NGAT: A Node-level Graph Attention Network for Long-term Stock Prediction

Mingchuan Zhao; Ruihai Dong; Valerio Poti; Yingjie Niu

REVIEW 4 major objections 5 minor 1 cited by

NGAT: A Node-level Graph Attention Network for Long-term Stock Prediction

T0 review · 4 major / 5 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 NGAT gives every company its own attention vector and reports better long-term return and volatility forecasts than shared-attention graph models on two datasets.

desk verdict Worth refereeing, but the missing split protocol and the malformed attention equation need to be fixed before the headline empirical gains can be trusted. read the letter →

arxiv 2507.02018 v1 pith:PTU7EKBX submitted 2025-07-02 q-fin.ST cs.AIcs.LG

Yingjie Niu , Mingchuan Zhao , Valerio Poti , Ruihai Dong This is my paper

classification q-fin.STcs.AIcs.LG

keywords GraphNeuralNetworksStockpredictionNode-levelattentionLong-termforecastingVolatilityCorporaterelationshipMomentumspillover

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

This paper argues that the usual next-day stock trend setup hides what graphs know, and replaces it with a long-horizon task: predict whether a stock's average return over the next $T$ days ($T\in\{1,5,10,21\}$) will be above or below its previous $T$-day average, and also predict the realized volatility of the next $T$ days. On top of LSTM sequential embeddings, the paper's NGAT layer gives each company its own attention vector instead of sharing one attention mechanism across all nodes, on the grounds that companies are heterogeneous in ways homogeneous graph benchmarks are not. The paper reports that NGAT outperforms LSTM, GCN, GAT, LSTM+GCN, TGC, and AD-GAT across horizons and tasks on the SPNews and ACL2018 datasets, with the largest gains over GAT, and that all models exceed 70% accuracy on long-horizon return classification. It also uses the same models on different graph constructions to argue that downstream task performance alone is an unreliable judge of graph quality.

What carries the argument

The carrying object is the node-level graph attention layer. For each node $i$ it computes masked coefficients $$\beta_{ij} = \frac{\exp(\hat{R}_{(i,j)}\,\mathrm{LeakyReLU}(a_i^\top [W_q h_i \| W_k h_j]))}{\sum_{m\in N_i} \exp(\hat{R}_{(i,m)}\,\mathrm{LeakyReLU}(a_i^\top [W_q h_i \| W_k h_m]))}$$ with node-specific attention parameters $a_i\in\mathbb{R}^{2F'\times F'}$, relation weights $\hat{R}$ from temporally decayed news or tweet co-occurrence, and multi-head aggregation. The output concatenates the transformed self embedding $W_v h_i$ with the neighbor aggregation, preserving node identity against over-smoothing. This layer is what makes the model portable: the rest of the pipeline (LSTM encoder, output layer) is unchanged from the baselines, so performance differences are attributed to the attention mechanism.

What would settle it

Run the released code's data splitter and check whether any day in a training sample's feature or target window also appears in the validation or test window of the same stock. If it does, re-run the comparison with a strictly chronological split; the central claim that node-level attention beats all baselines collapses if the advantage does not survive.

Watch

Extended reading notes

Core claim

NGAT assigns each node its own trainable attention mechanism $a_i$, so the weight a stock places on each neighbor is specific to that stock. The paper claims this node-level attention is why a graph model can finally beat a standalone LSTM across every horizon in both return classification and volatility regression, and it reports consistent NGAT wins in volatility forecasting on both datasets. A second claim is that the long-term task itself matters: all tested models exceed 70% accuracy when forecasting $T=1$ to $T=21$ average-return directions, far above the 50-60% typical of next-day prediction, which the paper reads as evidence that multi-period momentum spillovers are learnable from relational graphs. A third claim is cautionary: because NGAT's adaptive attention can compensate for noisy edges, comparing graph-construction methods by final prediction accuracy can be misleading.

Load-bearing premise

The reported numbers all assume the data split is temporal and leak-free, but the paper never says how validation and test days are separated from training days, so a random split that lets the model see a stock's future returns during training would inflate every metric.

Editorial extensions

If this is right

Any graph-based stock pipeline can swap in the NGAT layer without redesigning its LSTM encoder or output head, since NGAT is presented as a portable graph convolution layer.
At $T=21$ the model gives monthly return-direction and volatility estimates, which can feed directly into portfolio construction and risk budgeting.
The four return scenarios (surge, rebound, pullback, plunge) turn the classification output into a rule-based trading signal: hold or increase for $(LP, N+)$, sell or short for $(LN, N-)$.
Graph comparisons should not be judged by downstream accuracy alone, because a more adaptive model can hide graph construction defects.
Long-horizon forecasting lifts all baselines above 70% accuracy in the paper's runs, so multi-period targets deserve attention in future work.

Reading between the lines

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

Node-specific attention is a general recipe for relational domains with heterogeneous nodes, such as supply chains, interbank networks, or protein interaction graphs, because the same argument about non-identical node distributions applies beyond equities.
The graph-comparison finding implies that two graph constructions with equal downstream accuracy can still differ structurally; a natural test is to rank graphs by the stability of learned attention weights or by edge-ablation sensitivity.
A testable extension is to use NGAT's volatility forecast as a risk scaler for return-trend bets and measure whether the surge or plunge signal set beats an unconditional benchmark after transaction costs.
The paper's horizon pattern (return accuracy falls while volatility accuracy rises with $T$) suggests choosing $T$ per stock or sector rather than fixing one global horizon.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Here's the short version. The genuinely new bits are the long-term return/volatility task (forecasting average return and realized volatility over T=5/10/21 days) and node-specific attention, where each company gets its own attention mechanism instead of a shared one. The paper also makes a reasonable point that downstream task performance is not a clean way to compare graph constructions, and it demonstrates that by running both GAT and NGAT under static, correlation, and co-occurrence graphs. That graph-comparison discussion is the most thoughtful part.

The paper does several things right. The task reformulation is sensible: next-day prediction is indeed too short for momentum spillover, and including volatility as a target is useful. The empirical scope is broad—two datasets, six baselines, four horizons, classification and regression, plus scenario-level breakdowns. And there is a GitHub link, which is more than many papers in this area offer.

The soft spots are not small. First, Eq. 7 is not well-defined: with a_i in R^{2F' x F'}, the product a_i^T [W_q h_i || W_k h_j] is an F'-vector, but the softmax needs a scalar per edge. This looks like a typo—standard GAT uses a 2F'-dimensional vector—but as printed it makes the architecture impossible to implement from the text alone. Second, the train/validation/test split is never specified. The training setup section says only that all parameters are selected via grid search on validation performance. For daily stock data with labels computed over the next T days, a random split would let training labels overlap test-window returns, inflating every reported number. This is a load-bearing omission, not a stylistic one. Third, the claim in Section 6 that all models exceed 70% accuracy is factually wrong for several ACL2018 entries (e.g., GAT r-mean21 at 0.6789, LSTM at 0.6963). Finally, five-run averages without variance or significance tests weaken the 'significant improvements' language.

My overall read: the empirical claim that NGAT outperforms GAT and other baselines could be true, but the paper as printed does not give enough to check it. The split issue is the one that matters most. If the authors confirm a strict chronological split and fix Eq. 7, this becomes a solid applied paper.

Recommendation: send it to peer review rather than desk reject. The long-term task and the node-level attention idea are worth a careful look, and the load-bearing concerns are exactly what referees are for. The referee report should insist on the split protocol, an equation fix, and corrected accuracy statements.

Referee Report

4 major / 5 minor

Summary. This paper proposes a node-level graph attention network (NGAT) for long-horizon stock trend and volatility prediction. The authors define a classification target comparing future T-day average returns with the prior T-day average, and a regression target for realized volatility. Corporate relationship graphs are built from news/tweet co-occurrence with a temporal memory window. The NGAT layer assigns each node its own attention matrix, aggregates neighbor embeddings, and combines the result with a transformed self-embedding. Experiments on ACL2018 and SPNews compare NGAT against LSTM, GCN, GAT, LSTM+GCN, TGC, and AD-GAT across horizons T=1, 5, 10, 21, and also compare graph construction methods. The main claimed results are consistent improvements over baselines and a limitation of model-based graph comparison.

Significance. The long-horizon task formulation is a useful extension of the usual next-day prediction setup, and the idea of node-specific attention is interesting and potentially portable. The paper ships a public GitHub implementation and uses public datasets, which is a strength for reproducibility. If the architecture is corrected, the data split is made leakage-free, and the evaluation is statistically grounded, the empirical comparison would be a valuable data point for finance-GNN research. As it stands, however, the central empirical claim is not yet supported because the architecture is under-specified and the split protocol is absent.

major comments (4)

[Section 4.2, Eq. (7)] The attention parameter a_i is defined as a matrix a_i ∈ R^{2F'×F'}, but the expression a_i^T [W_q h_i || W_k h_j] is then a length-F' vector, so β_ij as written is a vector rather than a scalar attention coefficient. This makes the layer ill-defined. The standard GAT formulation uses a vector attention parameter, and the authors should either define a_i ∈ R^{2F'} or specify a reduction from the F'-dimensional vector to a scalar (for example, a learned linear map). The GitHub implementation may resolve the ambiguity, but the paper should be self-contained.
[Section 5.2] The paper states that all parameters are selected by grid search on validation performance, but it never defines the temporal split between training, validation, and test sets. Because the labels in Eqs. (1)-(2) use returns from days d+1 to d+T, a random split over days or stock-day pairs would allow future return information from test windows to leak into training, artificially inflating all reported metrics. The authors must specify a strict chronological split with a gap of at least T days between the last training label window and the first validation/test feature window, and confirm that hyperparameter selection is confined to the validation span.
[Section 6, Tables 1 and 3] All metrics are reported as five-run averages without standard deviations, confidence intervals, or significance tests, yet the text uses the phrase 'significant improvements' (for example, relative gains over GAT for T=21 on ACL2018 and T=10 on SPNews). As reported, these differences could be within run-to-run noise. Please provide per-seed results, standard deviations, and a paired significance test, or clearly state that the reported gains are not statistically tested.
[Section 6] The claim that 'all models exceed 70% accuracy' is contradicted by Table 1: for ACL2018 r-mean21, GAT, LSTM, and TGC report accuracies of 0.6789, 0.6963, and 0.6971 respectively. This factual error should be corrected and raises a need to audit the surrounding discussion for similar overstatements.

minor comments (5)

[Section 5.2] The hyperparameter 'relation threshold ∈ [0,1]' is listed, but the text does not explain how this threshold is applied to the aggregated relation attributes in Eq. (4). Please clarify the binarization or thresholding procedure.
[Table 2] Table 2 is difficult to parse: the header rows and entries are interleaved (for example, 'LSTM LN LP LN LP LN LP' followed by a sequence of 'N+' values). Please reformat with separate columns per scenario and horizon, or use a clearer layout.
[Equation (12)] Equation (12) has a mismatched parenthesis in the denominator, and the confusion matrix is written in a nonstandard order; please correct the formula.
[Equation (7)] The term 'LeakeyReLU' is a typo for 'LeakyReLU'.
[Section 4.2] The self-loop attribute R(s_i,s_i)_d is defined as the maximum relation over neighbors; the text should state explicitly whether this self-loop is included in the neighborhood N_i used in the normalization in Eq. (7), as this affects the attention weights.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular reasoning found: prediction targets are defined independently of the model and no fitted parameter is renamed as a prediction.

full rationale

The paper's derivation chain is self-contained. The classification label in Eq. (1) and the volatility regression target in Eq. (2) are computed directly from future log returns and are not functions of model parameters, so the empirical evaluation is not self-referential. The proposed NGAT layer is an architectural modification of GAT with node-specific attention vectors (Eqs. 5-7), trained with standard BCE/MSE losses; no parameter fitted to a subset of the target data is later reported as a prediction. The only self-citation is Reference [23] as the source of the SPNews dataset, but that is a data citation, not a load-bearing theoretical premise, and the datasets are public benchmarks also used for the ACL2018 comparisons. Statements such as 'all models exceed 70% accuracy' are factually inconsistent with Table 1 for some ACL2018 r-mean21 entries, and the omitted train/validation/test split raises a potential data-leakage concern, but those are correctness and experimental-validity issues, not circular reasoning. No step reduces to its own input by construction, so the circularity score is 0.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The empirical claim rests on standard supervised learning, but the paper never specifies the temporal split, so leakage is an unstated risk. All learned weights and six hyperparameters are fit on training or validation data; the node-level attention design is posited, not derived. There are no invented physical entities.

free parameters (6)

Per-node attention matrix a_i = learned; shape N x 2F' x F'
Each node receives its own attention parameters (Section 4.2); this is the central architectural change and a main source of model capacity.
Hidden dimension F' = selected from {16,32,64}
Chosen by grid search on validation performance (Section 5.2).
Attention heads M = selected from {1,4,8}
Grid-searched on validation performance (Section 5.2).
Memory window delta = optimal 5 from {1,2,3,5,7,10}
Controls temporal decay of edge weights in Eq. 4 and is tuned on validation (Section 5.2).
Relation threshold = not reported, from {0,1}
Grid-searched threshold for edges; the final value is not stated (Section 5.2).
Forecast horizon T = evaluated at {1,5,10,21}
Task-defining horizon; T=21 is highlighted as best balance after inspecting results (Sections 3 and 6).

assumptions (5)

domain assumption News and tweet co-occurrence defines economically meaningful inter-firm edges.
Section 5.1 constructs edges from S x S cliques per article; no independent validation that co-occurrence reflects momentum spillover, and Table 4 only evaluates graphs via downstream performance.
domain assumption Labels computed from future T-day returns (Eq. 1) are used without leakage.
No train/test split is specified, so temporal leakage would make all reported metrics optimistic; this assumption is load-bearing.
domain assumption The LSTM last hidden state is a sufficient representation of price and volume history.
Section 4.1 fixes the sequence encoder to LSTM and uses only the final time step; other encoders or pooling strategies are not tested.
domain assumption Message passing over corporate relationship graphs can capture momentum spillover.
The paper assumes inter-firm relations propagate return information, following related work; no causal or statistical test supports this beyond predictive accuracy.
domain assumption Stock returns and volatility are predictable from historical prices and relational data in a stationary way.
The supervised setup assumes the train/test relationship is stable; the paper does not discuss regime shifts, transaction costs, or changing market structure.

how reviews work

0 comments

Cite this review

Pith. "Pith review of NGAT: A Node-level Graph Attention Network for Long-term Stock Prediction." pith.science (2026). https://pith.science/paper/PTU7EKBX

@misc{pith2026250702018,
  author       = {Pith},
  title        = {Pith review of: NGAT: A Node-level Graph Attention Network for Long-term Stock Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PTU7EKBX}},
  note         = {Machine review of arXiv:2507.02018}
}

read the original abstract

Graph representation learning methods have been widely adopted in financial applications to enhance company representations by leveraging inter-firm relationships. However, current approaches face three key challenges: (1) The advantages of relational information are obscured by limitations in downstream task designs; (2) Existing graph models specifically designed for stock prediction often suffer from excessive complexity and poor generalization; (3) Experience-based construction of corporate relationship graphs lacks effective comparison of different graph structures. To address these limitations, we propose a long-term stock prediction task and develop a Node-level Graph Attention Network (NGAT) specifically tailored for corporate relationship graphs. Furthermore, we experimentally demonstrate the limitations of existing graph comparison methods based on model downstream task performance. Experimental results across two datasets consistently demonstrate the effectiveness of our proposed task and model. The project is publicly available on GitHub to encourage reproducibility and future research.

Figures

Figures reproduced from arXiv: 2507.02018 by the authors.

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

Relational Probing: LM-to-Graph Adaptation for Financial Prediction
cs.CL 2026-04 unverdicted novelty 6.0 of 10

Relational Probing replaces the LM output head with a trainable relation head that induces graphs from hidden states and optimizes them end-to-end for stock trend prediction, showing gains over co-occurrence baselines.

Reference graph

Works this paper leans on

32 extracted references · 26 canonical work pages · cited by 1 Pith paper

[1]

The Journal of Finance and Data Science4(2), 120–137 (2018)

Atkins, A., Niranjan, M., Gerding, E.: Financial news predicts stock market volatil- ity better than close price. The Journal of Finance and Data Science4(2), 120–137 (2018)

work page 2018
[2]

arXiv preprint arXiv:1707.03815 (2017)

Bojchevski, A., Günnemann, S.: Deep gaussian embedding of graphs: Unsupervised inductive learning via ranking. arXiv preprint arXiv:1707.03815 (2017)

arXiv 2017
[3]

Journal of Computational Science 2(1), 1–8 (2011)

Bollen, J., Mao, H., Zeng, X.: Twitter mood predicts the stock market. Journal of Computational Science 2(1), 1–8 (2011)

work page 2011
[4]

The Journal of finance 52(1), 57–82 (1997)

Carhart, M.M.: On persistence in mutual fund performance. The Journal of finance 52(1), 57–82 (1997)

work page 1997
[5]

In: Proceedings of the 27th ACM international conference on information and knowledge management

Chen, Y., Wei, Z., Huang, X.: Incorporating corporation relationship via graph convolutional neural networks for stock price prediction. In: Proceedings of the 27th ACM international conference on information and knowledge management. pp. 1655–1658 (2018)

work page 2018
[6]

Cheng, R., Li, Q.: Modeling the momentum spillover effect for stock prediction via attribute-drivengraphattentionnetworks.In:ProceedingsoftheAAAIConference on artificial intelligence. vol. 35, pp. 55–62 (2021)

work page 2021
[7]

arXiv preprint arXiv:1406.1078 (2014)

Cho, K., Van Merriënboer, B., Gulcehre, C., Bahdanau, D., Bougares, F., Schwenk, H., Bengio, Y.: Learning phrase representations using rnn encoder-decoder for sta- tistical machine translation. arXiv preprint arXiv:1406.1078 (2014)

arXiv 2014
[8]

Journal of Financial Econometrics7(2), 174–196 (2009)

Corsi, F.: A simple approximate long-memory model of realized volatility. Journal of Financial Econometrics7(2), 174–196 (2009)

work page 2009

Show all 32 references

[9]

Risks9(2) (2021)

Di Persio, L., Garbelli, M., Wallbaum, K.: Forward-looking volatility estimation for risk-managed investment strategies during the covid-19 crisis. Risks9(2) (2021)

2021
[10]

The Journal of Finance48(5), 1749–1778 (1993)

Engle, R.F., Ng, V.K.: Measuring and testing the impact of news on volatility. The Journal of Finance48(5), 1749–1778 (1993)

1993
[11]

Journal of financial economics 116(1), 1–22 (2015)

Fama, E.F., French, K.R.: A five-factor asset pricing model. Journal of financial economics 116(1), 1–22 (2015)

2015
[12]

ACM Transactions on Information Systems (TOIS) 37(2), 1–30 (2019)

Feng, F., He, X., Wang, X., Luo, C., Liu, Y., Chua, T.S.: Temporal relational ranking for stock prediction. ACM Transactions on Information Systems (TOIS) 37(2), 1–30 (2019)

2019
[13]

In: Proceedings of the thirteenth international conference on ar- tificial intelligence and statistics

Glorot, X., Bengio, Y.: Understanding the difficulty of training deep feedforward neural networks. In: Proceedings of the thirteenth international conference on ar- tificial intelligence and statistics. pp. 249–256. JMLR Workshop and Conference Proceedings (2010)

2010
[14]

Neural Computing and Applications32, 9713–9729 (2020)

Jin,Z.,Yang,Y.,Liu,Y.:Stockclosingpricepredictionbasedonsentimentanalysis and lstm. Neural Computing and Applications32, 9713–9729 (2020)

2020
[15]

Expert Systems with Applications 103, 25–37 (2018) 12 Yingjie Niu , Mingchuan Zhao, Valerio Poti, and Ruihai Dong

Kim, H.Y., Won, C.H.: Forecasting the volatility of stock price index: A hybrid model integrating lstm with multiple garch-type models. Expert Systems with Applications 103, 25–37 (2018) 12 Yingjie Niu , Mingchuan Zhao, Valerio Poti, and Ruihai Dong

2018
[16]

arXiv preprint arXiv:1412.6980 (2014)

Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)

2014 arXiv
[17]

arXiv preprint arXiv:1609.02907 (2016)

Kipf, T.N., Welling, M.: Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907 (2016)

2016 arXiv
[18]

ACM Transactions on Information Systems (TOIS) 34(2), 1–30 (2016)

Li, Q., Chen, Y., Jiang, L.L., Li, P., Chen, H.: A tensor-based information frame- work for predicting the stock market. ACM Transactions on Information Systems (TOIS) 34(2), 1–30 (2016)

2016
[19]

IEEE Transactions on Knowledge and Data Engineering 33(10), 3323–3337 (2020)

Li, Q., Tan, J., Wang, J., Chen, H.: A multimodal event-driven lstm model for stock prediction using online news. IEEE Transactions on Knowledge and Data Engineering 33(10), 3323–3337 (2020)

2020
[20]

Li, W., Bao, R., Harimoto, K., Chen, D., Xu, J., Su, Q.: Modeling the stock relation withgraphnetworkforovernightstockmovementprediction.In:Proceedingsofthe twenty-ninthinternationalconferenceoninternationaljointconferencesonartificial intelligence. pp. 4541–4547 (2021)

2021
[21]

Lo, A.W., MacKinlay, A.C.: When are contrarian profits due to stock market over- reaction? The Review of Financial Studies3(2), 175–205 (1990)

1990
[22]

Econometrica 59(2), 347–370 (1991)

Nelson, D.B.: Conditional heteroskedasticity in asset returns: A new approach. Econometrica 59(2), 347–370 (1991)

1991
[23]

In: Proceedings of the 5th ACM Interna- tional Conference on AI in Finance

Niu, Y., Lu, L., Dolphin, R., Poti, V., Dong, R.: Evaluating financial relational graphs: Interpretation before prediction. In: Proceedings of the 5th ACM Interna- tional Conference on AI in Finance. pp. 564–572 (2024)

2024
[24]

Journal of financial economics22(1), 27–59 (1988)

Poterba, J.M., Summers, L.H.: Mean reversion in stock prices: Evidence and im- plications. Journal of financial economics22(1), 27–59 (1988)

1988
[25]

In: 2017 inter- national conference on advances in computing, communications and informatics (icacci)

Selvin, S., Vinayakumar, R., Gopalakrishnan, E., Menon, V.K., Soman, K.: Stock price prediction using lstm, rnn and cnn-sliding window model. In: 2017 inter- national conference on advances in computing, communications and informatics (icacci). pp. 1643–1647. IEEE (2017)

2017
[26]

In: 2022 IEEE International Conference on Big Data (Big Data)

Soun, Y., Yoo, J., Cho, M., Jeon, J., Kang, U.: Accurate stock movement prediction with self-supervised learning from sparse noisy tweets. In: 2022 IEEE International Conference on Big Data (Big Data). pp. 1691–1700. IEEE (2022)

2022
[27]

In: 2020 2nd novel intelligent and leading emerging sciences conference (NILES)

Sunny, M.A.I., Maswood, M.M.S., Alharbi, A.G.: Deep learning-based stock price prediction using lstm and bi-directional lstm model. In: 2020 2nd novel intelligent and leading emerging sciences conference (NILES). pp. 87–92. IEEE (2020)

2020
[28]

arXiv preprint arXiv:1710.10903 (2017)

Veličković, P., Cucurull, G., Casanova, A., Romero, A., Lio, P., Bengio, Y.: Graph attention networks. arXiv preprint arXiv:1710.10903 (2017)

2017 arXiv
[29]

In: Proceedings of the 31st ACM international conference on information & knowledge management

Xiang, S., Cheng, D., Shang, C., Zhang, Y., Liang, Y.: Temporal and heterogeneous graph neural network for financial time series prediction. In: Proceedings of the 31st ACM international conference on information & knowledge management. pp. 3584–3593 (2022)

2022
[30]

Information Sciences607, 783–798 (2022)

Xu, C., Huang, H., Ying, X., Gao, J., Li, Z., Zhang, P., Xiao, J., Zhang, J., Luo, J.: Hgnn: Hierarchical graph neural network for predicting the classification of price-limit-hitting stocks. Information Sciences607, 783–798 (2022)

2022
[31]

In: Gurevych, I., Miyao, Y

Xu, Y., Cohen, S.B.: Stock movement prediction from tweets and historical prices. In: Gurevych, I., Miyao, Y. (eds.) Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 1970–

1970
[1979]

Association for Computational Linguistics, Melbourne, Australia (2018)

2018

Pith tools

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

[1] [1]

The Journal of Finance and Data Science4(2), 120–137 (2018)

Atkins, A., Niranjan, M., Gerding, E.: Financial news predicts stock market volatil- ity better than close price. The Journal of Finance and Data Science4(2), 120–137 (2018)

work page 2018

[2] [2]

arXiv preprint arXiv:1707.03815 (2017)

Bojchevski, A., Günnemann, S.: Deep gaussian embedding of graphs: Unsupervised inductive learning via ranking. arXiv preprint arXiv:1707.03815 (2017)

arXiv 2017

[3] [3]

Journal of Computational Science 2(1), 1–8 (2011)

Bollen, J., Mao, H., Zeng, X.: Twitter mood predicts the stock market. Journal of Computational Science 2(1), 1–8 (2011)

work page 2011

[4] [4]

The Journal of finance 52(1), 57–82 (1997)

Carhart, M.M.: On persistence in mutual fund performance. The Journal of finance 52(1), 57–82 (1997)

work page 1997

[5] [5]

In: Proceedings of the 27th ACM international conference on information and knowledge management

Chen, Y., Wei, Z., Huang, X.: Incorporating corporation relationship via graph convolutional neural networks for stock price prediction. In: Proceedings of the 27th ACM international conference on information and knowledge management. pp. 1655–1658 (2018)

work page 2018

[6] [6]

Cheng, R., Li, Q.: Modeling the momentum spillover effect for stock prediction via attribute-drivengraphattentionnetworks.In:ProceedingsoftheAAAIConference on artificial intelligence. vol. 35, pp. 55–62 (2021)

work page 2021

[7] [7]

arXiv preprint arXiv:1406.1078 (2014)

Cho, K., Van Merriënboer, B., Gulcehre, C., Bahdanau, D., Bougares, F., Schwenk, H., Bengio, Y.: Learning phrase representations using rnn encoder-decoder for sta- tistical machine translation. arXiv preprint arXiv:1406.1078 (2014)

arXiv 2014

[8] [8]

Journal of Financial Econometrics7(2), 174–196 (2009)

Corsi, F.: A simple approximate long-memory model of realized volatility. Journal of Financial Econometrics7(2), 174–196 (2009)

work page 2009