Paper: arXiv 2507.02018

Authors: Yingjie Niu, Mingchuan Zhao, Valerio Poti, Ruihai Dong

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.

Complexity vs Empirical Score

  • Math Complexity: 7.5/10
  • Empirical Rigor: 7.0/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper introduces a specialized Node-level Graph Attention Network (NGAT) with node-specific attention mechanisms, involving advanced graph theory and deep learning concepts, which drives a high math score. It demonstrates empirical rigor through multi-dataset experimentation, public code release on GitHub, and statistical evaluation of model performance, though it lacks the extreme data/implementation intensity of full-scale backtesting.

Research Flowchart

  flowchart TD
  A["Research Goal: Predict long-term stock price movements using corporate relationship graphs"] --> B
  
  subgraph B ["Methodology: Node-level Graph Attention Network NGAT"]
      B1["Node-level Attention"]
      B2["Graph Structure Learning"]
      B3["Long-term Prediction Task"]
  end
  
  C["Input Data: Corporate Relationship Graphs<br>(2 Datasets with Inter-firm Relations)"] --> D
  
  subgraph D ["Computational Process"]
      D1["Graph Embedding via NGAT"]
      D2["Temporal Pattern Extraction"]
      D3["Long-horizon Forecasting"]
  end
  
  B --> D
  D --> E
  
  subgraph E ["Key Findings & Outcomes"]
      E1["NGAT Outperforms Baseline Models"]
      E2["Long-term Task Reveals Relational Benefits"]
      E3["Limitations of Existing Graph Comparison Methods Exposed"]
      E4["Public GitHub Repository for Reproducibility"]
  end
  
  style A fill:#e1f5e1,stroke:#2e7d32
  style E fill:#fff3e0,stroke:#e65100

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes a long‑term stock prediction task and a Node‑level Graph Attention Network (NGAT) designed for corporate relationship graphs.

Source sentence
we propose a long‑term stock prediction task and develop a Node-level Graph Attention Network (NGAT) specifically tailored for corporate relationship graphs.

Assumes 2

  • Each company node has a distinct feature distribution, so a shared attention mechanism is unsuitable. model
    Source sentence
    assuming identical distributions across nodes becomes untenable.
  • The asset universe is fixed, making higher computational complexity acceptable. execution
    Source sentence
    NGAT incurs higher computational complexity, this remains acceptable under fixed asset universe settings common in institutional portfolios.

Data

  • SPNews Dataset · publicly available text-news public
    S&P500 companies
    Source sentence
    The SPNews Dataset [23] includes news and transaction data for S&P500 companies, with daily bulletins listing the latest eight news items per company.
  • ACL2018 Dataset · publicly available text-news public
    companies with ticker tags in tweets
    Source sentence
    The ACL2018 Dataset [31] contains tweets and transaction data spanning two years. Text data is derived from Twitter posts, where edges between companies are defined by the co‑occurrence of ticker tags.

Recognised sources: Twitter / X

Costs

Not applicable The paper reports no trading or execution result.

Finds 2

  • NGAT achieves higher return classification accuracy than baselines, with relative gains up to 8.12% and all models exceeding 70% accuracy. performance
    Source sentence
    NGAT achieves significant improvements over GAT, with relative gains of 4.03%, 5.90%, and 2.07% for T = 21 on ACL2018, and 2.08%, 8.12%, and 1.41% for T = 10 on SPNews.
  • NGAT consistently outperforms all baselines in volatility forecasting across both datasets. performance
    Source sentence
    NGAT consistently outperforms all baselines in volatility forecasting across both datasets.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 33 related papers). Coverage is partial, so this is neither replication nor refutation.

Cited by 1 paper in the archive (1 by independent authors): Retrieval-Augmented Diffusion Modeling for Stochastic … (used as a baseline)

Code

https://github.com/FreddieNIU/NGAT — the authors' own repository

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.

Cited References (40)

  1. Accurate stock movement prediction with self-supervised learning from sparse noisy tweets (2022)
  2. Stock Movement Prediction from Tweets and Historical Prices (2018)
  3. Efficient capital markets (1970)
  4. A tensor-based information framework for predicting the stock market (2016)
  5. Incorporating corporation relationship via graph convolutional neural networks for stock price prediction (2018)
  6. Temporal relational ranking for stock prediction (2019)
  7. A multimodal event-driven LSTM model for stock prediction using online news (2020)
  8. Long short-term memory (1997)
and 32 more
  1. Learning phrase representations using RNN encoder-decoder for statistical machine translation (2014) — link
  2. Modeling the momentum spillover effect for stock prediction via attribute-driven graph attention networks (2021)
  3. Temporal and heterogeneous graph neural network for financial time series prediction — → 2305.08740
  4. Pitfalls of graph neural network evaluation. arXiv 2018 (2018) — link
  5. Graph attention networks (2017) — link
  6. Adam: A method for stochastic optimization (2014) — link
  7. Understanding the difficulty of training deep feedforward neural networks (2010)
  8. Semi-supervised classification with graph convolutional networks (2016) — link
  9. Measuring and Testing the Impact of News on Volatility (1993)
  10. Modelling the persistence of conditional variances (1986)
  11. Conditional Heteroskedasticity in Asset Returns: A New Approach (1991)
  12. All in the family Nesting symmetric and asymmetric GARCH models (1995)
  13. A Simple Approximate Long-Memory Model of Realized Volatility (2009)
  14. Forecasting the volatility of stock price index: A hybrid model integrating LSTM with multiple GARCH-type models (2018)
  15. Stock price prediction using LSTM, RNN and CNN-sliding window model (2017)
  16. Deep learning-based stock price prediction using LSTM and bi-directional LSTM model (2020)
  17. Stock closing price prediction based on sentiment analysis and LSTM (2020)
  18. Modeling the stock relation with graph network for overnight stock movement prediction (2021)
  19. HGNN: Hierarchical graph neural network for predicting the classification of price-limit-hitting stocks (2022)
  20. Giving Content to Investor Sentiment: The Role of Media in the Stock Market (2007)
  21. Twitter mood predicts the stock market (2011)
  22. Financial news predicts stock market volatility better than close price (2018)
  23. Deep gaussian embedding of graphs: Unsupervised inductive learning via ranking (2017) — link
  24. Learning target-specific representations of financial news documents for cumulative abnormal return prediction (2018)
  25. Common risk factors in the returns on stocks and bonds (1993)
  26. A five-factor asset pricing model (2015)
  27. On persistence in mutual fund performance (1997)
  28. When are Contrarian Profits Due to Stock Market Overreaction? (1990)
  29. Mean reversion in stock prices: Evidence and implications (1988)
  30. Value versus growth: The international evidence (1998)
  31. Unit root tests in panel data: asymptotic and finite-sample properties (2002)
  32. Evaluating Financial Relational Graphs: Interpretation Before Prediction — → 2410.07216

2 of 40 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.