Pith. sign in

REVIEW 3 major objections 42 references

Community-aware propagation of LLM event signals through dynamic financial knowledge graphs produces incremental return predictability beyond direct firm news.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

In controlled simulations, community-aware propagation of LLM event signals on dynamic financial knowledge graphs recovers latent communities and prices incrementally beyond direct signals, though live alpha remains untested.

T0 review reviewed 2026-07-14 challenge →

load-bearing objection Careful simulation machinery for community-gated LLM event propagation; ranking inherits a DGP that hard-codes the same gate, and there is still no live market test. the 3 major comments →

arxiv 2607.10932 v1 pith:OUXD5JKF submitted 2026-07-12 stat.AP

LLM-Enhanced Dynamic Financial Knowledge Graphs for Cross-Entity Signal Propagation and alpha discovery

Lin Zhang This is my paper
classification stat.AP
keywords large language modelsknowledge graphscommunity detectioninformation diffusionlead-lag effectsalpha discoveryfinancial networks
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

Financial news is almost never about one firm alone: a capital-spending surprise or supply constraint travels through suppliers, customers, competitors, and technology ecosystems. Standard financial NLP scores only the mentioned firm and throws away the rest. This paper reframes the LLM as a measurement engine that extracts economic state innovations and the relationships among firms, builds a dynamic knowledge graph, detects evolving economic communities, and then diffuses the event signals more strongly inside those communities than across them. The resulting Propagated Information Surprise can be large for firms that never appear in a document. In controlled simulations that include noisy extraction, missing edges, and a mid-sample ecosystem split, community-aware propagation recovers the latent structure, prices the cross-entity signal incrementally to the direct signal, and ranks first among five nested benchmarks in rank IC and long–short Sharpe. A Russell-1000-calibrated re-run preserves the ordering at realistic magnitudes and quantifies the transaction-cost hurdle a live implementation must clear.

Core claim

When LLM-extracted economic state innovations are propagated through a dynamically detected financial knowledge graph with stronger within-community than cross-community weights, the resulting firm-level propagated signal predicts returns incrementally to the direct event signal and outperforms sentiment, direct events, static-graph, and uniform dynamic-graph benchmarks in controlled simulations with time-varying communities and realistic measurement noise.

What carries the argument

Community-aware signal propagation: event signals diffuse with weights λ_in inside detected economic communities and λ_out across them (maintained λ_in > λ_out), producing Community Information Surprise (CIS) and firm-level Propagated Information Surprise (PIS).

Load-bearing premise

The simulations hard-code that news actually travels more strongly inside the same economic communities the detector recovers than across them; if real markets do not diffuse that way, the ranking need not hold.

What would settle it

A live, point-in-time Russell 1000 study that extracts events and relationships from earnings calls, filings, and newswire, builds the dynamic graph, and runs Fama–MacBeth tests of the propagated coefficient: if the community-aware propagated signal is not priced after the direct signal and standard controls, or if estimated λ_in is not reliably larger than λ_out, the central claim fails.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Alpha can arise from modelling how information travels across latent economic communities, not only from discovering news first.
  • Dynamic community detection on text-derived graphs can surface emerging ecosystems before static sector taxonomies update.
  • PIS supplies a usable signal for never-mentioned or sparsely covered firms, concentrating where diffusion is slowest.
  • Standalone 10-day quintile rebalancing at realistic large-cap magnitudes faces a material transaction-cost hurdle, pushing implementation toward longer horizons, overlays, or low-coverage names.
  • Graph-extraction noise systematically attenuates feasible propagation coefficients, so extraction quality is first-order for measured alpha.

Where Pith is reading between the lines

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

  • If live LLM extraction noise is structured by document style or entity salience rather than roughly i.i.d., the attenuation formulas and community recovery rates may understate bias and the ranking could reverse.
  • The same pipeline could be stress-tested on smaller-cap or emerging-market universes where coverage is thinner and diffusion horizons longer, potentially raising the cost-adjusted Sharpe.
  • Event-type-dependent attention (supply-constraint vs demand-acceleration) is a natural next operator once live data volume supports training a graph-attention layer.
  • Widespread adoption of community-gated propagation would itself shorten diffusion horizons and erode the premium, exactly as earlier customer-momentum effects have partially arbitraged away.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 0 minor

Summary. The paper proposes an LLM-based pipeline that extracts firm-level economic state innovations and typed relationships from unstructured documents, builds a dynamic financial knowledge graph, detects communities with Louvain, and propagates event signals with a community gate λ_in > λ_out. It defines Community Information Surprise (CIS) and Propagated Information Surprise (PIS), derives an attenuation result for graph-extraction noise (Proposition 1), and tests five nested signals (sentiment, direct events, static-graph, dynamic-graph, community-aware) in controlled simulations with mid-sample ecosystem splits and noisy extraction. In the stylized design (N=300) community-aware propagation ranks first on rank IC and long–short Sharpe and is priced incrementally (median Fama–MacBeth t ≈ 3.7); a Russell-1000-calibrated design (N=1,000, sparser graph, heterogeneous coverage, smaller effects) preserves the ordering at compressed magnitudes (propagated priced in 80% of replications) while quantifying a transaction-cost hurdle. The paper supplies a point-in-time blueprint for a live Russell 1000 study and scopes claims as machinery validation rather than live alpha.

Significance. If the machinery transfers, the contribution is a coherent, testable bridge from LLM event extraction to cross-entity asset pricing that goes beyond document-level sentiment: innovation measurement, dynamic text-derived graphs, community-gated propagation, and formal Fama–MacBeth/portfolio tests. Strengths include the nested design, point-in-time discipline and graph burn-in, the errors-in-variables attenuation theory that matches observed λ shortfalls, explicit cost hurdles, and a complete live-study blueprint (prompts, leakage controls, data sources). The work is of clear interest to empirical asset pricing and financial NLP; its main limitation is that the ranking and incremental pricing are demonstrated under a DGP that encodes the same community gate the estimator assumes, so transfer remains an open empirical question the live study must answer.

major comments (3)
  1. Section 5.1 and Tables 8–9 hard-code community-gated diffusion (cin/cout = 4:1 stylized at 12/3 bps; Russell 8/2 bps) inside a stochastic-block structure that matches the maintained hypothesis of operator (8). The five-method ranking (Tables 1, 3, 5, 7) and incremental pricing of Propagated (Tables 2, 6) therefore show recovery under a DGP aligned with the estimator. The paper correctly scopes this as machinery validation, but the central ordering claim still rests on that match. A falsifying or stress DGP with λ_in = λ_out (or reverse gate, or non-block diffusion) should be reported so readers can see when the community-aware advantage disappears; without it the transfer claim is under-supported.
  2. The measurement layer uses i.i.d. signal and weight noise (Section 5.1: η ~ N(0,0.5^{2}), weight N(0,0.15^{2}), 20% missing / 10% spurious edges). Section 7 notes that live LLM errors are likely structured (style, salience, narrative). Because Proposition 1 and the feasible λ recovery (Section 5.3; Figure 4) rely on this noise model, at least one structured-noise or correlated-extraction stress should be shown; otherwise the attenuation match and community-recovery NMI (0.86 / 0.76) may overstate robustness for the live pipeline.
  3. All evidence is simulation; no realized-market results are reported. Section 6 supplies a careful Russell 1000 blueprint and a calibrated Monte Carlo, but the abstract and introduction still frame alpha discovery and cross-entity predictive power as delivered findings. The manuscript should either (i) include a limited live pilot (even a short post-cutoff window with one LLM and public text) or (ii) systematically soften claim language so that every ranking/pricing statement is explicitly conditional on the DGP until live evidence exists.

Circularity Check

0 steps flagged

No significant circularity: simulation recovers a known DGP under noise; claims scoped as machinery validation, not tautological prediction.

full rationale

The paper's load-bearing claims are empirical recovery and ranking results inside a controlled Monte Carlo whose truth is known by construction of the DGP (Section 5.1; Tables 8–9). The feasible estimator never observes true innovations u, true weights W, or true memberships except in explicitly labeled oracle diagnostics; it sees only noisy si,t, a noisy extracted graph (20% missing edges, 10% spurious, weight noise), and Louvain communities on that graph. Community recovery (mean NMI ≈ 0.86), attenuation of λ̂ (Proposition 1, standard errors-in-variables applied to the product regressor), incremental Fama–MacBeth pricing of Propagated (Tables 2, 6), and the five-method ordering are therefore non-tautological demonstrations that the pipeline works under the stated measurement imperfections. The DGP does encode cin/cout = 4 matching the maintained hypothesis λin > λout of operator (8), but the paper repeatedly scopes the exercise as machinery validation rather than a claim of live alpha (Abstract; §1; §6–8) and already quantifies compression under Russell-1000 calibration. No equation reduces a claimed prediction to a fitted input by definition, no uniqueness theorem is imported from the author, and no self-citation is load-bearing for the central result. The derivation chain is therefore self-contained against its own simulation benchmarks.

Axiom & Free-Parameter Ledger

7 free parameters · 6 axioms · 3 invented entities

The central simulation claim rests on a fully specified synthetic economy plus standard network and asset-pricing tools. Free parameters dominate: edge densities, response sizes, noise variances, propagation (K,γ,λ), and coverage heterogeneity are chosen by hand to match stylized or large-cap facts. Domain axioms include gradual diffusion, modularity communities as the right gate, and LLM-as-innovation-measurement. Invented entities are the CIS/PIS factors and the community-aware operator as an identified linear object; they are definitions with simulation handles, not new physical entities. No machine-checked proofs; independence of live alpha is explicitly not claimed.

free parameters (7)
  • λ_in, λ_out (community gate)
    A priori set to (1.0, 0.3) for M5; also estimated. Robustness table varies λ_out in [0,1]. Directly shapes the preferred method’s signal.
  • Propagation depth K and decay γ
    Fixed at K=2, γ=0.35 throughout main results; varied in robustness. Controls multi-hop mass of ˜s.
  • cin, cout, cown (true per-edge and own responses)
    Stylized 12/3/50 bps; Russell 8/2/40 bps. Encode the structural λ_in>λ_out the method is built to recover.
  • pin, pout (SBM edge probabilities)
    0.18/0.012 stylized; 0.10/0.004 Russell. Determine community recoverability and graph sparsity.
  • Extraction noise (miss 20%, spurious 10%, weight σ=0.15, signal η σ=0.5)
    Hand-set measurement layer; drives attenuation in Proposition 1 and feasible ˆλ.
  • Event intensity and heterogeneous coverage U(0.015,0.075)
    Russell calibration only; controls power and PIS relevance for low-coverage names.
  • Portfolio rebalance (10 days), costs (5/10 bps), half-lives
    Implementation parameters that determine whether net Sharpe is positive; not estimated from data.
axioms (6)
  • domain assumption Information diffuses gradually across economic links at horizons of days to weeks (Hong–Stein style limited attention).
    Section 2.1; underpins Hypotheses 1–2 and the multi-day neighbor kernel in the DGP.
  • domain assumption Latent economic communities gate diffusion more strongly within than across (λ_in > λ_out).
    Maintained in operator (8) and hard-coded in DGP cin/cout; Hypothesis 2.
  • domain assumption Modularity maximization (Louvain) on the extracted weighted graph recovers economically meaningful communities for propagation.
    Section 3.3; known resolution-limit pathologies acknowledged in §7.
  • domain assumption LLM outputs can be treated as noisy measurements of state innovations ΔState and of typed relationship edges.
    Equations (4)–(6); live quality untested, mocked by i.i.d. noise in simulation.
  • standard math Standard modularity, SBM, Fama–MacBeth, Newey–West, and errors-in-variables attenuation math.
    Used throughout §§3–5 and Appendix A without novel proof obligations beyond Proposition 1.
  • ad hoc to paper Simulation DGP with mid-sample community split and specified noise is an adequate stress test of the feasible pipeline.
    Section 5 design choice; validates machinery not market premium (§7).
invented entities (3)
  • Community Information Surprise (CIS) no independent evidence
    purpose: Aggregate state innovations within a detected community as a community-level factor.
    Defined in (9); simulation-only evidence; no external market test.
  • Propagated Information Surprise (PIS) no independent evidence
    purpose: Project CIS back to firms via community exposures so never-mentioned firms can receive signal.
    Defined in (10); subsumed by edge-level propagation in horse race but retained for sparse coverage.
  • Community-aware propagation operator with gate φ ∈ {λ_in, λ_out} no independent evidence
    purpose: Linear, identifiable alternative to uniform graph diffusion / GAT for cross-entity signals.
    Equation (8); testable via feasible regression (12); independent evidence only inside the paper’s DGP.

reviewed 2026-07-14 · how reviews work

0 comments
Cite this review

Pith. "Pith review of LLM-Enhanced Dynamic Financial Knowledge Graphs for Cross-Entity Signal Propagation and alpha discovery." pith.science (2026). https://pith.science/paper/OUXD5JKF

@misc{pith2026260710932,
  author       = {Pith},
  title        = {Pith review of: LLM-Enhanced Dynamic Financial Knowledge Graphs for Cross-Entity Signal Propagation and alpha discovery},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OUXD5JKF}},
  note         = {Machine review of arXiv:2607.10932}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Financial information rarely affects a single company in isolation. Earnings surprises, capital expenditure changes, supply constraints, and guidance revisions can propagate through networks of suppliers, customers, competitors, and technology ecosystems. Traditional financial NLP primarily measures document-level sentiment for the directly mentioned company and often ignores cross-entity information diffusion. This paper develops an LLM-based financial measurement and signal propagation framework. The LLM converts unstructured financial documents into structured economic state-change events and extracts explicit and implicit corporate relationships to construct a dynamic financial knowledge graph. Event signals are then propagated through the estimated network using a community-aware mechanism, allowing information to diffuse more strongly within dynamically detected economic communities than across community boundaries. We introduce Community Information Surprise, CIS, and Propagated Information Surprise, PIS, as network-based financial signals and develop corresponding econometric tests. Controlled simulations with time-varying economic communities show that the framework accurately recovers latent network structure, detects the emergence of new investment ecosystems, and generates propagated signals with incremental predictive power beyond sentiment and direct LLM event signals. Across repeated simulations, community-aware propagation achieves the strongest rank information coefficient and long-short Sharpe ratio among five nested benchmarks.A second Russell 1000 calibrated simulation confirms that the main results persist under sparser networks, heterogeneous news coverage, realistic large-cap volatility, and smaller effect sizes.

Figures

Figures reproduced from arXiv: 2607.10932 by Lin Zhang.

Figure 1
Figure 1. Figure 1: Extracted knowledge graph and detected communities. Spring-layout snapshots of the observed (noisily extracted) graph Wct in a representative replication, before (left) and after (right) the true ecosystem split at t = 375. Node colours are the communities detected by weighted Louvain on Wct at that refresh date; edges drawn with opacity proportional to weight. 5.3 The diffusion mechanism: event studies an… view at source ↗
Figure 2
Figure 2. Figure 2: Dynamic community recovery. Top: normalized mutual information between Louvain communities detected on the extracted graph and the true membership, at each monthly refresh; mean across 20 replications with 95% band. Bottom: mean detected number of communities Kˆ t against the true Kt . The dashed vertical line marks the true ecosystem split at t = 375. 0 2 4 6 8 10 12 14 Days after source event 0 2 4 6 8 1… view at source ↗
Figure 3
Figure 3. Figure 3: Cross-entity diffusion after source events. Sign-adjusted, market-adjusted cumulative average returns of graph neighbours following source events with |ui,t| ≥ 0.75, split by true community co-membership; neighbours with own events inside the window are excluded. Mean across 20 replications; shaded bands are 95% confidence intervals from cross-replication dispersion. 18 [PITH_FULL_IMAGE:figures/full_fig_p… view at source ↗
Figure 4
Figure 4. Figure 4: Feasible estimates of the community gate. Pooled edge-level regression (12) of neighbours’ 10-day-forward market-adjusted returns on the weighted source signal, split by detected community co-membership. Bars are means across 20 replications; whiskers are 95% confidence intervals. True per-edge responses are cin = 12 and cout = 3 bps; the shortfall is the measurement-error attenuation of Proposition 1. fir… view at source ↗
Figure 5
Figure 5. Figure 5: IC decay. Mean rank IC of each method against forward returns over horizons h = 1–20 trading days; averages across 20 replications with 95% bands. rejected in only 20% of replications), but strongly significant on the dynamic graph (median t = 3.4; rejection rate 85%) and the community-aware graph (median t = 3.7; rejection rate 90%), while the direct coefficient is essentially unchanged. This is Hypothesi… view at source ↗
Figure 6
Figure 6. Figure 6: Cumulative long–short performance net of 5 bps costs. Average cumulative portfolio value across 20 replications; the shaded band is the interquartile range for the community￾aware method. to the propagation horizon, netting against other signals, and liquid universes). The Sharpe dispersion is itself a finding: three-year backtests of modest-IC signals are extremely noisy, and in live research several non-… view at source ↗
Figure 7
Figure 7. Figure 7: Cross-entity diffusion at Russell-1000 calibration. Sign-adjusted, market￾adjusted cumulative average returns of graph neighbours after source events, split by true community co-membership, in the Russell-1000-calibrated configuration (N = 1,000; per-edge responses 8/2 bps). Mean across 10 replications with 95% bands. (smaller than in the stylized design, reflecting a more efficient large-cap market); and—… view at source ↗

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

42 extracted references · 5 linked inside Pith

  1. [1]

    Journal of Finance , year =

    Cohen, Lauren and Frazzini, Andrea , title =. Journal of Finance , year =

  2. [2]

    Journal of Finance , year =

    Menzly, Lior and Ozbas, Oguzhan , title =. Journal of Finance , year =

  3. [3]

    Review of Financial Studies , year =

    Hou, Kewei , title =. Review of Financial Studies , year =

  4. [4]

    Journal of Financial Economics , year =

    Hong, Harrison and Torous, Walter and Valkanov, Rossen , title =. Journal of Financial Economics , year =

  5. [5]

    and MacBeth, James D

    Fama, Eugene F. and MacBeth, James D. , title =. Journal of Political Economy , year =

  6. [6]

    , title =

    Hong, Harrison and Stein, Jeremy C. , title =. Journal of Finance , year =

  7. [7]

    , title =

    Tetlock, Paul C. , title =. Journal of Finance , year =

  8. [8]

    Journal of Finance , year =

    Loughran, Tim and McDonald, Bill , title =. Journal of Finance , year =

  9. [9]

    Journal of Financial Economics , year =

    Lopez-Lira, Alejandro and Tang, Yuehua , title =. Journal of Financial Economics , year =

  10. [10]

    arXiv preprint arXiv:1908.10063 , year =

    Araci, Dogu , title =. arXiv preprint arXiv:1908.10063 , year =

    Pith/arXiv arXiv 1908

  11. [11]

    and Wang, Hui and Yang, Yi , title =

    Huang, Allen H. and Wang, Hui and Yang, Yi , title =. Contemporary Accounting Research , year =

  12. [12]

    arXiv preprint arXiv:2303.17564 , year =

    Wu, Shijie and Irsoy, Ozan and Lu, Steven and Dabravolski, Vadim and Dredze, Mark and Gehrmann, Sebastian and Kambadur, Prabhanjan and Rosenberg, David and Mann, Gideon , title =. arXiv preprint arXiv:2303.17564 , year =

    Pith/arXiv arXiv

  13. [13]

    and Xiu, Dacheng , title =

    Chen, Yifei and Kelly, Bryan T. and Xiu, Dacheng , title =. 2023 , note =

    2023

  14. [14]

    Proceedings of the 5th ACM International Conference on AI in Finance (ICAIF '24) , year =

    Li, Xiaohui Victor and Sanna Passino, Francesco , title =. Proceedings of the 5th ACM International Conference on AI in Finance (ICAIF '24) , year =

  15. [15]

    Newman, Mark E. J. and Girvan, Michelle , title =. Physical Review E , year =

  16. [16]

    and Guillaume, Jean-Loup and Lambiotte, Renaud and Lefebvre, Etienne , title =

    Blondel, Vincent D. and Guillaume, Jean-Loup and Lambiotte, Renaud and Lefebvre, Etienne , title =. Journal of Statistical Mechanics: Theory and Experiment , year =

  17. [17]

    , title =

    Rosvall, Martin and Bergstrom, Carl T. , title =. Proceedings of the National Academy of Sciences , year =

  18. [18]

    and Laskey, Kathryn Blackmond and Leinhardt, Samuel , title =

    Holland, Paul W. and Laskey, Kathryn Blackmond and Leinhardt, Samuel , title =. Social Networks , year =

  19. [19]

    and Richardson, Thomas and Macon, Kevin and Porter, Mason A

    Mucha, Peter J. and Richardson, Thomas and Macon, Kevin and Porter, Mason A. and Onnela, Jukka-Pekka , title =. Science , year =

  20. [20]

    and Welling, Max , title =

    Kipf, Thomas N. and Welling, Max , title =. International Conference on Learning Representations (ICLR) , year =

  21. [21]

    Graph Attention Networks , booktitle =

    Veli. Graph Attention Networks , booktitle =. 2018 , note =

    2018

  22. [22]

    and Ying, Rex and Leskovec, Jure , title =

    Hamilton, William L. and Ying, Rex and Leskovec, Jure , title =. Advances in Neural Information Processing Systems 30 (NeurIPS) , year =

  23. [23]

    Physics Reports , year =

    Fortunato, Santo , title =. Physics Reports , year =

  24. [24]

    and West, Kenneth D

    Newey, Whitney K. and West, Kenneth D. , title =. Econometrica , year =

  25. [25]

    , title =

    Mantegna, Rosario N. , title =. European Physical Journal B , year =

  26. [26]

    and Ozdaglar, Asuman and Tahbaz-Salehi, Alireza , title =

    Acemoglu, Daron and Carvalho, Vasco M. and Ozdaglar, Asuman and Tahbaz-Salehi, Alireza , title =. Econometrica , year =

  27. [27]

    Global Business Networks , journal =

    Breitung, Christian and M. Global Business Networks , journal =. 2025 , volume =

    2025

  28. [28]

    SIGKDD 2023 Workshop on Robust NLP for Finance , year =

    Chen, Zihan and Zheng, Lei Nico and Lu, Cheng and Yuan, Jialu and Zhu, Di , title =. SIGKDD 2023 Workshop on Robust NLP for Finance , year =

    2023

  29. [29]

    Proceedings of the 6th ACM International Conference on AI in Finance (ICAIF '25) , year =

    Arun, Abhinav and Dimino, Fabrizio and Agarwal, Tejas Prakash and Sarmah, Bhaskarjit and Pasquali, Stefano , title =. Proceedings of the 6th ACM International Conference on AI in Finance (ICAIF '25) , year =

  30. [30]

    arXiv preprint arXiv:2604.19476 , year =

    Huang, Yikuan and Fan, Zheqi and Hu, Kaiqi and Ye, Yifan , title =. arXiv preprint arXiv:2604.19476 , year =

    Pith/arXiv arXiv

  31. [31]

    Review of Financial Studies , year =

    Hirshleifer, David and Peng, Lin and Wang, Qiguang , title =. Review of Financial Studies , year =

  32. [32]

    2020 , note =

    Schwenkler, Gustavo and Zheng, Hannan , title =. 2020 , note =

    2020

  33. [33]

    and Manela, Asaf and Xiu, Dacheng , title =

    Bybee, Leland and Kelly, Bryan T. and Manela, Asaf and Xiu, Dacheng , title =. Journal of Finance , year =

  34. [34]

    Management Science , year =

    Cong, Lin William and Liang, Tengyuan and Zhang, Xiao and Zhu, Wu , title =. Management Science , year =

  35. [35]

    and Xiu, Dacheng , title =

    Kelly, Bryan T. and Xiu, Dacheng , title =. Foundations and Trends in Finance , year =

  36. [36]

    and Poor, H

    Nie, Yuqi and Kong, Yaxuan and Dong, Xiaowen and Mulvey, John M. and Poor, H. Vincent and Wen, Qingsong and Zohren, Stefan , title =. arXiv preprint arXiv:2406.11903 , year =

    Pith/arXiv arXiv

  37. [37]

    ACM Transactions on Information Systems , year =

    Feng, Fuli and He, Xiangnan and Wang, Xiang and Luo, Cheng and Liu, Yiqun and Chua, Tat-Seng , title =. ACM Transactions on Information Systems , year =

  38. [38]

    arXiv preprint arXiv:1908.07999 , year =

    Kim, Raehyun and So, Chan Ho and Jeong, Minbyul and Lee, Sanghoon and Kim, Jinkyu and Kang, Jaewoo , title =. arXiv preprint arXiv:1908.07999 , year =

    Pith/arXiv arXiv 1908

  39. [39]

    Proceedings of the 31st ACM International Conference on Information & Knowledge Management (CIKM '22) , year =

    Xiang, Sheng and Cheng, Dawei and Shang, Chencheng and Zhang, Ying and Liang, Yuqi , title =. Proceedings of the 31st ACM International Conference on Information & Knowledge Management (CIKM '22) , year =

  40. [40]

    and Williams, Stacy and McDonald, Mark and Fenn, Daniel J

    Bazzi, Marya and Porter, Mason A. and Williams, Stacy and McDonald, Mark and Fenn, Daniel J. and Howison, Sam D. , title =. Multiscale Modeling & Simulation , year =

  41. [41]

    Applied Soft Computing , year =

    Wang, Ting and Guo, Jiale and Shan, Yuehui and Zhang, Yueyao and Peng, Bo and Wu, Zhuang , title =. Applied Soft Computing , year =

  42. [42]

    Finance Research Letters , year =

    Fan, Siyu and Wu, Yifei and Yang, Ruochen , title =. Finance Research Letters , year =

This paper was first reviewed by grok-4.5 on July 14, 2026.