MDGNN: Multi-Relational Dynamic Graph Neural Network for Comprehensive and Dynamic Stock Investment Prediction

Hao QianHongting ZhouQian ZhaoHao ChenHongxiang YaoJingwei WangZiqi LiuFei YuZhiqiang ZhangJun Zhou

article2024AAAI61 citations

Proposes a dynamic graph neural network framework that integrates diverse daily market relationships across industries, investment banks, and co-holding connections with a Transformer architecture to improve stock trend prediction.

Listen

Predicting stock price movements is a critical yet complex task in financial markets due to the dynamic and interconnected nature of market entities, macroeconomic trends, and corporate disclosures. Existing machine learning approaches often treat individual stocks in isolation or rely on static, single-relation network structures, failing to reflect how multi-entity relationships evolve over time. The main objective of the article is to present and evaluate a new framework, the Multi-Relational Dynamic Graph Neural Network, which models evolving daily interconnections among stocks, investment banks, and industry sectors to improve stock investment prediction.

The researchers evaluated their framework using real-world data from China's stock market across the CSI 100 and CSI 300 indices over a three-year evaluation window from January 2020 through February 2023. The approach constructs daily network snapshots connecting stocks to associated industries and investment banks using market trading indicators and text from financial news and analyst reports. A hierarchical graph layer aggregates these multi-faceted relationships, while a sequence-modeling module captures their temporal evolution over rolling lookback windows. The framework was benchmarked against ten baseline methods ranging from standard time-series models to dynamic graph architectures.

The findings show that the proposed framework consistently outperforms all baseline models across multiple performance and return metrics. On the broader CSI 300 dataset, the model achieved a cumulative return of 0.9828, more than doubling the performance of competitive dynamic baseline models. The model also achieved superior ranking accuracy and precision among top-ranked stocks, with the performance gains becoming more pronounced on larger datasets with denser relational connections. Ablation analyses confirmed that modeling multiple relationship pathways provided the single largest performance gain, with investment bank connections exerting a stronger predictive influence than industry connections alone.

These results demonstrate that incorporating multifaceted, evolving institutional relationships significantly enhances predictive performance and risk-adjusted return generation compared to traditional forecasting tools. By propagating information across shared ownership and industry links, quantitative investment strategies can better anticipate sector-wide movements and reduce prediction errors. However, decision-makers should note that the empirical evaluation was conducted exclusively on Chinese market indices under specific lookback windows, and adding excessive network complexity beyond optimal thresholds can lead to performance degradation. As a next step, the article suggests integrating contrastive learning techniques to further refine the framework's representations.

Alibaba Group Ant Group capital investment graph neural networks feedforward neural networks multi-relational dynamic graph neural network dynamic stock investment prediction stock trend forecasting

arXiv: 2402.06633

Paper: EvolveGCN: Evolving Graph Convolutional Networks for Dynamic Graphs, A. Pareja et al. (2019). It provides the foundational framework for evolving graph convolutional networks over time series, which MDGNN builds upon to capture daily evolving relational snapshots among market entities.
Paper: Heterogeneous Graph Transformer, Ziniu Hu et al. (2020). It introduces transformer-based message passing for heterogeneous graphs with multiple node and relationship types, establishing the multi-relational aggregation mechanisms used in financial networks.
Paper: Modeling Relational Data with Graph Convolutional Networks, Michael Schlichtkrull et al. (2018). It formulates relational graph convolutional networks for multi-relational knowledge graphs, defining the core principles of relation-specific message passing adapted by MDGNN.
Paper: Connecting the Dots: Multivariate Time Series Forecasting with Graph Neural Networks, Zonghan Wu et al. (2020). It establishes multivariate time series forecasting methods that combine temporal convolutions with graph neural networks to model inter-series dependencies.
Paper: Graph Transformer Networks, Seongjun Yun et al. (2019). It develops graph transformer architectures for learning multi-hop composite relations across heterogeneous networks without manual meta-path engineering.
Paper: FactorVAE: A Probabilistic Dynamic Factor Model Based on Variational Autoencoder for Predicting Cross-Sectional Stock Returns, Yitong Duan et al. (2022). It formalizes deep probabilistic factor modeling and cross-sectional return prediction on Chinese stock market indices, setting the financial forecasting setup used by the source.
Paper: Dynamic Graph Neural Networks Under Spatio-Temporal Distribution Shift, Zeyang Zhang et al. (2022). It analyzes dynamic graph neural network design under spatio-temporal distribution shifts, providing essential techniques for robust temporal modeling across changing market regimes.
Paper: Semi-Supervised Classification with Graph Convolutional Networks, Thomas N. Kipf et al. (2017). It introduces standard localized spectral graph convolutional networks, which serve as the fundamental graph layer for neighborhood information aggregation.

No sufficiently relevant recommendations were found.

Cover for MDGNN: Multi-Relational Dynamic Graph Neural Network for Comprehensive and Dynamic Stock Investment Prediction

Abstract

The stock market is a crucial component of the financial system, but predicting the movement of stock prices is challenging due to the dynamic and intricate relations arising from various aspects such as economic indicators, financial reports, global news, and investor sentiment. Traditional sequential methods and graph-based models have been applied in stock movement prediction, but they have limitations in capturing the multifaceted and temporal influences in stock price movements. To address these challenges, the Multi-relational Dynamic Graph Neural Network (MDGNN) framework is proposed, which utilizes a discrete dynamic graph to comprehensively capture multifaceted relations among stocks and their evolution over time. The representation generated from the graph offers a complete perspective on the interrelationships among stocks and associated entities. Additionally, the power of the Transformer structure is leveraged to encode the temporal evolution of multiplex relations, providing a dynamic and effective approach to predicting stock investment. Further, our proposed MDGNN framework achieves the best performance in public datasets compared with state-of-the-art (SOTA) stock investment methods.

Topics

capital investment graph neural networks feedforward neural networks

Keywords

multi-relational dynamic graph neural network dynamic stock investment prediction stock trend forecasting hierarchical multi-relational graph adaptive temporal-relational modeling graph attention intra-day graph

Introduction
Related Work
Preliminary
Algorithm Design
Intra-day Graph Snapshot
Inter-day Temporal Extraction Layer
Prediction Layer
Experiments
Experiment Setup
Experiment Result
Ablation Study
Case Study
Hyperparameter Study
Conclusion
References

Knowls

Knowl 1 — MDGNN architecture for dynamic stock prediction
model/method
MDGNN is a discrete dynamic graph neural network for stock investment prediction. For each trading day, it constructs a multi-relational graph containing stock, investment-bank, and industry entities; applies a hierarchical graph embedding module to obtain each stock’s daily representation; and applies a causal Transformer over a look-back sequence of daily representations to encode evolving stock and relation patterns. A prediction layer then converts the temporally enriched representation of each stock into a score interpreted as the probability of a positive return. The framework is designed to model both multifaceted relations within a day and temporal evolution across days.
Knowl 2 — Dynamic-graph stock prediction formulation
definition
Let a dynamic graph be a sequence $\mathcal{G}=\{G_1,G_2,\ldots,G_T\}$ of daily multi-relational snapshots, where $G_t=(V_t,E_t,R_t)$ , $V_t$ is the set of entities on trading day $t$ , $E_t$ is the set of edges, and $R_t$ specifies their relation types. For stock node $v_i^t\in V_t$ , let $p_{i,t}$ be its closing price and let $b_t$ be the benchmark-index return on day $t$ . The target is the stock’s excess return
$y_{i,t}=\frac{p_{i,t+1}-p_{i,t}}{p_{i,t}}-b_t.$
MDGNN treats prediction as node regression: it learns a scoring function $f(\mathcal{G};\Theta)$ with trainable parameters $\Theta$ by minimizing
$\mathcal{L}=\sum_{n\in\mathcal{N}_{\mathrm{train}}}\ell\!\left(Y_n,f(\mathcal{G}_n;\Theta)\right),$
where $\mathcal{N}_{\mathrm{train}}$ is the set of training samples, $Y_n$ is the ground-truth label or label vector for sample $n$ , and $\ell$ is the per-sample loss.
Knowl 3 — Daily multi-relational stock-market graph construction
model/method
MDGNN constructs a separate graph snapshot for each trading day from daily trading data and textual sources such as macroeconomic reports, financial news, financial statements, and investment-bank research reports. Financial lexicons and syntactic extraction methods are used to identify relation edges and their features. The graph contains stock, industry, and investment-bank nodes and includes several relation families: stock–stock relations based on sector, ownership, and co-holding; stock–bank relations representing buying, selling, research, and advisory activity; stock–industry relations encoding supply, demand, competition, and regulatory connections; and industry–industry relations. Because the data and extracted relations are rebuilt over time, the resulting graph sequence represents both changing stock attributes and changing inter-entity relationships.
Knowl 4 — Hierarchical multi-relational graph embedding
model/method
For each daily graph snapshot, MDGNN first propagates information along stock-centered meta-paths, including Stock–Stock ( $SS$ ), Stock–Bank–Stock ( $SBS$ ), and Stock–Industry–Industry–Stock ( $SIIS$ ). The node-update attention uses both neighboring-node features and edge features. For target node $i$ , neighbor $j\in\mathcal{N}_i$ , node representations $h_i,h_j$ , edge representation $e_{ij}$ , trainable projection $W$ , trainable attention vector $a$ , and concatenation operator $\Vert$ , the attention score and normalized weight are
$\beta_{ij}=a^{\mathsf T}[Wh_i\Vert Wh_j\Vert We_{ij}],\qquad \alpha_{ij}^{(k)}=\frac{\exp(\operatorname{LeakyReLU}(\beta_{ij}))}{\sum_{j'\in\mathcal{N}_i}\exp(\operatorname{LeakyReLU}(\beta_{ij'}))},$
where $k\in\{1,\ldots,K\}$ indexes the $K$ attention heads. The head-aggregated representation is
$h_i=\sigma\!\left(\frac{1}{K}\sum_{k=1}^{K}\sum_{j\in\mathcal{N}_i}\alpha_{ij}^{(k)}W_kh_j\right),$
where $W_k$ is the head-specific projection and $\sigma$ is an activation function. Let $h_i^{(1)},h_i^{(2)},h_i^{(3)}$ be the representations produced by the $SS$ , $SBS$ , and $SIIS$ meta-paths. MDGNN adaptively fuses them as
$h_{v_i}=\sigma\!\left(\sum_{r=1}^{3}\operatorname{Softmax}_{r}\!\left(Wh_i^{(r)}\right)h_i^{(r)}\right),$
where $\operatorname{Softmax}_{r}$ produces normalized attention coefficients over the three meta-paths. Stacking $L$ such graph layers allows successive layers to incorporate increasingly distant relational information; the final-layer representation of stock $v$ on day $t$ is denoted $h_{v,t}$ .
Knowl 5 — Causal Transformer for evolving graph representations
model/method
For a stock $v$ on trading day $t$ , MDGNN forms the sequence $H_{v,t-\delta:t}=\{h_{v,t'}\mid t-\delta\le t'\le t\}$ from the current and preceding $\delta$ trading days, where $\delta$ is the temporal window and each $h_{v,t'}$ is the daily graph embedding. Trainable matrices $W_Q,W_K,W_V$ transform this sequence into queries, keys, and values:
$Q=W_QH_{v,t-\delta:t},\qquad K=W_KH_{v,t-\delta:t},\qquad V=W_VH_{v,t-\delta:t}.$
The temporal attention output is
$Z=\operatorname{softmax}\!\left(\frac{QK^{\mathsf T}}{\sqrt{d}}+mP+M\right)V,$
where $d$ is the key/query dimensionality, $P$ is the fixed ALiBi relative-position bias, $m$ is its slope parameter, and $M$ is a causal forward mask that prevents a time position from attending to subsequent positions. The ALiBi term penalizes attention between more distant query–key positions, thereby favoring recent events without introducing learned positional embeddings. The current-day output vector is denoted $z_{v,t}$ .
Knowl 6 — Stock-return prediction layer
equation
For stock $v$ on day $t$ , MDGNN maps the current-day temporal representation $z_{v,t}\in\mathbb{R}^{d}$ to a positive-return probability with
$\hat{y}_{v,t}=\sigma\!\left(W_1z_{v,t}+b_1\right),$
where $W_1\in\mathbb{R}^{1\times d}$ and $b_1\in\mathbb{R}$ are trainable parameters and $\sigma$ is the sigmoid function. The scalar $\hat{y}_{v,t}$ is used as the stock-ranking score and estimates whether the stock will yield a positive return on trading day $t$ .
Knowl 7 — Chinese-market datasets and rolling backtest
experimental setup
The evaluation uses Chinese CSI100 and CSI300 constituent-stock datasets. Each stock has 42 normalized features: 25 market-performance features, including opening price, closing price, percentage change, volatility, and turnover rate; 12 valuation features, including P/E, P/B, and P/S ratios; 4 categorical company features; and 1 institutional-consensus-expectation feature. The graph statistics are:
Could not parse LaTeX table
The reported backtest spans 01/01/2020 to 02/31/2023. Training is repeated every six months, producing seven models: each model uses the preceding six months for training, the preceding month for validation, and fixed parameters for prediction during the following six months. Evaluation uses Information Coefficient (IC), Information Ratio (IR), cumulative return (CR), and Precision@30. IC measures ranking quality, IR is excess portfolio return divided by tracking error, CR is accumulated portfolio return, and Precision@30 measures whether the top 30 predicted stocks outperform the benchmark. All models use the same splits, are trained for up to 500 epochs with early stopping, and are implemented in PyTorch on an Nvidia V100 GPU. MDGNN uses hidden size 128, two graph layers, and a temporal window of 10 trading days.
Knowl 8 — Benchmark performance on CSI100 and CSI300
empirical result
Across both datasets and all four metrics, MDGNN achieved higher reported mean performance than every compared baseline, including MLP, LSTM, Transformer, GAT, GCN, RGCN, HAN, HGT, EvolveGCN, and HTGNN. The comparison below gives MDGNN and the strongest baseline value for each metric; the strongest baseline can differ across metrics.
Could not parse LaTeX table
The advantage is especially pronounced on CSI300 for IC and cumulative return. The authors attribute the stronger performance on the larger graph to the additional institutional and industry nodes, which provide more paths for information propagation. The results also indicate that heterogeneous and temporal graph information is more effective than using only intrinsic stock features, temporal features, or a static homogeneous graph.
Knowl 9 — Component ablation of MDGNN
empirical result
On CSI300, removing any of MDGNN’s principal components reduced performance. The full model and four variants are:
Could not parse LaTeX table
The largest degradation occurs when the meta-path module is removed, supporting the contribution of explicitly modeling multiple relation paths. Removing temporal extraction also substantially lowers cumulative return, while removing edge weights or hierarchical aggregation produces smaller but consistent losses.
Knowl 10 — Contribution of individual relation families
empirical result
A CSI300 relation ablation evaluates stock–stock ( $SS$ ), stock–bank ( $SB$ ), stock–industry ( $SI$ ), and industry–industry ( $II$ ) edges. Connections between different node types are bidirectional by default; removing an edge family also removes any meta-path that requires it. The reported results are:
Could not parse LaTeX table
Adding stock–bank edges to stock–stock edges improves all metrics more than adding stock–industry edges alone. Combining stock, bank, and industry relations yields a further improvement, and the complete relation set performs best, supporting the use of investment-bank and industry information as complementary signals to direct stock–stock relations.

Coverage note — The qualitative Chengdu Bank case study and the window-size/GNN-depth sensitivity plots were omitted because they are supporting analyses rather than load-bearing components of the proposed model or its main quantitative validation.

References

1.Chen, Q.; and Robert, C. Y. 2021a. Graph-Based Learning for Stock Movement Prediction with Textual and Relational Data. In The Journal of Financial Data Science.
2.Chen, Q.; and Robert, C.-Y. 2021b. Graph-based learning for stock movement prediction with textual and relational data. arXiv preprint arXiv:2107.10941.
3.Chung, J.; Gulcehre, C.; Cho, K.; and Bengio, Y. 2014. Empirical Evaluation of Gated Recurrent Neural Networks on Sequence Modeling. arXiv:1412.3555.
4.Devlin, J.; Chang, M.-W.; Lee, K.; and Toutanova, K. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. 4171–4186.
5.Fan, Y.; Ju, M.; Zhang, C.; Zhao, L.; and Ye, Y. 2021. Heterogeneous Temporal Graph Neural Network. arXiv:2110.13889.
6.Feng, F.; Chen, H.; He, X.; Ding, J.; Sun, M.; and Chua, T.-S. 2019a. Enhancing Stock Movement Prediction with Adversarial Training. In Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence, IJCAI-19, 5843–5849. International Joint Conferences on Artificial Intelligence Organization.
7.Feng, F.; He, X.; Wang, X.; Luo, C.; Liu, Y.; and Chua, T.-S. 2019b. Temporal relational ranking for stock prediction. ACM Transactions on Information Systems (TOIS), 37(2): 1–30.
8.Gu, S.; Kelly, B.; and Xiu, D. 2020. Empirical asset pricing via machine learning. The Review of Financial Studies, 33(5): 2223–2273.
9.Han, Y.; Kim, J.; and Enke, D. 2023. A machine learning trading system for the stock market based on N-period Min-Max labeling using XGBoost. Expert Systems with Applications, 211: 118581.
10.Hochreiter, S.; and Schmidhuber, J. 1997. Long Short-Term Memory. Neural Computation, 9(8): 1735–1780.
11.Hu, Z.; Dong, Y.; Wang, K.; and Sun, Y. 2020. Heterogeneous Graph Transformer. In Proceedings of The Web Conference 2020, WWW ’20, 2704–2710. New York, NY, USA: Association for Computing Machinery. ISBN 9781450370233.
12.Jain, L. C.; and Medsker, L. R. 1999. Recurrent Neural Networks: Design and Applications. USA: CRC Press, Inc., 1st edition. ISBN 0849371813.
13.Jarvelin, K.; and Kekäläinen, J. 2000. IR Evaluation Methods for Retrieving Highly Relevant Documents. In Proceedings of the 23rd Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR ’00, 41–48. New York, NY, USA: Association for Computing Machinery. ISBN 1581132263.
14.Kipf, T. N.; and Welling, M. 2017. Semi-Supervised Classification with Graph Convolutional Networks. In 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net.
15.Li, Z.; Yang, D.; Zhao, L.; Bian, J.; Qin, T.; and Liu, T.-Y. 2019. Individualized Indicator for All: Stock-Wise Technical Indicator Optimization with Stock Embedding. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, KDD ’19, 894–902. New York, NY, USA: Association for Computing Machinery. ISBN 9781450362016.
16.Lin, H.; Zhou, D.; Liu, W.; and Bian, J. 2021. Learning Multiple Stock Trading Patterns with Temporal Routing Adaptor and Optimal Transport. In Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining, KDD ’21, 1017–1026. New York, NY, USA: Association for Computing Machinery. ISBN 9781450383325.
17.Nagel, S. 2021. Machine learning in asset pricing, volume 8. Princeton University Press.
18.Nelson, D. M.; Pereira, A. C.; and De Oliveira, R. A. 2017. Stock market’s price movement prediction with LSTM neural networks. In 2017 International joint conference on neural networks (IJCNN), 1419–1426. Ieee.
19.Pareja, A.; Domeniconi, G.; Chen, J.; Ma, T.; Suzumura, T.; Kanezashi, H.; Kaler, T.; Schardl, T.; and Leiserson, C. 2020. EvolveGCN: Evolving Graph Convolutional Networks for Dynamic Graphs. Proceedings of the AAAI Conference on Artificial Intelligence, 34(04): 5363–5370.
20.Press, O.; Smith, N. A.; and Lewis, M. 2022. Train Short, Test Long: Attention with Linear Biases Enables Input Length Extrapolation. arXiv:2108.12409.
21.Roy, S. S.; Mittal, D.; Basu, A.; and Abraham, A. 2015. Stock market forecasting using LASSO linear regression model. In Afro-European Conference for Industrial Advancement: Proceedings of the First International Afro-European Conference for Industrial Advancement AECIA 2014, 371–381. Springer.
22.Sawhney, R.; Agarwal, S.; Wadhwa, A.; Derr, T.; and Shah, R. R. 2021. Stock Selection via Spatiotemporal Hypergraph Attention Network: A Learning to Rank Approach. Proceedings of the AAAI Conference on Artificial Intelligence, 35(1): 497–504.
23.Schlichtkrull, M.; Kipf, T. N.; Bloem, P.; van den Berg, R.; Titov, I.; and Welling, M. 2017. Modeling Relational Data with Graph Convolutional Networks. arXiv:1703.06103.
24.Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones, L.; Gomez, A. N.; Kaiser, L. u.; and Polosukhin, I. 2017. Attention is All you Need. In Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc.
25.Velicković, P.; Cucurull, G.; Casanova, A.; Romero, A.; Liò, P.; and Bengio, Y. 2018. Graph Attention Networks. International Conference on Learning Representations.
26.Wang, H.; Li, S.; Wang, T.; and Zheng, J. 2021a. Hierarchical Adaptive Temporal-Relational Modeling for Stock Trend Prediction. In Zhou, Z.-H., ed., Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, IJCAI-21, 3691–3698. International Joint Conferences on Artificial Intelligence Organization. Main Track.
27.Wang, H.; Li, S.; Wang, T.; and Zheng, J. 2021b. Hierarchical Adaptive Temporal-Relational Modeling for Stock Trend Prediction. In IJCAI, 3691–3698.
28.Wang, H.; Wang, T.; Li, S.; Zheng, J.; Guan, S.; and Chen, W. 2022. Adaptive Long-Short Pattern Transformer for Stock Investment Selection. In Raedt, L. D., ed., Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, IJCAI-22, 3970–3977. International Joint Conferences on Artificial Intelligence Organization. Main Track.
29.Wang, H.; Wang, T.; and Li, Y. 2020. Incorporating Expert-Based Investment Opinion Signals in Stock Prediction: A Deep Learning Framework. Proceedings of the AAAI Conference on Artificial Intelligence, 34(01): 971–978.
30.Xu, W.; Liu, W.; Wang, L.; Xia, Y.; Bian, J.; Yin, J.; and Liu, T.-Y. 2021a. HIST: A Graph-based Framework for Stock Trend Forecasting via Mining Concept-Oriented Shared Information. arXiv preprint arXiv:2110.13716.
31.Xu, W.; Liu, W.; Wang, L.; Xia, Y.; Bian, J.; Yin, J.; and Liu, T.-Y. 2021b. Hist: A graph-based framework for stock trend forecasting via mining concept-oriented shared information. arXiv preprint arXiv:2110.13716.
32.Zhang, L.; Aggarwal, C.; and Qi, G.-J. 2017. Stock Price Prediction via Discovering Multi-Frequency Trading Patterns. In Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD ’17, 2141–2149. New York, NY, USA: Association for Computing Machinery. ISBN 9781450348874.

Citation

MLA

Qian, H., et al. “MDGNN: Multi-Relational Dynamic Graph Neural Network for Comprehensive and Dynamic Stock Investment Prediction”. arXiv, 2024, http://arxiv.org/abs/2402.06633v1.

APA

Qian, H., Zhou, H., Zhao, Q., Chen, H., Yao, H., Wang, J., Liu, Z., Yu, F., Zhang, Z., & Zhou, J. (2024). MDGNN: Multi-Relational Dynamic Graph Neural Network for Comprehensive and Dynamic Stock Investment Prediction. arXiv. http://arxiv.org/abs/2402.06633v1

Chicago

Qian, H., H. Zhou, Q. Zhao, et al. 2024. “MDGNN: Multi-Relational Dynamic Graph Neural Network for Comprehensive and Dynamic Stock Investment Prediction”. arXiv. http://arxiv.org/abs/2402.06633v1.

Harvard

Qian, H. et al. (2024) “MDGNN: Multi-Relational Dynamic Graph Neural Network for Comprehensive and Dynamic Stock Investment Prediction”, arXiv [Preprint]. Available at: http://arxiv.org/abs/2402.06633v1.

Vancouver

1. Qian H, Zhou H, Zhao Q, Chen H, Yao H, Wang J, Liu Z, Yu F, Zhang Z, Zhou J (2024) MDGNN: Multi-Relational Dynamic Graph Neural Network for Comprehensive and Dynamic Stock Investment Prediction. arXiv

BibTeX

@article{qian2024mdgnn,
  title = {MDGNN: Multi-Relational Dynamic Graph Neural Network for Comprehensive and Dynamic Stock Investment Prediction},
  author = {Qian, Hao and Zhou, Hongting and Zhao, Qian and Chen, Hao and Yao, Hongxiang and Wang, Jingwei and Liu, Ziqi and Yu, Fei and Zhang, Zhiqiang and Zhou, Jun},
  year = {2024},
  journal = {arXiv},
  url = {http://arxiv.org/abs/2402.06633v1},
  eprint = {2402.06633}
}

Metadata:arXiv

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF

MDGNN: Multi-Relational Dynamic Graph Neural Network for Comprehensive and Dynamic Stock Investment Prediction

Abstract

Topics

Keywords

Table of Contents

Knowls

Knowl 1 — MDGNN architecture for dynamic stock prediction

Knowl 2 — Dynamic-graph stock prediction formulation

Knowl 3 — Daily multi-relational stock-market graph construction

Knowl 4 — Hierarchical multi-relational graph embedding

Knowl 5 — Causal Transformer for evolving graph representations

Knowl 6 — Stock-return prediction layer

Knowl 7 — Chinese-market datasets and rolling backtest

Knowl 8 — Benchmark performance on CSI100 and CSI300

Knowl 9 — Component ablation of MDGNN

Knowl 10 — Contribution of individual relation families

References

Citation

BibTeX

Access the Paper