Scholarlink

<p>Stock price forecasting is a key task in the financial technology (FinTech) sector, as it supports informed investment decision-making and enhances risk management strategies. However, accurate forecasting of stock prices remains highly challenging due to the unpredictable, noisy, and ever-changing nature of financial markets, which are influenced by complex macroeconomic factors and rapid regime changes. In recent years, graph neural networks (GNNs) have been shown to be useful algorithms for modeling and forecasting complex stock interactions. Nonetheless, current GNN-based models primarily ignore the dynamic nature of financial markets, frequently relying on static adjacency matrices and not taking into consideration regime-driven movements such as bull, bear, and sideways markets. Hence, in this study, a Hidden Markov Model (HMM) assisted regime-aware learnable graph neural networks (RL-GNN) is proposed that explicitly models market regime dynamics through static graphs and is enriched with learnable graph correlations via dynamic graphs. Initially, the HMM module helps to determine underlying market regimes and builds regime-specific stock static correlation graphs that depict bull, bear, and sideways-based stock correlations. Additionally, unlike existing works that use fixed graph structures, the proposed framework incorporates a learnable graph correlation layer, allowing the model to refine stock relationships during training adaptively. By fusing regime-based static graphs with a trainable dynamic graph structure learning, the proposed RL-GNN captures both macro regime shifts and micro relational adjustments, leading to more accurate stock forecasts. An experimental study on the S&amp;P 500, FTSE 100, and SZSE-500 demonstrates that the proposed RL-GNN outperforms state-of-the-art approaches in terms of obtaining the lowest errors in MAPE and RMSPE, which are around 1.884% and 2.569%, respectively.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Regime-aware static and dynamic graph structure learning for stock price forecasting

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月25日

查看全文
  • Fizza Bukhari,
  • Amna Sarwar,
  • Asma Sattar,
  • Zahoor ur Rehman,
  • Sungwoo Park,
  • Seungmin Rho

摘要

Stock price forecasting is a key task in the financial technology (FinTech) sector, as it supports informed investment decision-making and enhances risk management strategies. However, accurate forecasting of stock prices remains highly challenging due to the unpredictable, noisy, and ever-changing nature of financial markets, which are influenced by complex macroeconomic factors and rapid regime changes. In recent years, graph neural networks (GNNs) have been shown to be useful algorithms for modeling and forecasting complex stock interactions. Nonetheless, current GNN-based models primarily ignore the dynamic nature of financial markets, frequently relying on static adjacency matrices and not taking into consideration regime-driven movements such as bull, bear, and sideways markets. Hence, in this study, a Hidden Markov Model (HMM) assisted regime-aware learnable graph neural networks (RL-GNN) is proposed that explicitly models market regime dynamics through static graphs and is enriched with learnable graph correlations via dynamic graphs. Initially, the HMM module helps to determine underlying market regimes and builds regime-specific stock static correlation graphs that depict bull, bear, and sideways-based stock correlations. Additionally, unlike existing works that use fixed graph structures, the proposed framework incorporates a learnable graph correlation layer, allowing the model to refine stock relationships during training adaptively. By fusing regime-based static graphs with a trainable dynamic graph structure learning, the proposed RL-GNN captures both macro regime shifts and micro relational adjustments, leading to more accurate stock forecasts. An experimental study on the S&P 500, FTSE 100, and SZSE-500 demonstrates that the proposed RL-GNN outperforms state-of-the-art approaches in terms of obtaining the lowest errors in MAPE and RMSPE, which are around 1.884% and 2.569%, respectively.