Pith. sign in

REVIEW 4 major objections 7 minor 46 references

From Bias to Behavior: Learning Bull-Bear Market Dynamics with Contrastive Modeling

T0 review · 4 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read B4 claims that modeling news as a tug-of-war between bulls and bears improves market trend prediction and yields interpretable bias signals.

desk verdict A novel contrastive architecture for bull/bear financial modeling whose interpretability claim is undercut by supervision from next-day labels; the predictive evidence is also weaker than presented. read the letter →

arxiv 2507.14182 v1 pith:GR2PBR74 submitted 2025-07-12 cs.LG cs.AIstat.ML

Xiaotong Luo , Shengda Zhuo , Min Chen , Lichun Li , Ruizhao Lu , Wenqi Fan , Shuqiang Huang , Yin Tang This is my paper
classification cs.LGcs.AIstat.ML
keywords MarketRepresentationQuantitativeTradingContrastiveLearningBull-BearDynamicsInvestorBiasSentimentAnalysisFinancialNewsPriceTrendPrediction
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 tries to establish that financial market trends are better predicted when news is read as a competitive interaction between bullish and bearish investor interpretations, rather than as a single sentiment stream. It introduces B4, a model that embeds price sequences and news into a shared latent space, makes the bull-bear opposition explicit through sentiment markers, and pairs temporally adjacent samples to preserve momentum. On a backtest over 110 US stocks across 10 industries, the authors report that B4 beats seven baselines on cumulative and annual returns while keeping drawdowns low. They also argue that the attention difference between bullish and bearish embeddings measures investor bias, that this bias migrates across topics and time, and that its structural shifts often precede price reversals.

What carries the argument

The central object is the bias matrix $\mathbf{A}_{BU} - \mathbf{A}_{BE}$, the difference between attention maps derived from the [UP] and [DOWN] sentiment markers in the final layer of a sentiment-aware encoder; this difference is the model's operational definition of investor bias. The Price-to-Concept module aligns price windows with text prototypes via cross-attention, and the Inertial Pairing module constructs positive and negative samples using the condition that same-trend neighbors within a temporal span $\Delta$ are positives. The Dual Competition Mechanism then applies contrastive losses that push the competition representation of the winning camp toward the market representation, together with a cross-entropy loss that quantifies bullish versus bearish dominance, all balanced by a weight $\alpha$.

What would settle it

Compute B4's bias series on a held-out period and correlate it with an independent bullishness measure, such as the weekly AAII investor sentiment survey, the CBOE put/call ratio, or options-implied skew, testing leads of 1 to 20 trading days; if the bias series shows near-zero correlation or does not lead subsequent price returns, the interpretability claim is falsified and the advantage reduces to a supervised contrastive classifier.

Watch

Extended reading notes

Core claim

The central claim is that investor bias, defined as the difference between how bullish and bearish attention weight the same news, is a measurable and predictive quantity that drives market behavior. B4 operationalizes this by prepending [UP] and [DOWN] tokens to news, encoding the augmented text with a sentiment-aware language model, and reading off the attention maps those markers produce; the bias matrix $\mathbf{A}_{BU} - \mathbf{A}_{BE}$ then captures which features each camp attends to. An inertial pairing module builds positive and negative samples from same-trend and opposite-trend neighbors, and a dual competition mechanism with contrastive losses aligns the camp representations with the realized market representation. The paper reports that this joint design outperforms seven baselines in a backtest on 110 US stocks across 10 industries, and that the resulting bias time series shows interpretable patterns: symmetric but intense attention accompanies volatile chop, polarizing bias drives directional trends, and sudden bearish intrusion into formerly bullish topics coincides with price corrections.

Load-bearing premise

The load-bearing premise is that prepending [UP]/[DOWN] tokens to news and reading out their final-layer embeddings gives a faithful, causally prior measure of real investor bias, but the training loss in Eq. (13) explicitly aligns the camp chosen by the next-day label with the market representation, so the bias time series are partly manufactured by supervision.

Editorial extensions

If this is right

  • If the bias metric is causally prior to price moves, regime shifts could be detected from attention migration before the price reacts, giving an early-warning signal for trend reversals.
  • The model's interpretability layer connects behavioral-finance concepts, such as bias polarization and sentiment collapse, to concrete, computable time series derived from news attention.
  • The reported backtest results suggest that treating sentiment as two competing forces rather than a scalar signal can improve both returns and drawdown control in stock-level trading strategies.
  • The ablations indicate that the dual competition mechanism, bias simulation, inertial pairing, and price-to-concept alignment each contribute to the reported gains, with the contrastive competition playing the largest role.

Reading between the lines

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

  • Because Eq. (13) explicitly trains the camp chosen by the next-day label to align with the market representation, the 'bias' time series are partly manufactured by supervision; treating them as independent psychological measurements of investors requires out-of-sample validation against external stance data.
  • A strong test would be to correlate B4's bias series with an independent bullishness index, such as a survey-based sentiment measure or a put/call ratio, on held-out periods; high correlation would support the interpretability claim, and near-zero correlation would suggest the model is a supervised contrastive classifier wearing a behavioral label.
  • The attention-migration metrics could be extended to cross-asset or portfolio-level monitoring, a direction the paper itself lists as future work, and doing so would test whether industry-level bias flows generalize beyond single-stock case studies.
  • If the bias mechanism is real, the framework could also be applied to derivative pricing or systemic risk scenarios, where asymmetric interpretation of the same event matters, though that extension is not demonstrated here.
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.

Referee Report

4 major / 7 minor

Summary. The paper proposes B4, a contrastive learning framework for stock trend prediction that explicitly separates bullish and bearish investor interpretations. Price windows are mapped into a language-model semantic space via a Price-to-Concept module, news text is augmented with [UP]/[DOWN] markers, and the final-layer embeddings of these markers are used as competing bull/bear representations. Inertial Pairing builds positive/negative samples from next-day price direction with a momentum window, and a dual competition loss aligns the bull/bear embeddings with market representations. The authors evaluate on the CMIN dataset (110 US stocks, 2018-2021) against seven baselines and report cumulative returns, annual return, max drawdown, and Calmar ratio. They further define attention-based 'bias' metrics and present case studies claiming that bias dynamics foreshadow price moves.

Significance. If the predictive result held, B4 would be a useful addition to multimodal financial forecasting because it makes the bull-bear dichotomy explicit and couples it with a momentum-aware contrastive objective. The paper also has constructive features: the loss functions are fully specified, the dataset is public, and the ablation and hyperparameter studies address the main components. However, the second headline contribution -- that the attention-derived bias is an interpretable, causally prior measure of investor behavior -- is not supported as stated, because the bias representations are trained against next-day labels, making the lead-lag narratives partly circular. The performance claims are also not verifiable from the reported tables due to missing error bars, missing a buy-and-hold baseline, absent transaction costs, and internal numerical inconsistencies. As presented, the central claims require substantial additional evidence and a deconfounded evaluation protocol.

major comments (4)
  1. [§5, Eqs. (8), (11)-(13), Definitions 1-2] The interpretability claim is circular. The trend label y_t in Eq. (8) is the next-day direction; the cross-entropy loss in Eq. (13) explicitly trains the [UP]/[DOWN] embeddings so that the one matching y_t has the highest dot product with h_Mar, and Eqs. (11)-(12) align same-trend samples. The Attention Score and Bias in Definitions 1-2 (Section 5.1) are then read out from these same supervised attention maps. The lead-lag narratives in Section 5.3 ('bias collapses before a price drop') therefore restate the training objective rather than providing independent evidence that investor bias precedes price moves. To support the claim, the authors should retrain without L_CE or with shuffled labels and show that the bias series no longer leads price, and quantify the lead-lag against a null model.
  2. [Tables 1-2, 7 and §4.2] The performance claims are not verifiable as reported. Tables 1 and 7 give point estimates with no error bars, number of seeds, or significance tests; the backtest omits transaction costs, position sizing, and a buy-and-hold baseline. There is also an internal inconsistency: Table 1 reports a B4 Calmar ratio of 7.778 for EQIX, but the text in Section 4.2 states that CEformer's 4.597 is 'significantly higher than B4's 2.470', a number that appears nowhere in Table 1. The authors should correct the numbers and rerun the comparison with costs and a market baseline before the superiority claim can be assessed.
  3. [Definition 2, Section 5.1] Definition 2 calls Bias the 'absolute difference' in Attention Scores, but the displayed formula Bias = AS_bull - AS_bear is signed, and the subsequent case study in Section 5.3 relies on the sign (bullish versus bearish dominance). This is not just a typo: the metric's definition must be fixed, and the sign convention must be specified, before any directional bias interpretation can be evaluated.
  4. [Figs. 7-10 and Tables 4-5] The topic labels used in the interpretability section are inconsistent. Section 5.1 compares IAS for Z7 and Z6 from Table 4, while Figure 9's right panel uses Z-labels with Table 5 explanations (e.g., 'Z6 (Schall Law Firm class action)' is z6 in Table 5) and Table 4's caption still reads 'cf. Fig. ??.'. Because the case narratives in Section 5.3 depend on these topic codes, the reader cannot currently map the claimed bias patterns to concrete topics.
minor comments (7)
  1. [Section 1] The organization paragraph is copied from an unrelated paper about fusing voice interactions and knowledge graph information; it should be rewritten to describe the actual paper structure.
  2. [Figures 3, 6, 11] Several figure labels contain corrupted glyphs such as '/uni00000013' and similar escape sequences, making the figures unreadable; these need to be regenerated with a working TTF font.
  3. [Algorithm 1] The line '𝐄price ← Embed(𝐗Aug_t)' should almost certainly be '𝐄text ← Embed(𝐗Aug_t)', since the augmented text is being embedded rather than the price input.
  4. [Eq. (7) and Definition 1] Equation (7) defines A_BU and A_BE as outer-product matrices, while Definition 1 treats A_s,tau,z^{BU/BE} as scalar attention weights for a topic; the reduction from matrix to topic-level scalar is not explained.
  5. [Contributions, Section 1] The paper says '110 trades across 10 industries' in the contributions bullet but the dataset description and tables refer to 110 stocks; the wording should be corrected.
  6. [Section 4.3] The sentence 'B4 w/o BS shows improved performance' is ambiguous and appears to contradict the following clause about Bias Simulation enhancing profitability; it should be reworded to state that removing BS degrades performance.
  7. [Table 1 caption] The caption lists baseline order 'ESPMP, DUALformer, and SCLformer' but the table columns are ordered 'DUALformer, CEformer, SCLformer, ESPMP'; the caption should match the table.

Circularity Check

1 steps flagged · score 6.0 of 10

Bias series in Section 5 is read out from the same [UP]/[DOWN] embeddings that Eq. 13 supervises with next-day labels, so the interpretability claim is partly circular; the predictive backtest remains externally benchmarked.

  1. fitted input called prediction [Section 3.2, Eq. (13); Section 5.1, Definitions 1-2; Section 5.3, Fig. 10]
    "𝐡⋆ Comp,i denotes the columns of 𝐡Comp,i corresponding to the ground-truth label of 𝐡Mar,i. ... LCE=−1N∑i∈Nlog exp(LCM(i,i))exp(LUM(i,i))+exp(LEM(i,i)), where LCM(i,i)=𝐡⋆Comp,i⋅𝐡Mar,i, LUM(i,i)=𝐡BU,i⋅𝐡Mar,i, LEM(i,i)=𝐡BE,i⋅𝐡Mar,i. ... Definition 1. (AttentionScore,AS) ... ASbull/bears,τ,z=1000⋅|||ABU/BEs,τ,z|||, where ABU and ABE are the bullish and bearish attention weights derived from B4 model. Definition 2. (Bias) Biass,τ,z=ASbulls,τ,z−ASbears,τ,z."

    Eq. (13) is a cross-entropy loss that maximizes the dot product of the [UP]/[DOWN] marker embedding matching the ground-truth next-day direction y_t (Eq. 8) with the market representation, while suppressing the opposite marker. The Section 5 bias metric is then defined as the difference of attention maps built from these same h_BU and h_BE embeddings (Eqs. 6-7, Definitions 1-2). The paper interprets the resulting series as an independent behavioral signal that 'preempts' or 'triggers' price moves (Section 5.3, Fig. 10). But the model was explicitly rewarded for making the bull/bear embeddings align with the next-day label, so the bias time series is partly a re-statement of its own training signal rather than an independent cause.

full rationale

B4's engineering is largely self-contained: the model is trained on the public CMIN dataset and compared against seven baselines on returns, drawdown, and Calmar, so the predictive-accuracy claim is externally falsifiable and not circular. The circular step is the second headline contribution: the 'bias' time series. Eq. (13) supervises the [UP]/[DOWN] marker embeddings so that the marker matching the next-day label has higher similarity to the market representation, and Section 5 defines Bias as the difference of attention maps from those very embeddings. Narratives such as 'the bias collapsed and triggered a rapid price drop' (Section 5.3) therefore partly describe the model's own supervision rather than an independent behavioral measurement. This does not make the backtest results circular, but it makes the interpretability claim partially forced by construction. The internal reporting inconsistencies (EQIX Calmar 2.470 in Table 1 vs 7.778 in Table 2, mismatched Z/z topic labels, and the Section 1 organization sentence copied from an unrelated fusion paper) are correctness and provenance issues rather than additional circularity; they are noted here because they further reduce confidence in the paper's stated claims. Overall: score 6, partial circularity in the interpretability component, no forced circularity in the prediction benchmark itself.

Assumptions & free parameters 4 free parameters · 5 assumptions · 2 invented entities

The central claims rest on a small set of hyperparameters (alpha, Delta, tau, window length) and on domain assumptions that equate model-internal attention patterns with real investor psychology. The model is a supervised contrastive learner whose 'bias' outputs are trained against next-day direction labels, so the interpretability findings in Section 5 are not independent evidence. No code, error bars, or backtest details are provided.

free parameters (4)
  • alpha (loss weight) = not reported per stock; grid searched over {0.1, 0.3, 0.5, 0.7, 0.9}
    Eq (14) balances contrastive alignment and cross-entropy; chosen on a validation set separately for each stock.
  • MomentumSpan Delta = not reported per stock; grid searched over {-2,-1,0,1,2,±1,±2}
    Eqs (9)-(10) define positive/negative samples; the best configuration per stock is not listed in the paper.
  • temperature tau = not specified
    Used in the contrastive losses Eqs (11)-(12); no value or tuning procedure is given.
  • lookback window delta = not specified
    Price window length in the Price-to-Concept module is a free design choice; no value is reported.
assumptions (5)
  • domain assumption Next-day close direction is a valid and sufficient label for market trend.
    Eq (8) defines y_t from x_{t+1} versus x_t; all training and evaluation depend on this binary proxy for bull/bear regimes.
  • ad hoc to paper Prepending [UP]/[DOWN] tokens to news and extracting their embeddings yields a faithful representation of real bull and bear investor interpretations.
    Eqs (2), (6), and (7); no external validation against actual investor surveys or trades.
  • ad hoc to paper Difference of bull and bear attention maps A_BU - A_BE is a valid operationalization of cognitive bias.
    Section 5.1, Definition 2; the paper assumes attention divergence equals bias without independent measurement.
  • ad hoc to paper Momentum persists across a small temporal neighborhood Delta, so same-label samples within |i-j|<=Delta are positives.
    Eqs (9)-(10); no empirical test of the momentum span assumption is provided.
  • domain assumption BERT pretrained weights transfer to financial news semantics after per-stock fine-tuning.
    Section 4.1 D; a standard transfer-learning assumption that is not tested for this domain.
invented entities (2)
  • Bullish and bearish force embeddings (h_BU, h_BE)
    purpose: Represent the two competing investor camps in latent space and drive the dual competition loss.
    These vectors are internal model parameters trained with next-day direction labels (Eqs 8, 13); no external data on real investor beliefs supports them.
  • Bias matrix (A_BU - A_BE)
    purpose: Interpretable measure of divergence between bull and bear attention, used in Section 5 to explain price moves.
    Derived from the supervised attention weights; the claimed bias-price link is partly manufactured by training on future returns.

how reviews work

0 comments
Cite this review

Pith. "Pith review of From Bias to Behavior: Learning Bull-Bear Market Dynamics with Contrastive Modeling." pith.science (2026). https://pith.science/paper/GR2PBR74

@misc{pith2026250714182,
  author       = {Pith},
  title        = {Pith review of: From Bias to Behavior: Learning Bull-Bear Market Dynamics with Contrastive Modeling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GR2PBR74}},
  note         = {Machine review of arXiv:2507.14182}
}
read the original abstract

Financial markets exhibit highly dynamic and complex behaviors shaped by both historical price trajectories and exogenous narratives, such as news, policy interpretations, and social media sentiment. The heterogeneity in these data and the diverse insight of investors introduce biases that complicate the modeling of market dynamics. Unlike prior work, this paper explores the potential of bull and bear regimes in investor-driven market dynamics. Through empirical analysis on real-world financial datasets, we uncover a dynamic relationship between bias variation and behavioral adaptation, which enhances trend prediction under evolving market conditions. To model this mechanism, we propose the Bias to Behavior from Bull-Bear Dynamics model (B4), a unified framework that jointly embeds temporal price sequences and external contextual signals into a shared latent space where opposing bull and bear forces naturally emerge, forming the foundation for bias representation. Within this space, an inertial pairing module pairs temporally adjacent samples to preserve momentum, while the dual competition mechanism contrasts bullish and bearish embeddings to capture behavioral divergence. Together, these components allow B4 to model bias-driven asymmetry, behavioral inertia, and market heterogeneity. Experimental results on real-world financial datasets demonstrate that our model not only achieves superior performance in predicting market trends but also provides interpretable insights into the interplay of biases, investor behaviors, and market dynamics.

Figures

Figures reproduced from arXiv: 2507.14182 by the authors.

Figure 1
Figure 1. Light green triangles represent “bullish dominance” scenarios, and light red triangles represent “bearish dominance” scenarios, indicating the competitive interactions between bears and bulls. Red circles denote “bear markets”, while green circles denote “bull markets”, reflecting the outcomes of this competition corresponding to downward and upward trends, respectively. In postprocessing of B4, data points are more… view at source ↗
Figure 2
Figure 2. The overall architecture of B4 model for dynamic market modeling. The framework is designed to address key questions regarding the integration of temporal data and the mitigation of bias to extract insights from news events, as well as capturing price momentum and reflecting bull-bear competition to produce actionable trading insights. techniques [19] improve robustness, but still treat behavior implicitly or linear… view at source ↗
Figure 3
Figure 3. Performance Comparison of magnitude 𝜂 of CL. 1. Annual Return: The annual return is calculated by tak￾ing the geometric average of the return over the period and subtracting one. It represents the average annual profit or loss from an investment: Annual Return = (𝑅 𝑛 )1 𝑡 − 1, where 𝑅 is the total return over the period, 𝑛 is the number of sub-periods ( e.g., years), and 𝑡 is the number of years. 2. Cumulative Retur… view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Ablation study on key components of the proposed method, evaluating the impact of removing Price2Concept (P2C), Bias Simulation (BS), Inertial Pairing (IP), and Dual Competition Mechanism (DCM). Overall, B4 strikes a balance between high returns and low risk. Although …
Figure 6
Figure 6. Figure 6: The performance of the 𝛼 and Δ on the Annual Return metric for AAPL and ECL stocks. performance becomes significantly influenced by changes in Δ, showing greater volatility. This suggests that under high signal extraction weights, the model’s sensitivity to momen￾tum s…
Figure 5
Figure 5. Figure 5: B4 under different combinations of hyperparameters. Larger bubbles mean higher performances. 4.4. Hyper-parameters Combinations (RQ3) We further analyze the effects of the two key hyper￾parameters introduced in Implementation Details: the loss￾balancing weight 𝛼 and th…
Figure 7
Figure 7. Figure 7: The bias of stock and Industry. Left: Variations in attention patterns by 𝐴𝑆 distribution between bullish and bearish investors. Right: Differential attention, showed by 𝐼𝐴𝑆𝑇 𝑒𝑐ℎ, to various topics within a specific industry (e.g., Technology Sector) across bullish and…
Figure 8
Figure 8. Figure 8: (Left) depicts the landscape of attention flow and biases shift over time. Each flow line represents the mag￾nitude of 𝐴𝑀, uncovering several important patterns: (1) Attention flows as hot topics emerge or evolve. (2) The attention to a topic at time 𝜏 is influenced by…
Figure 10
Figure 10. Figure 10: The classification codes for topics are presented in Appendix 5. Volatility under Symmetric but Intense Attention. In some cases, both bulls and bears focus intensely on the same high-stakes themes, such as lawsuits (z1, z2), resulting in a low net bias but a high tot…
Figure 9
Figure 9. Figure 9: (Right) displays a heatmap of the Bias Migration Propensity (𝐵𝑀𝑃 ), reflecting how bullish and bearish per￾spectives diverge across industries and topics. Green indi￾cates topics where bulls dominate attention migration; red signals bear dominance. Z1 Z2 Z3 Z4 Z5Z6 Z7 …
Figure 11
Figure 11. Figure 11: The comparative performance of eight distinct models on different stock over a series of time steps. The models include B4, LSTM, Stocknet, Transformer, CE-DNN, and ESPMP, along with DUAL-DNN and SCL-DNN highlighted in the legend. Each line represents the progression …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 35 canonical work pages

  1. [1]

    The influence of risk perception, risk tolerance, overconfidence, and loss aversion towards investment decision making

    Ainia, N.S.N., Lutfi, L., 2019. The influence of risk perception, risk tolerance, overconfidence, and loss aversion towards investment decision making. Journal of Economics, Business, & Accountancy Ventura 21, 401–413

    work page 2019

  2. [2]

    The pricing of options and corporate liabilities

    Black, F., Scholes, M., 1973. The pricing of options and corporate liabilities. Journal of political economy 81, 637–654

    work page 1973

  3. [3]

    Latent dirichlet allocation

    Blei, D.M., Ng, A.Y., Jordan, M.I., 2003. Latent dirichlet allocation. J. Mach. Learn. Res. 3, 993–1022

    work page 2003

  4. [4]

    Generalized autoregressive conditional het- eroskedasticity

    Bollerslev, T., 1986. Generalized autoregressive conditional het- eroskedasticity. Journal of econometrics 31, 307–327

    work page 1986

  5. [5]

    Arch models

    Bollerslev, T., 1994. Arch models. Handbook of Econometrics 4

    work page 1994

  6. [6]

    Journal of Banking & Finance 24, 1651–1679

    Chang,E.C.,Cheng,J.W.,Khorana,A.,2000.Anexaminationofherd behavior in equity markets: An international perspective. Journal of Banking & Finance 24, 1651–1679

    work page 2000

  7. [7]

    Dual Contrastive Learning: Text Classification via Label-Aware Data Augmentation

    Chen, Q., Zhang, R., Zheng, Y., Mao, Y., 2022. Dual contrastive learning:Textclassificationvialabel-awaredataaugmentation. arXiv preprint arXiv:2201.08702

    work page Pith review arXiv 2022

  8. [8]

    Financial time se- ries forecasting with multi-modality graph neural network

    Cheng, D., Yang, F., Xiang, S., Liu, J., 2022. Financial time se- ries forecasting with multi-modality graph neural network. Pattern Recognition 121, 108218. doi: https://doi.org/10.1016/j.patcog. 2021.108218

    arXiv 2022

Show all 46 references
  1. [9]

    Sentiment analysis on stock social mediaforstockpricemovementprediction

    Derakhshan, A., Beigy, H., 2019. Sentiment analysis on stock social mediaforstockpricemovementprediction. Engineeringapplications of artificial intelligence 85, 569–578

    2019

  2. [10]

    Tradexpert: Revolutionizing trading with mixture of expert llms

    Ding, Q., Shi, H., Liu, B., 2024. Tradexpert: Revolutionizing trading with mixture of expert llms. arXiv preprint arXiv:2411.00782

    2024 arXiv

  3. [11]

    Du,K.,Mao,R.,Xing,F.,Cambria,E.,2024. Explainablestockprice movement prediction using contrastive learning, in: Proceedings of the 33rd ACM International Conference on Information and Knowl- edgeManagement,AssociationforComputingMachinery,NewYork, NY, USA. p. 529–537. URL: htt...

    2024

  4. [12]

    Mf-clr: Multi-frequency contrastive learning representation for time series, in: International Conference on Machine Learning

    Duan, J., Zheng, W., Du, Y., Wu, W., Jiang, H., Qi, H., 2024. Mf-clr: Multi-frequency contrastive learning representation for time series, in: International Conference on Machine Learning. URL: https: //api.semanticscholar.org/CorpusID:272330491

    2024

  5. [13]

    Modality-aware transformer for financial time series forecasting, in: Proceedings of the 5th ACM International Conference on AI in Finance, pp

    Emami Gohari, H., Dang, X.H., Shah, S.Y., Zerfos, P., 2024. Modality-aware transformer for financial time series forecasting, in: Proceedings of the 5th ACM International Conference on AI in Finance, pp. 677–685

    2024

  6. [14]

    A sentiment-enhanced hybrid model for crude oil price forecasting

    Fang, Y., Wang, W., Wu, P., Zhao, Y., 2023. A sentiment-enhanced hybrid model for crude oil price forecasting. Expert Systems with Applications 215, 119329

    2023

  7. [15]

    Temporal relational ranking for stock prediction

    Feng, F., He, X., Wang, X., Luo, C., Liu, Y., Chua, T.S., 2019. Temporal relational ranking for stock prediction. ACM Transactions on Information Systems (TOIS) 37, 1–30

    2019

  8. [16]

    Stockformer: Learning hybrid trading machines with predictive coding., in: IJCAI, pp

    Gao, S., Wang, Y., Yang, X., 2023. Stockformer: Learning hybrid trading machines with predictive coding., in: IJCAI, pp. 4766–4774

    2023

  9. [17]

    The random- walk hypothesis of stock market behavior a

    Godfrey, M.D., Granger, C.W., Morgenstern, O., 1964. The random- walk hypothesis of stock market behavior a. Kyklos 17, 1–30

    1964

  10. [18]

    URL: https://arxiv.org/abs/2203.05794, arXiv:2203.05794

    Grootendorst,M.,2022.Bertopic:Neuraltopicmodelingwithaclass- based tf-idf procedure. URL: https://arxiv.org/abs/2203.05794, arXiv:2203.05794

    2022 arXiv

  11. [19]

    Predictive performance of denoising algorithms in s&p 500 and bitcoin returns

    Gulay, E., Akgun, O.B., Bekiroglu, K., Duru, O., 2025. Predictive performance of denoising algorithms in s&p 500 and bitcoin returns. Expert Systems with Applications 260, 125400

    2025

  12. [20]

    Long short-term memory

    Hochreiter, S., Schmidhuber, J., 1997. Long short-term memory. NeuralComput.9,1735–1780. URL: https://doi.org/10.1162/neco. 1997.9.8.1735, doi:10.1162/neco.1997.9.8.1735

    1997 doi

  13. [21]

    Hou, M., Xu, C., Liu, Y., Liu, W., Bian, J., Wu, L., Li, Z., Chen, E., Liu,T.Y.,2021. Stocktrendpredictionwithmulti-granularitydata:A contrastivelearningapproachwithadaptivefusion,in:Proceedingsof the30thACMInternationalConferenceonInformation&Knowledge Management, pp. 700–709

    2021

  14. [22]

    Finbert: A large language model for extracting information from financial text

    Huang, A.H., Wang, H., Yang, Y., 2023. Finbert: A large language model for extracting information from financial text. Contemporary Accounting Research 40, 806–841

    2023

  15. [23]

    Bi-level attention graph neural networks

    Iyer, R.G., Wang, W., Sun, Y., 2023. Bi-level attention graph neural networks. URL: https://arxiv.org/abs/2304.11533, arXiv:2304.11533

    2023 arXiv

  16. [24]

    Behavioral portfolio selection in con- tinuous time

    Jin, H., Yu Zhou, X., 2008. Behavioral portfolio selection in con- tinuous time. Mathematical Finance: An International Journal of : Preprint submitted to Elsevier Page 13 of 17 Mathematics, Statistics and Financial Economics 18, 385–426

    2008

  17. [25]

    Time-llm: Time series forecastingbyreprogramminglargelanguagemodels

    Jin, M., Wang, S., Ma, L., Chu, Z., Zhang, J.Y., Shi, X., Chen, P.Y., Liang, Y., Li, Y.F., Pan, S., et al., 2023. Time-llm: Time series forecastingbyreprogramminglargelanguagemodels. arXivpreprint arXiv:2310.01728

    2023 arXiv

  18. [26]

    Supervised contrastive learning

    Khosla, P., Teterwak, P., Wang, C., Sarna, A., Tian, Y., Isola, P., Maschinot, A., Liu, C., Krishnan, D., 2021. Supervised contrastive learning. URL:https://arxiv.org/abs/2004.11362, arXiv:2004.11362

    2021 arXiv

  19. [27]

    Finreport: Explainable stock earnings forecasting via news factor analyzing model, in: Companion Proceedings of the ACM on Web Conference 2024, pp

    Li, X., Shen, X., Zeng, Y., Xing, X., Xu, J., 2024a. Finreport: Explainable stock earnings forecasting via news factor analyzing model, in: Companion Proceedings of the ACM on Web Conference 2024, pp. 319–327

    2024

  20. [28]

    Finreport: Explainable stock earnings forecasting via news factor analyzing model

    Li, X., Shen, X., Zeng, Y., Xing, X., Xu, J., 2024b. Finreport: Explainable stock earnings forecasting via news factor analyzing model. Companion Proceedings of the ACM on Web Conference 2024 URL:https://api.semanticscholar.org/CorpusID:268247835

    2024

  21. [29]

    The extra value of online investor sentiment measures on forecasting stock return volatility: A large- scale longitudinal evaluation based on chinese stock market

    Lin, P., Ma, S., Fildes, R., 2024. The extra value of online investor sentiment measures on forecasting stock return volatility: A large- scale longitudinal evaluation based on chinese stock market. Expert Systems with Applications 238, 121927

    2024

  22. [30]

    Focal loss for dense object detection, in: 2017 IEEE International Conference onComputerVision(ICCV),pp.2999–3007

    Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollár, P., 2017. Focal loss for dense object detection, in: 2017 IEEE International Conference onComputerVision(ICCV),pp.2999–3007. doi: 10.1109/ICCV.2017. 324

    2017 doi

  23. [31]

    Autotimes: Autoregressive time series forecasters via large language models

    Liu, Y., Qin, G., Huang, X., Wang, J., Long, M., 2024. Autotimes: Autoregressive time series forecasters via large language models. arXiv preprint arXiv:2402.02370

    2024 arXiv

  24. [32]

    Causality- guidedmulti-memoryinteractionnetworkformultivariatestockprice movement prediction, in: Annual Meeting of the Association for Computational Linguistics

    Luo, D., Liao, W., Li, S., Cheng, X., Yan, R., 2023. Causality- guidedmulti-memoryinteractionnetworkformultivariatestockprice movement prediction, in: Annual Meeting of the Association for Computational Linguistics. URL:https://api.semanticscholar.org/ CorpusID:259370608

    2023

  25. [33]

    Mm- itransformer: A multimodal approach to economic time series fore- casting with textual data

    Mou, S., Xue, Q., Chen, J., Takiguchi, T., Ariki, Y., 2025. Mm- itransformer: A multimodal approach to economic time series fore- casting with textual data. Applied Sciences 15, 1241

    2025

  26. [34]

    Sentimentanalysisoffinancial news and its impact on the stock market, in: 2024 2nd World Confer- ence on Communication & Computing (WCONF), IEEE

    Patil,A.,Sharma,H.,Sinha,A.,2024. Sentimentanalysisoffinancial news and its impact on the stock market, in: 2024 2nd World Confer- ence on Communication & Computing (WCONF), IEEE. pp. 1–5

    2024

  27. [35]

    Pei,Y.,Cartlidge,J.,Mandal,A.,Gold,D.,Marcilio,E.,Mazzon,R.,

  28. [36]

    Msmf: Multi-scale multi-modal fusion for enhanced stock market prediction

    Qin, J., 2024. Msmf: Multi-scale multi-modal fusion for enhanced stock market prediction. arXiv preprint arXiv:2409.07855

    2024 arXiv

  29. [37]

    Attentionisallyouneed

    Vaswani,A.,Shazeer,N.,Parmar,N.,Uszkoreit,J.,Jones,L.,Gomez, A.N.,Kaiser,L.,Polosukhin,I.,2023. Attentionisallyouneed. URL: https://arxiv.org/abs/1706.03762, arXiv:1706.03762

    2023 arXiv

  30. [38]

    Market-gan: Adding control to financial market data generation with semantic context, in: Proceedings of the AAAI Conference on Artificial Intelligence, pp

    Xia, H., Sun, S., Wang, X., An, B., 2024. Market-gan: Adding control to financial market data generation with semantic context, in: Proceedings of the AAAI Conference on Artificial Intelligence, pp. 15996–16004

    2024

  31. [39]

    3584–3593

    Xiang,S.,Cheng,D.,Shang,C.,Zhang,Y.,Liang,Y.,2022.Temporal and heterogeneous graph neural network for financial time series prediction,in:Proceedingsofthe31stACMinternationalconference on information & knowledge management, pp. 3584–3593

    2022

  32. [40]

    Xu, Y., Cohen, S.B., 2018. Stock movement prediction from tweets and historical prices, in: Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics, Association for Computational Linguistics, Melbourne, Australia. pp. 1970–1979

    2018

  33. [41]

    A large-scale microblog dataset and stock movementpredictionbasedonsupervisedcontrastivelearningmodel

    Yang, S., Tang, D., 2024. A large-scale microblog dataset and stock movementpredictionbasedonsupervisedcontrastivelearningmodel. Neurocomputing 584, 127583. URL:https://api.semanticscholar. org/CorpusID:268670132

    2024

  34. [42]

    Financialtimeseriesforecastingusingcnnandtransformer

    Zeng, Z., Kaur, R., Siddagangappa, S., Rahimi, S., Balch, T., Veloso, M.,2023. Financialtimeseriesforecastingusingcnnandtransformer. arXiv preprint arXiv:2304.04912 1

    2023 arXiv

  35. [43]

    Multimodal sentiment analysis based on fusion methods: A survey

    Zhu, L., Zhu, Z., Zhang, C., Xu, Y., Kong, X., 2023. Multimodal sentiment analysis based on fusion methods: A survey. Information Fusion 95, 306–325

    2023

  36. [44]

    Stock movement prediction with mul- timodal stable fusion via gated cross-attention mechanism

    Zong, C., Zhou, H., 2024. Stock movement prediction with mul- timodal stable fusion via gated cross-attention mechanism. URL: https://arxiv.org/abs/2406.06594, arXiv:2406.06594

    2024 arXiv

  37. [45]

    A novel deep reinforcement learning based automated stock trading system using cascaded lstm networks

    Zou, J., Lou, J., Wang, B., Liu, S., 2024. A novel deep reinforcement learning based automated stock trading system using cascaded lstm networks. Expert Systems with Applications 242, 122801. : Preprint submitted to Elsevier Page 14 of 17 Table 3 Summary of Variables in Attent...

    2024

  38. [2025]

    URL: https://arxiv.org/abs/2504.13522, arXiv:2504.13522

    Cross-modal temporal fusion for financial market forecasting. URL: https://arxiv.org/abs/2504.13522, arXiv:2504.13522

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.