Pith. sign in

REVIEW 2 major objections 5 minor 40 references

FactorGCL: A Hypergraph-Based Factor Model with Temporal Residual Contrastive Learning for Stock Returns Prediction

T0 review · 2 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read FactorGCL claims that decomposing stock returns into prior factors, mined hidden factors, and idiosyncratic alpha, with a temporal contrastive regularizer, yields the most accurate and profitable predictions on China A-shares tested here.

desk verdict Genuinely novel architecture, but the headline numbers are compromised by a likely lookahead leak in the contrastive loss that the paper never addresses. read the letter →

arxiv 2502.05218 v1 pith:XQORNFDN submitted 2025-02-05 q-fin.ST cs.AIcs.LG

Yitong Duan , Weiran Wang , Jian Li This is my paper · ORCID
classification q-fin.STcs.AIcs.LG
keywords factormodelhypergraphneuralnetworkstockreturnpredictioncontrastivelearninghiddenminingtemporalresidualChinaA-sharesinformationcoefficient
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 a factor model can be made more predictive by letting the model mine hidden factors from what is left after known industry factors have done their work. It proposes FactorGCL, which decomposes predicted returns into three pieces—prior beta from expert-designed industry factors, hidden beta from data-mined factors, and an idiosyncratic alpha—and builds each piece with hypergraph convolutions, treating stocks as nodes and factors as hyperedges. To keep the mined factors from fitting noise, it adds a temporal residual contrastive loss that asks the alpha residuals from historical and future windows of the same stock to agree. On China A-shares data from 2014 to 2023, the method reports the highest information coefficient and ICIR among thirteen models and the most profitable top-K backtests, which matters because it suggests data-driven factors can complement expert factors in real trading.

What carries the argument

The central machinery is a cascading residual hypergraph architecture built on hypergraph convolution, the message-passing rule that aggregates node features along hyperedges. Stocks are nodes; the 83 secondary-industry factors are prior hyperedges, and mined hidden factors are generated as soft hyperedges $\beta_h^{(i,j)} = \mathrm{Sigmoid}(e_r^{(i)} \cdot c^{(j)T})$ from residual embeddings $e_r = e_s - e_p$. Each stage removes the previous stage's influence before the next begins, so predictions are $\hat{y} = w_{o1} e_p + w_{o2} e_h + w_{o3} e_\alpha + b_o$. The temporal residual contrastive loss, InfoNCE applied to $e_\alpha$ computed from historical and future windows, is the mechanism that pushes the hidden factors to be both effective and comprehensive by enforcing temporal consistency of idiosyncratic residuals.

What would settle it

Re-run the same China A-shares experiment on an identical train/validation/test split, but truncate or discard any sample whose next-20-days future window crosses a split boundary, or insert a 20-trading-day gap between periods. If FactorGCL's reported IC and backtest advantage over the best baseline shrinks to within noise, the temporal contrastive loss is borrowing future information rather than mining persistent hidden factors.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that residual decomposition plus hypergraph message passing recovers hidden factors that expert industry factors leave unexplained, and that contrasting the leftover idiosyncratic residuals across time keeps those factors effective out of sample. Concretely, the model computes prior embeddings $e_p$ from the industry-factor hypergraph, subtracts them from stock features to get residuals $e_r$, generates soft hidden-factor exposures from similarity between $e_r$ and learned prototypes, and obtains $\alpha$ from the remaining residual. The temporal contrastive loss applies InfoNCE to the $\alpha$ embeddings $e_\alpha$ from historical and future data, pulling same-stock pairs together and pushing different-stock pairs apart. The reported result is that FactorGCL achieves the best IC and ICIR at every prediction horizon ($\Delta t = 1,5,10,20$) among all compared baselines, and that each component—prior $\beta$, hidden $\beta$, $\alpha$, and contrastive loss—contributes to the final score in ablations.

Load-bearing premise

The load-bearing assumption is that the future 20-day window used in the contrastive loss does not leak test-period information into training or validation, yet the paper does not state whether samples near split boundaries are truncated or excluded.

Editorial extensions

If this is right

  • Each component of the cascade earns its keep: removing the prior beta module, hidden beta module, alpha module, or contrastive loss lowers IC at every tested horizon, so the decomposition itself is the source of the gain.
  • Hidden factors help most in moderation: performance rises as the number of hidden factors $M$ grows from 0 to 64 but falls at $M=128$, indicating an overfitting regime that practitioners should tune.
  • The model's advantage persists across prediction horizons from 1 to 20 trading days, with longer horizons showing the largest IC, so the multi-period objective gives a more reliable cross-sectional ranking.
  • When used as a stock-selection signal with a TopK=30 strategy on CSI300 and CSI500 constituents, the method produces higher annualized return, information ratio, and return-over-maximum-drawdown than all compared baselines under 0.3% transaction costs.

Reading between the lines

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

  • If the split-boundary leakage concern is resolved, the temporal residual contrastive idea would likely transfer to other cross-sectional prediction problems where expert factors are incomplete, not just stock returns.
  • The contrastive loss can be viewed as a temporal-invariance regularizer on alpha; a cheaper non-contrastive consistency penalty might capture part of the gain, but the paper does not test that.
  • The same cascading residual decomposition could be applied to risk-factor mining and portfolio optimization, directions the authors list as future work.
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

2 major / 5 minor

Summary. The paper proposes FactorGCL, a hypergraph-based nonlinear factor model that decomposes stock returns into prior beta, hidden beta, and individual alpha components via a cascading residual architecture. Hidden factors are generated as soft hyperedges using learnable prototypes, and a temporal residual contrastive loss (Eqs. 9-11) aligns alpha embeddings computed from historical and future windows. The authors report IC, ICIR, and backtest results on China A-shares data from 2014 to 2023, claiming that FactorGCL outperforms existing state-of-the-art baselines and mines effective hidden factors for stock return prediction.

Significance. If the empirical claims hold, the architecture is a relevant contribution to machine-learning-based factor models: the cascading residual decomposition is principled, the hypergraph treatment of factors is sensible, and the temporal contrastive objective is well motivated. The paper also benchmarks against a broad set of baselines and reports multiple financial metrics. However, the central empirical claim is currently not reliable because the future window used in the contrastive loss can cross temporal split boundaries, and the paper does not describe any boundary handling rule. The hidden-factor contribution is also asserted rather than directly evaluated. With a clean temporal split and direct factor-level analysis, the work could become a useful contribution.

major comments (2)
  1. [Methodology, Eqs. (9)-(11); Experiments, 'Experiment Settings'] The temporal residual contrastive loss computes future alpha embeddings e'_alpha from x', described as the next 20 days of data, through modules that share parameters with the prediction path. For any training or validation sample within 20 trading days of a split boundary, x' extends into the validation or test period, so information from those periods can enter gradient updates and early stopping. The same issue affects the multi-period labels y^(l) in Eq. (12), whose windows can also cross into the next split. The paper only states that the dataset is split 5 years / 1 year / 2 years in temporal order with a rolling method; it never states that boundary samples are dropped or that a gap is inserted. Because every reported IC, ICIR, and backtest number in Tables 1-3 depends on this split, the authors must either specify a boundary rule that prevents any future window from crossing into the next split or rerun the experiments under such a rule.
  2. [Hidden Beta Module; Main Results] The paper claims that FactorGCL 'mines effective hidden factors,' but the hidden factor prototypes c^(i) and the hidden exposure matrix beta_h are never analyzed or validated directly. The only evidence offered is aggregate prediction improvement and backtest profitability, which conflates the quality of the extracted factors with the quality of the overall model fit. To support this contribution, the authors should provide direct out-of-sample evidence, such as factor-mimicking portfolio returns, correlations with known factors, stability of the learned hidden factors across time, or an economic interpretation of the extracted exposures.
minor comments (5)
  1. [Eq. (11)] The definition of the projection p is incomplete: 'p(x) = is a 2-layer MLP' appears to have a missing expression. Please clarify the exact architecture, including whether normalization or a specific output dimension is used.
  2. [Preliminaries, Eq. (2); Experiment Settings] The notation for future returns is inconsistent: Eq. (2) defines y_t = (price_{t+Delta t} - price_t) / price_t, while the experiments use y_t = (price_{t+Delta t + 1} - price_{t+1}) / price_{t+1}. Please reconcile the two definitions.
  3. [Investment Simulation, Figure 7] Figure 7 shows backtest curves starting on 2020-08-06, although the test period is stated as 01/01/2020 to 06/30/2023. Please explain the discrepancy.
  4. [Main Results, Table 1] The paper reports only point estimates of IC and ICIR without standard errors or significance tests. Given the modest differences between methods, the authors should add statistical significance measures such as Newey-West adjusted t-statistics or bootstrap confidence intervals.
  5. [Figure 6] Please clarify how the M=0 setting removes the hidden beta module while keeping the rest of the architecture intact, since Figure 6 reports a nonzero IC for M=0.

Circularity Check

1 steps flagged · score 6.0 of 10

Temporal residual contrastive learning uses future data x′ that overlaps the label window, so the reported out-of-sample predictions are partly forced by construction; no boundary-sample exclusion is stated.

  1. fitted input called prediction [Methodology, Temporal Residual Contrastive Learning, Eqs. (9)-(11) and Experiment Settings]
    "given future data x′ ∈ R^{N×T′×D} ... we use the prior factor exposure β and hidden factor exposure βh extracted from historical data to calculate the future alpha embedding e′α: e′s = ϕ′feat(x′); e′α = e′s − ϕ′prior(e′s, β) − ϕ′hidden(e′r, βh) ... The length of historical squence data x is T = 60, and the length of future data x′ is T ′ = 20. ... the future return is calculated by the formula yt = pricet+∆t+1−pricet+1 / pricet+1."

    In Eq. (10), the future alpha embedding e′α is computed from x′, a 20-day future window. The labels yt are returns over ∆t = 1, 5, 10, 20 days starting at t+1, i.e., the same calendar window as x′. The contrastive loss (Eq. 11) uses e′α and, through the total objective (Eq. 13), updates all shared parameters (ϕfeat, ϕprior, ϕhidden). Thus the model is trained on the exact time window it is later evaluated on. The paper describes the split only as a temporal 5y/1y/2y rolling method and never states that boundary samples are dropped or a gap is inserted; consequently, any training or validation sample within 20 days of a split boundary has x′ reaching into the next split, including the test period.

full rationale

The principal circular step is the temporal overlap between the contrastive-learning input x′ and the prediction labels. Equation (10) builds the future alpha embedding from 20 days of future market data, and the label formula for ∆t = 20 uses returns over exactly that same future window; the contrastive loss then backpropagates through the shared modules, so the model is fitted to the prediction-window data. Because the split description gives no exclusion of boundary samples, test-period information can enter training and early stopping, undermining the independent content of the SOTA claims in Tables 1 and 3. The rest of the derivation is not circular: the cascading residual decomposition of prior beta, hidden beta, and alpha is an architectural choice, and the hidden factors are learned latent representations whose predictive contribution is at least nominally tested out-of-sample. The self-citation to FactorVAE [10] is a baseline, not a load-bearing argument. The score reflects that one load-bearing empirical claim is compromised by construction, though it is not a definitional equivalence of the whole method.

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

The model's central claim depends on several hyperparameters whose values are not reported (M, tau, gamma, H), on the standard factor decomposition, and crucially on the unstated assumption that future data used in the contrastive loss does not cross the temporal split boundaries.

free parameters (5)
  • Number of hidden factors M
    The number of hidden factor prototypes is a hyperparameter; the paper reports performance for different M but does not state the final chosen M or how it was selected.
  • Temperature tau
    Temperature parameter in InfoNCE loss (Eq. 11), not specified in the main text.
  • Contrastive weight gamma
    Weight balancing MSE and contrastive loss in Eq. (13), not specified in the main text.
  • Hidden dimension H
    Dimension of stock feature embeddings and factor prototypes; the GRU hidden state size is not stated in the main text.
  • Prediction periods L
    Number of forward prediction periods in multi-label loss; the paper lists four periods but also backtests at delta t = 10, so the exact set used in Eq. (12) is ambiguous.
assumptions (4)
  • domain assumption Stock returns follow a factor structure decomposable into prior beta, hidden beta, and idiosyncratic alpha
    Eqs. (2)-(4) assume the standard factor model decomposition, which is a modeling choice, not derived.
  • standard math Hypergraph convolution from Feng et al. [13] correctly propagates factor information
    The message passing rule in Eq. (1) is taken as given from prior literature.
  • ad hoc to paper Future data windows used in contrastive loss are contained within the same temporal split
    Eqs. (9)-(10) use x' (next 20 days) for every training sample; no gapping or truncation is described, so this assumption is unstated and load-bearing.
  • domain assumption IC and ICIR are appropriate metrics for cross-sectional prediction in this setting
    The paper evaluates with information coefficient, a standard but not exhaustive measure of prediction quality.
invented entities (2)
  • Hidden factor prototypes c^(i)
    purpose: Learnable vectors used to generate soft hyperedges for hidden factors via similarity with residual embeddings.
    The prototypes are optimized end-to-end; no external evidence establishes they correspond to real economic factors.
  • Hidden factors (soft hyperedges)
    purpose: Data-driven factors that supplement the 83 industry factors; they are defined as Sigmoid dot products of residuals and prototypes.
    The paper only shows that removing the module degrades prediction; it does not examine the factors' content, stability, or economic meaning.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FactorGCL: A Hypergraph-Based Factor Model with Temporal Residual Contrastive Learning for Stock Returns Prediction." pith.science (2026). https://pith.science/paper/XQORNFDN

@misc{pith2026250205218,
  author       = {Pith},
  title        = {Pith review of: FactorGCL: A Hypergraph-Based Factor Model with Temporal Residual Contrastive Learning for Stock Returns Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XQORNFDN}},
  note         = {Machine review of arXiv:2502.05218}
}
read the original abstract

As a fundamental method in economics and finance, the factor model has been extensively utilized in quantitative investment. In recent years, there has been a paradigm shift from traditional linear models with expert-designed factors to more flexible nonlinear machine learning-based models with data-driven factors, aiming to enhance the effectiveness of these factor models. However, due to the low signal-to-noise ratio in market data, mining effective factors in data-driven models remains challenging. In this work, we propose a hypergraph-based factor model with temporal residual contrastive learning (FactorGCL) that employs a hypergraph structure to better capture high-order nonlinear relationships among stock returns and factors. To mine hidden factors that supplement human-designed prior factors for predicting stock returns, we design a cascading residual hypergraph architecture, in which the hidden factors are extracted from the residual information after removing the influence of prior factors. Additionally, we propose a temporal residual contrastive learning method to guide the extraction of effective and comprehensive hidden factors by contrasting stock-specific residual information over different time periods. Our extensive experiments on real stock market data demonstrate that FactorGCL not only outperforms existing state-of-the-art methods but also mines effective hidden factors for predicting stock returns.

Figures

Figures reproduced from arXiv: 2502.05218 by the authors.

Figure 1
Figure 1. Stock price trends vary across different sectors and [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Brief illustration of FactorGCL. Most current ML-based methods extract factors from mar￾ket data without effectively leveraging prior human experi￾ence, which may lead to overfitting of the extracted factors to market noise rather than capturing effective patterns. This limitation presents a significant obstacle to the application of ML in factor models. To address this obstacle, we propose a hypergraph-based factor… view at source ↗
Figure 3
Figure 3. Overview of the cascading residual hypergraph ar [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Illustration of the information propagation process [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Illustration of the temporal residual contrastive [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: The performances of FactorGCL with different [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

40 extracted references · 30 canonical work pages

  1. [1]

    Almeida, C.; and Freire, G. 2023. Which (Nonlinear) Factor Models? Available at SSRN 4421179

    work page 2023

  2. [2]

    Z.; and Koltun, V

    Bai, S.; Kolter, J. Z.; and Koltun, V . 2018. An empir- ical evaluation of generic convolutional and recurrent networks for sequence modeling. arXiv.arXiv preprint arXiv:1803.01271, 10

    arXiv 2018

  3. [3]

    Bai, S.; Zhang, F.; and Torr, P. H. 2021. Hypergraph convolution and hypergraph attention. Pattern Recog- nition, 110: 107637

    work page 2021

  4. [4]

    Bansal, R.; and Yaron, A. 2004. Risks for the long run: A potential resolution of asset pricing puzzles. The journal of Finance, 59(4): 1481–1509

    work page 2004

  5. [5]

    Caron, M.; Misra, I.; Mairal, J.; Goyal, P.; Bojanowski, P.; and Joulin, A. 2020. Unsupervised learning of vi- sual features by contrasting cluster assignments. Ad- vances in neural information processing systems , 33: 9912–9924

    work page 2020

  6. [6]

    Chen, X.; and He, K. 2021. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 15750–15758

    work page 2021

  7. [7]

    Cho, K.; Van Merri ¨enboer, B.; Gulcehre, C.; Bah- danau, D.; Bougares, F.; Schwenk, H.; and Bengio, Y . 2014. Learning phrase representations using RNN encoder-decoder for statistical machine translation. arXiv preprint arXiv:1406.1078

    arXiv 2014

  8. [8]

    Cui, C.; Li, X.; Zhang, C.; Guan, W.; and Wang, M. 2023. Temporal-relational hypergraph tri-attention networks for stock trend prediction. Pattern Recogni- tion, 143: 109759

    work page 2023

Show all 40 references
  1. [9]

    Daniel, K.; Hirshleifer, D.; and Sun, L. 2020. Short- and long-horizon behavioral factors. The review of fi- nancial studies, 33(4): 1673–1736

    2020

  2. [10]

    Duan, Y .; Wang, L.; Zhang, Q.; and Li, J. 2022. Fac- torvae: A probabilistic dynamic factor model based on variational autoencoder for predicting cross-sectional stock returns. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, 4468–4476

    2022

  3. [11]

    Eugene, F.; and French, K. 1992. The cross-section of expected stock returns. Journal of Finance, 47(2): 427–465

    1992

  4. [12]

    F.; and French, K

    Fama, E. F.; and French, K. R. 2021. Multifactor ex- planations of asset pricing anomalies . University of Chicago Press

    2021

  5. [13]

    Feng, Y .; You, H.; Zhang, Z.; Ji, R.; and Gao, Y . 2019. Hypergraph neural networks. In Proceedings of the AAAI conference on artificial intelligence, volume 33, 3558–3565

    2019

  6. [14]

    Gu, S.; Kelly, B.; and Xiu, D. 2021. Autoencoder asset pricing models. Journal of Econometrics, 222(1): 429– 450

    2021

  7. [15]

    Han, H.; Xie, L.; Chen, S.; and Xu, H. 2023. Stock trend prediction based on industry relationships driven hypergraph attention networks. Applied Intelligence, 53(23): 29448–29464

    2023

  8. [16]

    He, Z.; and Krishnamurthy, A. 2013. Intermediary as- set pricing. American Economic Review, 103(2): 732– 70

    2013

  9. [17]

    Hou, M.; Xu, C.; Liu, Y .; Liu, W.; Bian, J.; Wu, L.; Li, Z.; Chen, E.; and Liu, T.-Y . 2021. Stock trend predic- tion with multi-granularity data: A contrastive learning approach with adaptive fusion. In Proceedings of the 30th ACM International Conference on Information & Kn...

    2021

  10. [18]

    T.; Pruitt, S.; and Su, Y

    Kelly, B. T.; Pruitt, S.; and Su, Y . 2019. Characteristics are covariances: A unified model of risk and return. Journal of Financial Economics, 134(3): 501–524

    2019

  11. [19]

    Levin, A. 1995. Stock selection via nonlinear multi- factor models. Advances in Neural Information Pro- cessing Systems, 8

    1995

  12. [20]

    Li, J.; Zhou, P.; Xiong, C.; and Hoi, S. C. 2020. Proto- typical contrastive learning of unsupervised represen- tations. arXiv preprint arXiv:2005.04966

    2020 arXiv

  13. [21]

    Li, X.; Cui, C.; Cao, D.; Du, J.; and Zhang, C. 2022. Hypergraph-based reinforcement learning for stock portfolio selection. In ICASSP 2022-2022 IEEE Inter- national Conference on Acoustics, Speech and Signal Processing (ICASSP), 4028–4032. IEEE

    2022

  14. [22]

    Li, X.; Li, Z.; Shi, C.; Xu, Y .; Du, Q.; Tan, M.; Huang, J.; and Lin, W. 2024. AlphaFin: Benchmarking Finan- cial Analysis with Retrieval-Augmented Stock-Chain Framework. arXiv preprint arXiv:2403.12582

    2024 arXiv

  15. [23]

    Lin, S.; Liu, C.; Zhou, P.; Hu, Z.-Y .; Wang, S.; Zhao, R.; Zheng, Y .; Lin, L.; Xing, E.; and Liang, X. 2022. Prototypical graph contrastive learning. IEEE transac- tions on neural networks and learning systems , 35(2): 2747–2758

    2022

  16. [24]

    Lintner, J. 1975. The valuation of risk assets and the selection of risky investments in stock portfolios and capital budgets. In Stochastic optimization models in finance, 131–155. Elsevier

    1975

  17. [25]

    Liu, Y .; Zheng, Y .; Zhang, D.; Chen, H.; Peng, H.; and Pan, S. 2022. Towards unsupervised deep graph struc- ture learning. In Proceedings of the ACM Web Confer- ence 2022, 1392–1403

    2022

  18. [26]

    Oord, A. v. d.; Li, Y .; and Vinyals, O. 2018. Repre- sentation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748

    2018 arXiv

  19. [27]

    Qin, Y .; Song, D.; Chen, H.; Cheng, W.; Jiang, G.; and Cottrell, G. 2017. A dual-stage attention-based recur- rent neural network for time series prediction. arXiv preprint arXiv:1704.02971

    2017 arXiv

  20. [28]

    Sawhney, R.; Agarwal, S.; Wadhwa, A.; Derr, T.; and Shah, R. R. 2021. Stock selection via spatiotempo- ral hypergraph attention network: A learning to rank approach. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, 497–504

    2021

  21. [29]

    Sawhney, R.; Agarwal, S.; Wadhwa, A.; and Shah, R. R. 2020. Spatiotemporal hypergraph convolu- tion network for stock movement forecasting. In 2020 IEEE International Conference on Data Mining (ICDM), 482–491. IEEE

    2020

  22. [30]

    Sharpe, W. F. 1964. Capital asset prices: A theory of market equilibrium under conditions of risk. The jour- nal of finance, 19(3): 425–442

    1964

  23. [31]

    Su, H.; Wang, X.; Qin, Y .; and Chen, Q. 2024. Atten- tion based adaptive spatial–temporal hypergraph con- volutional networks for stock price trend prediction. Expert Systems with Applications, 238: 121899

    2024

  24. [32]

    Treynor, J. L. 1961. Toward a theory of market value of risky assets

    1961

  25. [33]

    Uddin, A.; and Yu, D. 2020. Latent factor model for asset pricing. Journal of Behavioral and Experimental Finance, 27: 100353

    2020

  26. [34]

    Vaswani, A. 2017. Attention is all you need. arXiv preprint arXiv:1706.03762

    2017 arXiv

  27. [35]

    Veli ˇckovi´c, P.; Cucurull, G.; Casanova, A.; Romero, A.; Lio, P.; and Bengio, Y . 2017. Graph attention net- works. arXiv preprint arXiv:1710.10903

    2017 arXiv

  28. [36]

    Xia, H.; Ao, H.; Li, L.; Liu, Y .; Liu, S.; Ye, G.; and Chai, H. 2024. CI-STHPAN: Pre-trained At- tention Network for Stock Selection with Channel- Independent Spatio-Temporal Hypergraph. In Pro- ceedings of the AAAI Conference on Artificial Intelli- gence, volume 38, 9187–9195

    2024

  29. [37]

    Xu, W.; Liu, W.; Wang, L.; Xia, Y .; Bian, J.; Yin, J.; and Liu, T.-Y . 2021. Hist: A graph-based framework for stock trend forecasting via mining concept-oriented shared information. arXiv preprint arXiv:2110.13716

    2021 arXiv

  30. [38]

    Zhang, L.; Aggarwal, C.; and Qi, G.-J. 2017. Stock price prediction via discovering multi-frequency trad- ing patterns. In Proceedings of the 23rd ACM SIGKDD international conference on knowledge discovery and data mining, 2141–2149

    2017

  31. [39]

    Zhou, D.; Huang, J.; and Sch ¨olkopf, B. 2006. Learning with hypergraphs: Clustering, classification, and em- bedding. Advances in neural information processing systems, 19

    2006

  32. [40]

    Zou, J.; Cao, H.; Liu, L.; Lin, Y .; Abbasnejad, E.; and Shi, J. Q. 2022. Astock: A new dataset and automated stock trading based on stock-specific news analyzing model. arXiv preprint arXiv:2206.06606

    2022 arXiv

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.