统计与数据科学学院知识库

随着金融科技的快速发展，金融市场数据呈现出前所未有的增长，高频交易、智能投研、量化交易系统及数字化金融基础设施的完善，使金融数据高度多源异构，涵盖历史交易数据、空间关联数据和新闻文本数据等结构化与非结构化信息。传统预测方法主要依赖时序数据，忽视了市场内部的联动关系与外部文本情绪影响，导致预测精度和稳定性受限。股票市场作为复杂的非线性系统，具有强时空特性和情绪驱动特征，单一数据源难以全面刻画市场动态。近年来，深度学习在金融时序分析预测、风险管理和投资决策中展现出卓越能力，尤其在“人工智能+金融”政策推动下，结合多源异构数据进行股票类资产价格预测，为精准投资决策和风险预警提供了新机遇。基于此，本文围绕深度学习驱动的多源异构数据融合预测框架，研究历史交易数据、空间关联数据和新闻文本数据的特征提取与融合策略，以多维度刻画市场动态。多源异构数据融合不仅弥补了单一数据源的信息缺失，提升模型的适应性和鲁棒性，同时依托深度学习强大的特征提取能力，挖掘数据间的复杂交互，为市场趋势预测提供更科学的决策依据。

在深入理解股票市场多源信息结构的基础上，基于不同的数据可得性与应用场景，本文构建了三种不同市场环境情形下的股票类资产价格预测模型。首先，在指数类资产难以依赖外部特征且股票市场具有多重周期性与非线性特征的情形下，提出自加权集成门控循环单元模型，结合变分模态分解对股指序列进行分解，并通过自加权集成损失函数优化子序列预测误差，以提升股指序列预测精度。其次，在股票市场内部股票空间联动效应显著的情形下，采用超图注意力网络与循环神经网络深度嵌合，实现股票资产空间关联信息与时序信息的实时融合，从而更准确地刻画股票市场内部的运行状态，满足股票资产的预测需求。最后，在宏观经济新闻、公司公告和政策事件对投资者情绪产生显著影响的情形下，提出自注意力残差门控循环单元模型，充分挖掘新闻文本与历史交易数据的交互特性，提高短期及事件驱动型预测的精准度。上述预测情形并非相互孤立，通过这种逐步扩展的策略，既能形成清晰的技术演进脉络，也能在实践中根据数据可得性和预测需求灵活选择合适的方法。

为成功构建并应用上述三种预测情形，本文在展开融合预测前进行了关键性准备工作，即完成了股票关联网络构建与中文金融大语言模型训练，分别从空间关联与文本情绪两个维度夯实多源异构数据融合的基础。通过股票张量分解及图与超图稀疏化处理，构建出更具解释力的股票关联网络，为后续时空特征的实时融合提供高质量网络输入；同时，依托大规模中文金融文本语料，训练并微调面向金融领域的中文大语言模型，使新闻舆情与政策解读信息得以精准量化，为后续文本情绪与历史交易数据的深度融合提供有力支撑。本文具体工作如下：

（1）针对现有股票关联网络构建中单一数据源的局限性和高阶关联的缺失问题，本文提出了基于张量分解的相似度矩阵构建方法以及图邻接矩阵和超图关联矩阵的稀疏化策略。通过整合历史交易数据与基本面数据，挖掘其互补信息，提升了对股票资产复杂空间关系的建模能力，并捕捉了多个股票资产之间的高阶交互关系。以沪深300指数部分成分股为例，构建的股票关联网络展现了显著的拓扑特征：①强关联性与高频互动，有效表征股票间复杂空间关系；②关键节点在网络中的重要作用，提升信息传递效率；③高效的信息传播能力，支持股票市场动态变化的捕捉与分析。

（2）针对传统文本模型在金融新闻领域存在的语境理解局限性，以及中文金融大语言模型的资源稀缺的现状，本文设计并训练了一款专门面向中文金融新闻文本情绪分类任务的大语言模型。通过系统分析三大中文金融文本数据库，构建了适用于遮蔽语言模型、下一句预测及情绪分类训练任务的训练样本集，基于BERT（Bidirectional Encoder Representations from Transformers）框架的预训练机制与微调原理，将领域自适应微调技术应用于金融大语言模型训练，最终构建出适配中国金融市场特性的金融大语言模型。该模型为后续中文股票新闻文本情绪特征提取和语义理解研究提供了可靠的模型支持。

（3）针对在基于历史交易数据的股指预测中现有分解集成方法未能有效处理子序列预测损失差异的问题，本文提出了一种基于变分模态分解与自加权集成门控循环单元（Gated Recurrent Unit, GRU）的股指时序预测模型。该模型首先对股指时间序列进行变分模态分解，得到多个子序列，然后利用GRU网络对每个子序列进行独立预测。同时，模型采用自加权损失函数自适应地调整各子序列的预测损失权重，从而优化模型整体预测性能。实验结果表明：①自加权集成损失函数显著提升了模型整体预测效果；②通过选择最优的超参数值，模型能够在不同数据集上实现最佳预测性能，验证了该方法的强适应性。该模型为股指时间序列预测提供了新思路，并为高精度预测模型的开发奠定了坚实基础。

（4）针对现有股票价格预测模型中，时序与空间特征采用串联架构难以实时融合、易破坏时间连续性的问题，本文提出了一种基于股票时空数据实时融合的超图注意力门控循环单元预测模型。该模型基于前文构建的股票关联网络，利用超图注意力网络提取空间特征，并将其嵌入到门控循环单元网络，实现了时空特征的实时融合，提升了模型预测精度。实验结果表明：①与传统串联架构相比，本文设计的嵌入式时空特征融合架构显著提高了模型预测精度；②该模型在预测精度上优于传统计量、机器学习和深度学习模型，展示了其在时空预测中的应用潜力；③与基准模型对比，该模型在长期预测中表现出更高的稳定性。该模型架构具有较高的应用价值，且可推广至其他时空预测领域。

（5）针对现有股票新闻情绪特征和时序特征融合预测模型中特征交互效率低和特征融合不充分的问题，本文提出了一种基于文本和历史交易数据融合的自注意力残差门控循环单元预测模型。该模型通过自注意力机制有效融合股票新闻情绪特征和时序特征，并通过残差连接将融合后的特征与原特征结合，输入门控循环单元提取时序特征后输出预测值。实验结果表明：①该模型在新闻文本情绪和时序特征融合预测方面展现了显著的有效性和长期预测能力，提供了有益的借鉴；②间接证明本文训练的中文金融大语言模型具备较强的股票新闻情感分类能力，具有较高的实际应用价值。

该研究不仅从方法论层面探讨了深度学习驱动的多源异构数据融合技术，而且通过实证分析验证了其在股票类资产价格预测中的良好成效，为股票市场的智能化预测、风险管理及量化交易等方面提供了有力支撑。研究结果表明，这种融合技术能够显著提升预测的准确性、稳定性和适应性，为今后智能金融研究与投资策略的优化带来了重要参考。

With the rapid development of financial technology, financial market data have experienced unprecedented growth. The improvement of high-frequency trading, intelligent investment research, quantitative trading systems, and digital financial infrastructure has made financial data highly multi-sourced and heterogeneous, encompassing both structured and unstructured information such as historical transaction data, spatial correlation data, and news text data. Traditional forecasting methods rely predominantly on time-series data while overlooking the internal linkage effects of markets and external sentiment impacts from textual information, leading to constraints in prediction accuracy and stability. As a complex nonlinear system, the stock market is characterized by pronounced spatiotemporal attributes and sentiment-driven features, making it difficult for a single data source to comprehensively capture market dynamics. In recent years, deep learning has demonstrated exceptional capabilities in financial time-series forecasting, risk management, and investment decision-making. Especially under the impetus of the “AI + Finance” policy, integrating multi-sourced heterogeneous data to predict the prices of stock-like assets has created new opportunities for precise investment decisions and risk warnings. In view of this, this dissertation centers on a deep learning-driven multi-sourced heterogeneous data fusion forecasting framework, examining feature extraction and fusion strategies for historical transaction data, spatial correlation data, and news text data to represent market dynamics from multiple dimensions. Such multi-sourced heterogeneous data fusion not only compensates for information gaps inherent in single-source data and enhances model adaptability and robustness, but also leverages the powerful feature extraction capabilities of deep learning to uncover complex interactions among different data sources, thus providing a more scientific basis for market trend forecasting.

Building on a thorough understanding of the multi-sourced information structure in the stock market, and taking into account varying data availability and application scenarios, this dissertation constructs three prediction models for stock-like asset prices under different market conditions. First, in situations where index-based assets cannot rely significantly on external features and where the stock market exhibits multiple cyclical and nonlinear characteristics, a self-weighted ensemble gated recurrent unit model is proposed. This model uses variational mode decomposition to decompose stock index sequences, then optimizes sub-sequence prediction errors via a self-weighted ensemble loss function, thereby improving the predictive accuracy of stock index series. Second, under conditions where strong spatial linkage effects exist among stocks within the stock market, a deep integration of a hypergraph attention network and a recurrent neural network is employed. This approach achieves real-time fusion of spatial correlation information and time-series data, thereby providing a more accurate depiction of internal market states and meeting the predictive requirements for stock assets. Finally, in cases where macroeconomic news, corporate announcements, and policy events have a significant influence on investor sentiment, this dissertation proposes a self-attention residual gated recurrent unit model. By fully exploring the interactions between news text data and historical transaction data, the model enhances accuracy in short-term and event-driven forecasts. These predictive scenarios are not mutually exclusive. Through a strategy of progressive expansion, the methodology outlines a clear technological evolution path while allowing for flexible selection of suitable methods in practice based on data availability and specific predictive needs.

In order to effectively construct and apply the above three predictive scenarios, this dissertation carries out critical preparatory work before conducting the integrated forecasts—namely, the development of stock association networks and the training of a Chinese financial large language model—thereby strengthening the foundations of multi-sourced heterogeneous data fusion from spatial correlation and textual sentiment dimensions. By employing tensor decomposition for stocks alongside graph and hypergraph sparsification, a more interpretable stock association network is constructed, furnishing high-quality input for subsequent real-time fusion of spatiotemporal features. Meanwhile, relying on a large corpus of Chinese financial texts, a Chinese financial large language model is trained and fine-tuned to enable precise quantification of news sentiment and policy interpretation information, thus providing strong support for the deep fusion of textual sentiment and historical transaction data. Specifically, the main contributions of this dissertation are as follows:

(1) To address the limitations of single-data-source stock association network construction and the absence of high-order correlations, this study proposes a tensor decomposition-based similarity matrix construction method as well as sparsification strategies for graph adjacency matrices and hypergraph incidence matrices. By integrating historical transaction data with fundamental data and exploiting their complementary information, the approach enhances modeling of the complex spatial relationships among stock assets and captures high-order interactions between multiple stocks. Using some component stocks of the CSI 300 Index as an example, the resulting stock association network exhibits notable topological features: (i) strong interconnectivity and high-frequency interactions that effectively characterize complex spatial relationships among stocks; (ii) key nodes that play vital roles in the network, thereby improving information transmission efficiency; and (iii) robust information dissemination capabilities supporting the capture and analysis of the stock market’s dynamic changes.

(2) To address the contextual understanding limitations of traditional text models in the domain of financial news and the current scarcity of Chinese financial large language models, this dissertation designs and trains a large language model specifically aimed at Chinese financial news sentiment classification tasks. By systematically analyzing three major Chinese financial text databases, a training sample set is constructed for masked language modeling, next-sentence prediction, and sentiment classification tasks. Based on the BERT (Bidirectional Encoder Representations from Transformers) framework’s pretraining mechanism and fine-tuning principles, domain-adaptive fine-tuning technologies are applied to the training of financial large language models, ultimately developing a financial large language model suited to the characteristics of China’s financial market. This model provides reliable support for subsequent research on extracting sentiment features from Chinese stock news text and deeper semantic understanding.

(3) To address the issue that existing decomposition-based ensemble methods in stock index forecasting, grounded in historical transaction data, have not effectively handled discrepancies in sub-sequence prediction loss, this dissertation introduces a stock index time-series forecasting model based on variational mode decomposition (VMD) and a self-weighted ensemble gated recurrent unit (GRU). The model first employs VMD to decompose the stock index time series into multiple sub-series, then uses the GRU network to predict each sub-series independently. Meanwhile, the model incorporates a self-weighted loss function that adaptively adjusts the sub-series’ weighting in the final prediction, thus optimizing overall predictive performance. Experimental results show that: (i) the self-weighted ensemble loss function significantly boosts the model’s overall predictive accuracy; and (ii) by selecting optimal hyperparameters, the model achieves the best predictive performance on different datasets, validating its robust adaptability. This model offers a new perspective on stock index time-series forecasting and lays a solid foundation for the development of high-precision predictive models.

(4) To address the problems in existing stock price spatiotemporal forecasting models, where sequential architectures for temporal and spatial features cannot achieve real-time integration and disrupt time-series features, this dissertation proposes a real-time fused hypergraph attention gated recurrent unit (GRU) forecasting model for stock spatiotemporal data. Building on the previously constructed stock association network, the model uses a hypergraph attention network to extract spatial features and embeds them into a GRU network, thereby enabling real-time spatiotemporal feature fusion and improving predictive accuracy. Experimental results demonstrate: (i) compared with conventional sequential architectures, the embedded spatiotemporal feature fusion architecture significantly enhances predictive accuracy; (ii) the model outperforms conventional statistical, machine learning, and deep learning approaches, showcasing its application potential in spatiotemporal forecasting; and (iii) the model exhibits greater stability in long-term forecasting compared with benchmark models. This model architecture holds high practical value and can be generalized to other spatiotemporal forecasting fields.

(5) To address the low feature interaction efficiency and insufficient feature fusion in existing models that integrate stock news sentiment features and time-series features, this dissertation proposes a self-attention mechanism-based residual gated recurrent unit prediction model that fuses textual and historical trading data. By leveraging a self-attention mechanism, the model effectively fuses stock news sentiment features and time-series features; through residual connections, the fused features are combined with the original features, which are then input to the GRU to capture time-series dependencies and produce final predictions. Experimental results demonstrate: (i) the model shows significant effectiveness and long-term predictive capability in integrating news sentiment and time-series features, offering valuable insights; and (ii) it indirectly verifies the strong capacity of the Chinese financial large language model trained herein for classifying sentiments in stock news, confirming its high practical value.

This study not only discusses, at the methodological level, deep learning-driven multi-sourced heterogeneous data fusion technology, but also verifies its strong performance in predicting prices of stock-like assets through empirical analysis, thereby providing robust support for intelligent prediction, risk management, and quantitative trading in stock markets. The findings indicate that this fusion technology can significantly enhance the accuracy, stability, and adaptability of forecasts, furnishing an important reference for the future development of intelligent finance research and optimization of investment strategies.