is applied carefully, with four null models (shuffle, block, AR(1), IAAFT) and multiple-testing correction.

Significance. The paper addresses a well-posed empirical question—whether the temporal persistence of news sentiment has changed over four decades—using a well-established scaling-analysis toolkit. The two-regime DFA framework with crossover detection is well-motivated, and the comparison against four null models (including IAAFT) is methodologically sound. The finding that short-scale persistence rises while long-scale organization remains stable is a specific, falsifiable empirical claim. The rolling Hurst exponent analysis with Newey-West standard errors is appropriate. The endogenous-memory model is transparently presented as descriptive rather than uniquely identified. The paper's scope is appropriate for the journal's readership in computational social science and complex systems.

major comments (1)
  1. The most important concern is the confound between declining measurement noise (from increasing article volume over time) and the reported rise in short-scale persistence. The daily sentiment index is estimated via day fixed effects from a varying number of articles per day (Eq. 1). If article volume increased substantially over 45 years, the day effects are estimated more precisely in later periods, mechanically increasing autocorrelation and the Hurst exponent even if the underlying sentiment process is unchanged. The paper acknowledges 'changes in article volume' as an uncontrolled factor (p. 3, p. 12) but does not test whether volume trends drive the H-trend. This is load-bearing for the central claim because all reported patterns—declining volatility (0.020→0.014), fewer zero-crossings (0.47→0.35), rising short-scale H, and the improved Michigan correlation (0.59→0.74)—are equally一致
minor comments (7)
  1. The bimodality coefficient BC = (ν² + 1)/κ is used without citing its source or discussing its known sensitivity to sample size. A reference (e.g., Pfister et al., 2013) and a note on the sample sizes involved would help readers.
  2. In the burst analysis (§ on Heavy-tailed dynamics), the power-law exponents γ ≈ 2.5–4.0 are reported without specifying the estimation method (MLE, Clauset et al., or Hill estimator) or the range over which the power law holds. Please specify the fitting procedure.
  3. The mechanistic model (Eq. 2–3) includes ω (state-dependent volatility) in the axiom ledger but ω is not reported among the calibrated parameters in the main text. Is ω fixed or fitted? If fixed, at what value? If fitted, what is the estimate?
  4. Fig. 2A: the raw index DFA curve yields H ≈ 0 at large scales, described as 'unrealistically low.' It would help to show (in SI) the fluctuation curves before and after each preprocessing step (deseasonalization, Hampel filtering) so readers can see what artifact is removed by each step.
  5. The paper references 'see SI' multiple times (crossover selection, sensitivity checks, long-scale trends, burst duration fits) but no SI appears to be attached. Please ensure SI is included with the submission.
  6. Table 1: the notation mixes α and H. The caption says 'Hurst exponents' but the text sometimes refers to α. Using one symbol consistently (or stating the mapping explicitly in the caption) would improve readability.
  7. The term 'mesoscopic' in the section title 'A mesoscopic endogenous-memory model' is not standard in this context. Consider 'minimal' or 'two-component' instead.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for a careful and constructive reading of the manuscript. The referee's central concern about the confound between declining measurement noise and rising persistence is well-taken and important. We address it below and commit to a revision that directly tests the robustness of our findings to article-volume trends.

read point-by-point responses

  1. Referee: The most important concern is the confound between declining measurement noise (from increasing article volume over time) and the reported rise in short-scale persistence. The daily sentiment index is estimated via day fixed effects from a varying number of articles per day (Eq. 1). If article volume increased substantially over 45 years, the day effects are estimated more precisely in later periods, mechanically increasing autocorrelation and the Hurst exponent even if the underlying sentiment process is unchanged. The paper acknowledges 'changes in article volume' as an uncontrolled factor but does not test whether volume trends drive the H-trend. This is load-bearing for the central claim because all reported patterns—declining volatility, fewer zero-crossings, rising short-scale H, and the improved Michigan correlation—are equally consistent with declining measurement noise.

    Authors: The referee raises a valid and important concern. We acknowledge that the manuscript identifies article-volume changes as an uncontrolled factor but does not directly test whether the observed trends in persistence, volatility, zero-crossing rate, and bimodality are driven by declining measurement noise rather than genuine changes in the sentiment process. This is a genuine gap in the analysis, and we agree it is load-bearing for the central claim. We will address it in the revision as follows. First, we will obtain the daily article counts underlying the sentiment index and report the volume trend over time. Second, we will perform a subsampling robustness check: for each rolling window, we will re-estimate the DFA exponents using only a fixed number of articles per day (matching the lower volume of earlier periods), so that measurement precision is approximately equalized across eras. If the rising-H trend survives this test, it cannot be attributed solely to declining noise. Third, as a complementary check, we will add calibrated measurement noise to later-period estimates to match the precision of earlier periods and re-estimate the rolling Hurst exponents. We note that while the volume-noise confound is a plausible alternative explanation for some of our findings—particularly declining volatility and rising short-scale H—certain patterns are less straightforwardly explained by declining noise alone. Specifically, the increasing bimodality coefficient (stronger separation between positive and negative states) is not an obvious consequence of reduced measurement noise, which would typically concentrate the distribution around its mean rather than increase separation between modes. Similarly, the asymmetry between positive and negative burst durations (negative bursts revision: no

Circularity Check

1 steps flagged

Model calibration fits to empirical Hurst exponents and reproduces them; paper is transparent. Central empirical claim is independent.

specific steps
  1. fitted input called prediction [Methods — Mechanistic model calibration; Figure 3; main text around Eq. 2-3]
    "We select the parameter set that best reproduces the empirical rolling Hurst exponents of both the sentiment index and its first differences by minimizing a global dimensionless loss function using L = |bias_x|/σ_x + |bias_Δx|/σ_Δx + (1 − corr_x) + (1 − corr_Δx)... The calibrated values show that the model can reproduce the observed scaling patterns, but they should not be seen as unique estimates of the true underlying process."

    The model's six parameters (d1,d2,d3,φ1,φ2,φ3) are calibrated by minimizing a loss function that directly measures deviation from the empirical rolling Hurst exponents. Figure 3 then shows the model 'reproduces' those same exponents. This is a fit-to-target: the model output (rolling H) is the quantity the parameters were optimized to match. However, the paper is explicitly transparent: it calls the model 'descriptive,' says parameters 'should not be seen as unique estimates,' and states 'The model is simple and not uniquely identified.' The model is not used to generate out-of-sample predictions or to claim the empirical trends are predicted. The central empirical claim (rising Hurst exponents) is measured directly from data via DFA and does not depend on the model. Thus this is a minor,-

full rationale

The paper's central empirical claim — that short-scale Hurst exponents of U.S. economic news sentiment have increased, particularly after 2009 — is derived directly from DFA applied to an externally constructed sentiment index (Shapiro et al.). This measurement does not depend on the author's model or prior work. The endogenous-memory model (Eqs. 2-3) is calibrated to reproduce the empirical rolling Hurst exponents and then shown to reproduce them, which is a fit-to-target rather than a genuine prediction. However, the paper is transparent about this: it explicitly states the model is 'descriptive,' 'not uniquely identified,' and that parameters 'should not be seen as unique estimates.' The model serves as a mechanistic interpretation, not as evidence for the empirical claim. Self-citations (refs [16], [30]) are used for general background on opinion dynamics and are not load-bearing for the central results. No step in the empirical derivation chain reduces to its own inputs by construction. The circularity is confined to the model-fitting exercise, which the paper honestly frames as descriptive rather than predictive.

Axiom & Free-Parameter Ledger

7 free parameters · 4 axioms · 1 invented entities

The model introduces 6-7 free parameters fitted to reproduce the empirical Hurst exponents. The decomposition into slow and fast components is an ad hoc modeling choice. The primary axiom, temporal comparability of the sentiment index, is a domain assumption that the paper itself flags as not fully satisfied.

free parameters (7)
  • d1 (pre-web memory parameter) = 0.28
    Memory parameter for the fractional Gaussian noise slow component in the pre-web period; fitted to match empirical rolling Hurst exponents.
  • d2 (web memory parameter) = 0.36
    Memory parameter for the web period; fitted to match empirical rolling Hurst exponents.
  • d3 (social media memory parameter) = 0.45
    Memory parameter for the social media period; fitted to match empirical rolling Hurst exponents.
  • phi1 (pre-web feedback coefficient) = -0.03
    AR(1) feedback coefficient for the fast shock component in the pre-web period; fitted to match empirical rolling Hurst exponents of first differences.
  • phi2 (web feedback coefficient) = 0.04
    AR(1) feedback coefficient for the web period; fitted to match empirical rolling Hurst exponents of first differences.
  • phi3 (social media feedback coefficient) = 0.26
    AR(1) feedback coefficient for the social media period; fitted to match empirical rolling Hurst exponents of first differences.
  • omega (state-dependent volatility)
    Controls the dependence of shock amplitude on previous state in the fast component; stated as 'weak (omega > 0)' but exact fitted value not reported in main text.
axioms (4)
  • domain assumption The Shapiro et al. sentiment index validly measures economic news sentiment and is comparable across 45 years despite changes in media infrastructure.
    The entire analysis rests on this index being a meaningful and comparable measure. The paper acknowledges this is not fully guaranteed (section 'News data and sentiment index').
  • standard math DFA is an appropriate method for quantifying long-range dependence in this sentiment time series.
    DFA is a standard method for this purpose, widely used in the physics and physiology literature.
  • domain assumption The three media-era periodization (pre-web, web, social media) is a meaningful descriptive framework.
    The paper explicitly states this should be seen as 'descriptive periodization rather than as causal identification' (Conclusions).
  • ad hoc to paper The sentiment process can be decomposed into a slow latent component and a fast shock component.
    This decomposition (Eq. 2) is the basis of the mechanistic model. It is a modeling choice, not derived from first principles.
invented entities (1)
  • Slow latent sentiment component xL_t no independent evidence
    purpose: Represents the slowly evolving background sentiment or macro-narrative in the mechanistic model.
    This is a model construct. It is not directly observed; only the aggregate sentiment index is observed. The paper does not provide an external falsifiable handle on this component.

pith-pipeline@v1.1.0-glm · 18099 in / 2601 out tokens · 348194 ms · 2026-07-08T12:46:20.327727+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

50 extracted references · 50 canonical work pages · 6 internal anchors

  1. [1]

    Amit Goldenberg and James J. Gross. Digital emotion contagion.Trends Cogn. Sci., 24(4): 316–328, 2020. doi: 10.1016/j.tics.2020.01.009

    work page doi:10.1016/j.tics.2020.01.009 2020

  2. [2]

    Robert J. Shiller. Narrative economics.Am. Econ. Rev., 107(4):967–1004, 2017. doi: 10.1257/ aer.107.4.967

    work page 2017

  3. [3]

    Nimark and Stefan Pitschner

    Kristoffer P. Nimark and Stefan Pitschner. News and noise: Narrative information and bond risk premia.J. Econ. Theory, 181:160–196, 2019. doi: 10.1016/j.jet.2019.02.001

    work page doi:10.1016/j.jet.2019.02.001 2019

  4. [4]

    Heston and Nitish R

    Steven L. Heston and Nitish R. Sinha. News vs. sentiment: Predicting stock returns from news stories.Financ. Anal. J., 73(3):67–83, 2017. doi: 10.2469/faj.v73.n3.4

    work page doi:10.2469/faj.v73.n3.4 2017

  5. [5]

    Feida Mai, Zhan Shan, Qing Bai, Xin Wang, and Roger H. L. Chiang. How does social media impact Bitcoin value? a test of the silent majority hypothesis.J. Manag. Inf. Syst., 35(1): 19–52, 2018. doi: 10.1080/07421222.2018.1440774

    work page doi:10.1080/07421222.2018.1440774 2018

  6. [6]

    McCombs and Donald L

    Maxwell E. McCombs and Donald L. Shaw. The agenda-setting function of mass media.Public Opin. Q., 36(2):176–187, 1972. doi: 10.1086/267990

    work page doi:10.1086/267990 1972

  7. [7]

    Competitive authoritarianism and media freedom: The dynamics of mass media in serbia.Probl

    Nebojˇ sa Vladisavljevi´ c. Competitive authoritarianism and media freedom: The dynamics of mass media in serbia.Probl. Post-Communism, 62(3):145–158, 2015. doi: 10.1080/10758216. 2015.1010906

    work page doi:10.1080/10758216 2015

  8. [8]

    Institutional trust and media use in times of cultural backlash: A cross- national study in nine european countries.Int

    Marc Verboord. Institutional trust and media use in times of cultural backlash: A cross- national study in nine european countries.Int. J. Press/Politics, 30(3):752–774, 2025. doi: 10.1177/19401612231187568

    work page doi:10.1177/19401612231187568 2025

  9. [9]

    Baker, Nicholas Bloom, and Steven J

    Scott R. Baker, Nicholas Bloom, and Steven J. Davis. Measuring economic policy uncertainty. Q. J. Econ., 131(4):1593–1636, 2016. doi: 10.1093/qje/qjw024

    work page doi:10.1093/qje/qjw024 2016

  10. [10]

    Adam Hale Shapiro, Moritz Sudhof, and Daniel J. Wilson. Measuring news sentiment. Working Paper 2017-01, Federal Reserve Bank of San Francisco, 2020

    work page 2017

  11. [11]

    Buldyrev, Shlomo Havlin, Fredrik Liljeros, and Hern´ an A

    Diego Rybski, Sergey V. Buldyrev, Shlomo Havlin, Fredrik Liljeros, and Hern´ an A. Makse. Communication activity in a social network: Relationship between long-term correlations and inter-event clustering.Sci. Rep., 2:560, 2012. doi: 10.1038/srep00560. 15

    work page doi:10.1038/srep00560 2012

  12. [12]

    Eugene Stanley

    Yukie Sano, Yuki Sano, Hideki Takayasu, Misako Takayasu, Shlomo Havlin, and H. Eugene Stanley. Long-term periodic cycles and memories of collective emotion in online social media. PLOS ONE, 14(3):e0213843, 2019. doi: 10.1371/journal.pone.0213843

    work page doi:10.1371/journal.pone.0213843 2019

  13. [13]

    Collective emotions and social resilience in the digital traces after a terrorist attack.Psychol

    David Garc´ ıa and Bernard Rim´ e. Collective emotions and social resilience in the digital traces after a terrorist attack.Psychol. Sci., 30(4):617–628, 2019. doi: 10.1177/0956797619831964

    work page doi:10.1177/0956797619831964 2019

  14. [14]

    Alexandre Bovet and Hern´ an A. Makse. Influence of fake news in twitter during the 2016 US presidential election.Nat. Commun., 10:7, 2019. doi: 10.1038/s41467-018-07761-2

    work page doi:10.1038/s41467-018-07761-2 2016

  15. [15]

    Springer Briefs in Complexity

    M´ arton Karsai, Nicola Perra, and Alessandro Vespignani.Bursty Human Dynamics. Springer Briefs in Complexity. Springer, Cham, 2018. doi: 10.1007/978-3-319-68540-3

    work page doi:10.1007/978-3-319-68540-3 2018

  16. [16]

    Fatemeh Zarei, Yerali Gandica, and Luis E. C. Rocha. Bursts of communication increase opinion diversity in the temporal deffuant model.Sci. Rep., 14(1):2222, 2024. doi: 10.1038/ s41598-024-52458-w

    work page 2024

  17. [17]

    Oxford University Press, New York, 2018

    Jos´ e van Dijck, Thomas Poell, and Martijn de Waal.The Platform Society: Public Values in a Connective World. Oxford University Press, New York, 2018. ISBN 9780190889777

    work page 2018

  18. [18]

    Platformisation.Internet Policy Rev., 8(4): 1–13, 2019

    Thomas Poell, David Nieborg, and Jos´ e van Dijck. Platformisation.Internet Policy Rev., 8(4): 1–13, 2019. doi: 10.14763/2019.4.1425

    work page doi:10.14763/2019.4.1425 2019

  19. [19]

    Twitter mood predicts the stock market.J

    Johan Bollen, Huina Mao, and Xiaojun Zeng. Twitter mood predicts the stock market.J. Comput. Sci., 2(1):1–8, 2011. doi: 10.1016/j.jocs.2010.12.007

    work page doi:10.1016/j.jocs.2010.12.007 2011

  20. [20]

    Eugene Stanley

    Tobias Preis, Helen Susannah Moat, and H. Eugene Stanley. Quantifying trading behavior in financial markets using Google trends.Sci. Rep., 3:1684, 2013. doi: 10.1038/srep01684

    work page doi:10.1038/srep01684 2013

  21. [21]

    The impact of microblogging sentiment on stock returns: Evidence from the Portuguese stock market.Decis

    Nuno Oliveira, Paulo Cortez, and Nelson Areal. The impact of microblogging sentiment on stock returns: Evidence from the Portuguese stock market.Decis. Support Syst., 85:62–74,

    work page

  22. [22]

    doi: 10.1016/j.dss.2016.12.004

    work page doi:10.1016/j.dss.2016.12.004 2016

  23. [23]

    Soroka, Dominik Stecula, and Christopher Wlezien

    Stuart N. Soroka, Dominik Stecula, and Christopher Wlezien. It’s (change in) the (future) economy, stupid: Economic indicators, the media, and public opinion.Am. J. Political Sci., 59 (2):457–474, 2015. doi: 10.1111/ajps.12145

    work page doi:10.1111/ajps.12145 2015

  24. [24]

    Westwood

    Shanto Iyengar and Sean J. Westwood. Fear and loathing across party lines: New evidence on group polarization.Am. J. Political Sci., 59(3):690–707, 2015. doi: 10.1111/ajps.12152

    work page doi:10.1111/ajps.12152 2015

  25. [25]

    Joseph Engelberg and Christopher A. Parsons. The causal impact of media in financial markets. J. Finance, 66(1):67–97, 2011. doi: 10.1111/j.1540-6261.2010.01626.x

    work page doi:10.1111/j.1540-6261.2010.01626.x 2011

  26. [26]

    Eytan Bakshy, Solomon Messing, and Lada A. Adamic. Exposure to ideologically diverse news and opinion on Facebook.Science, 348(6239):1130–1132, 2015. doi: 10.1126/science.aaa1160

    work page doi:10.1126/science.aaa1160 2015

  27. [27]

    Seth Flaxman, Sharad Goel, and Justin M. Rao. Filter bubbles, echo chambers, and online news consumption.Public Opin. Q., 80(S1):298–320, 2016. doi: 10.1093/poq/nfw006

    work page doi:10.1093/poq/nfw006 2016

  28. [28]

    The spread of true and false news online.Science, 359(6380):1146–1151, 2018

    Soroush Vosoughi, Deb Roy, and Sinan Aral. The spread of true and false news online.Science, 359(6380):1146–1151, 2018. doi: 10.1126/science.aap9559. 16

    work page doi:10.1126/science.aap9559 2018

  29. [29]

    Bail, Lisa P

    Christopher A. Bail, Lisa P. Argyle, Taylor W. Brown, John P. Bumpus, Haohan Chen, M. B. Hunzaker, Jaemin Lee, Marcus Mann, Friedolin Merhout, and Alexander Volfovsky. Exposure to opposing views on social media can increase political polarization.Proc. Natl. Acad. Sci. U.S.A., 115(37):9216–9221, 2018. doi: 10.1073/pnas.1804840115

    work page doi:10.1073/pnas.1804840115 2018

  30. [30]

    Guess, Brendan Nyhan, and Jason Reifler

    Andrew M. Guess, Brendan Nyhan, and Jason Reifler. Selective exposure to misinformation: Evidence from the consumption of fake news during the 2016 U.S. presidential campaign. European Research Council Working Paper, 1, 2018

    work page 2016

  31. [31]

    Nicola Perra and Luis E. C. Rocha. Modelling opinion dynamics in the age of algorithmic personalisation.Sci. Rep., 9(1):7261, 2019. doi: 10.1038/s41598-019-43830-2

    work page doi:10.1038/s41598-019-43830-2 2019

  32. [32]

    Rewarding Engagement and Personalization in Popularity-Based Rankings Amplifies Extremism and Polarization

    Jacopo D’Ignazi, Andreas Kaltenbrunner, Ga¨ el Le Mens, Fabrizio Germano, and Vicen¸ c G´ omez. Rewarding engagement and personalization in popularity-based rankings amplifies extremism and polarization.arXiv preprint, 2025. doi: 10.48550/arXiv.2510.24354

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2510.24354 2025

  33. [33]

    Peterson-Salahuddin and Nicholas Diakopoulos

    Brita Y. Peterson-Salahuddin and Nicholas Diakopoulos. Negotiated autonomy: Journalistic role performance, audience metrics, and social media algorithms.Digit. Journal., 8(6):772–789,

    work page

  34. [34]

    doi: 10.1080/21670811.2020.1733761

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1080/21670811.2020.1733761 2020

  35. [35]

    Guide to automated journalism

    Andreas Graefe. Guide to automated journalism. Technical report, Tow Center for Digital Journalism, Columbia University, 2016

    work page 2016

  36. [36]

    Predictions of Quasar Clustering: Redshift, Luminosity and Selection Dependence

    Matt Carlson. The robotic reporter: Automated journalism and the redefinition of labor, compositional forms, and journalistic authority.Digit. Journal., 3(3):416–431, 2015. doi: 10.1080/21670811.2014.976412

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1080/21670811.2014.976412 2015

  37. [37]

    Land Cover Mapping Using Ensemble Feature Selection Methods

    Tal Montal and Zvi Reich. I, robot. you, journalist. who is the author? authorship, bylines and full disclosure in automated journalism.Digit. Journal., 5(7):829–849, 2017. doi: 10.1080/ 21670811.2016.1209083

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  38. [38]

    Form Factors and Dyson-Schwinger Equations

    Nicole Blanchett Neheli. News by numbers: The evolution of analytics in journalism.Digit. Journal., 6(8):1041–1051, 2018. doi: 10.1080/21670811.2018.1504626

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1080/21670811.2018.1504626 2018

  39. [39]

    Do metrics drive news decisions? political news journalists’ exposure, evaluation, and use of audience data.Journal

    Jeroen Lamot, Steve Paulussen, and Peter Van Aelst. Do metrics drive news decisions? political news journalists’ exposure, evaluation, and use of audience data.Journal. Stud., 22(15): 2063–2080, 2021. doi: 10.1080/1461670X.2021.1916986

    work page doi:10.1080/1461670x.2021.1916986 2063

  40. [40]

    Peng, Shlomo Havlin, H

    C.-K. Peng, Shlomo Havlin, H. Eugene Stanley, and Ary L. Goldberger. Quantification of scaling exponents and crossover phenomena in nonstationary heartbeat time series.Chaos, 5 (1):82–87, 1995. doi: 10.1063/1.166141

    work page doi:10.1063/1.166141 1995

  41. [41]

    Kantelhardt, Eva Koscielny-Bunde, Henio H

    Jan W. Kantelhardt, Eva Koscielny-Bunde, Henio H. A. Rego, Shlomo Havlin, and Armin Bunde. Detecting long-range correlations with detrended fluctuation analysis.Phys. A, 295 (3–4):441–454, 2001. doi: 10.1016/S0378-4371(01)00144-3

    work page doi:10.1016/s0378-4371(01)00144-3 2001

  42. [42]

    Daily news sentiment index

    Federal Reserve Bank of San Francisco. Daily news sentiment index. www.frbsf. org/research-and-insights/data-and-indicators/daily-news-sentiment-index/ , Ac- cessed: 22 October 2025

    work page 2025

  43. [43]

    Four best practices for measuring news 17 sentiment using ’off-the-shelf’ dictionaries: A large-scale p-hacking experiment.Computational Communication Research, 3(1):1–27, 2021

    Chung-hong Chan, Joseph Bajjalieh, Loretta Auvil, Hartmut Wessler, Scott Althaus, Kasper Welbers, Wouter van Atteveldt, and Marc Jungblut. Four best practices for measuring news 17 sentiment using ’off-the-shelf’ dictionaries: A large-scale p-hacking experiment.Computational Communication Research, 3(1):1–27, 2021. doi: 10.5117/CCR2021.1.001.CHAN

    work page doi:10.5117/ccr2021.1.001.chan 2021

  44. [44]

    Hutto and Eric Gilbert

    Clayton J. Hutto and Eric Gilbert. VADER: A parsimonious rule-based model for sentiment analysis of social media text. InProceedings of the Eighth International AAAI Conference on Weblogs and Social Media, 2014. doi: 10.1609/icwsm.v8i1.14550

    work page doi:10.1609/icwsm.v8i1.14550 2014

  45. [45]

    Deep Just-In-Time Inconsistency Detection Between Comments and Source Code

    Tim Loughran and Bill McDonald. When is a liability not a liability? textual analysis, dictionaries, and 10-Ks.The Journal of Finance, 66(1):35–65, 2011. doi: 10.1111/j.1540-6261. 2010.01625.x

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1111/j.1540-6261 2011

  46. [46]

    Mining and summarizing customer reviews

    Minqing Hu and Bing Liu. Mining and summarizing customer reviews. InProceedings of the Tenth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 168–177, 2004. doi: 10.1145/1014052.101407

    work page doi:10.1145/1014052.101407 2004

  47. [47]

    Social influence and unfollowing accelerate the emergence of echo chambers.J

    Kazutoshi Sasahara, Wen Chen, Hao Peng, Giovanni Luca Ciampaglia, Alessandro Flammini, and Filippo Menczer. Social influence and unfollowing accelerate the emergence of echo chambers.J. Comput. Soc. Sci., 3(1):381–402, 2020. doi: 10.1007/s42001-020-00084-7

    work page doi:10.1007/s42001-020-00084-7 2020

  48. [48]

    Universality, criticality and complexity of information propagation in social media

    Daniele Notarmuzi, Claudio Castellano, Alessandro Flammini, Dario Mazzilli, and Filippo Radicchi. Universality, criticality and complexity of information propagation in social media. Nat. Commun., 13:7558, 2022. doi: 10.1038/s41467-022-35295-4

    work page doi:10.1038/s41467-022-35295-4 2022

  49. [49]

    Detrended fluctuation analysis as a regression framework: Estimating dependence at different scales.Phys

    Ladislav Kristoufek. Detrended fluctuation analysis as a regression framework: Estimating dependence at different scales.Phys. Rev. E, 91(2):022802, 2015. doi: 10.1103/PhysRevE.91. 022802

    work page doi:10.1103/physreve.91 2015

  50. [50]

    Improved surrogate data for nonlinearity tests.Phys

    Thomas Schreiber and Andreas Schmitz. Improved surrogate data for nonlinearity tests.Phys. Rev. Lett., 77(4):635–638, 1996. doi: 10.1103/PhysRevLett.77.635. 18

    work page doi:10.1103/physrevlett.77.635 1996