毕业生知识库

自证券市场诞生以来，股票预测就一直是金融投资和预测科学领域的重要研究问题。开发高效的预测方法对于理解金融市场运行规律、提升资产管理效率和防范系统性风险具有重要价值。近年来，以深度学习方法为代表的人工智能技术已经被广泛用于股票市场预测，这类技术主要通过挖掘股票量价时间序列内部的波动模式和财经文本中的金融语义等信息来判断股票行情的走势。然而随着金融市场的不断发展，金融系统的内部和外部关联性日益增强，在政治、经济、商业和投资活动等多种显式和隐式因素的交织作用下，股票价格之间逐渐呈现出强相关性和紧耦合性等复杂现象，这使得构建有效的深度学习预测模型变得愈发困难。

为了应对上述困难，最近基于图表示学习的股票市场行情预测方法通过使用网络表征股票间关联，将股票预测问题转换为图上的预测问题，并使用图表示学习方法从中挖掘有价值的信号。这些方法能够对股票市场进行整体性建模，在一定程度上提升了预测表现。但是现有的基于图表示学习的预测方法所使用的股票关联关系较为简单，不能考虑股票市场内部关联的复杂性和股票市场外部投资者的预测等能够精细刻画市场结构的高价值信息，这限制了图表示学习方法的预测能力。鉴于此，本论文从复杂网络的视角出发来解析股票市场的特性，构建了动态联动网络、多层关联网络、潜在依赖网络和分析师-股票交互网络，来充分表征股票与股票之间以及投资者与股票之间两种类型的市场结构。针对这四类金融网络的特性，本论文设计了一系列图表示学习方法提取金融网络中对市场行情预测有价值的信息，从而提升股票市场预测效果。

本论文的主要研究贡献包括：

1. 针对股票价格之间的动态联动效应建模困难的问题，本论文提出了面向动态网络的归纳式图表示学习方法。该方法通过量化股票价格之间波动的相似程度，构建了能够刻画联动效应的动态股票联动网络；随后该方法设计了一种混合注意力动态图神经网络，可以同时建模股票之间的动态联动和股票量价数据内部复杂的时序依赖。最后，该方法开发了面向动态网络的批梯度下降训练技术来降低模型的训练代价。多个市场上的股票预测对比实验验证了所提方法的有效性。

2. 针对现有的图表示学习预测方法无法有效建模股票之间的多维度关联这一困难，本论文提出了多层网络图表示学习股票预测方法。为了建模网络上目标股票的量价特征在时空上与相关股票的复杂依赖，该方法设计了时空聚合机制以获得能够保持相应网络金融含义的层内表示。其次，该方法引入了深度交叉网络来融合多个层的股票表示，从而获得更综合的层间表示。同时，为了提升模型预测结果在投资决策中的价值，该方法设计了多任务预测模块兼顾趋势和回报两种预测任务。在中国A股市场上开展的预测实验验证了所提方法的有效性。

3. 针对现有的图表示学习预测方法无法挖掘市场内部活动所导致的股票价格之间的潜在依赖这一困难，本论文提出了从历史数据中建模股票之间潜在关系的图表示学习预测方法。该方法使用端到端的方式，将稀疏自注意力机制学到的注意力权值矩阵作为股票之间的潜在依赖。随后将每个时刻获得的潜在依赖网络和股票特征输入到循环图神经网络中，建模股票市场的动态演化。为了获得更好的可扩展性，该方法还提出了基于K-means的分簇自注意力机制，将原生自注意力的时空复杂度由平方降为线性。在中美两个股票市场上的预测实验和仿真交易验证了所提方法的有效性。

4. 针对现有的图表示学习预测方法仍然面临的数据时间跨度短暂、高维稀疏和样本外泛化能力不够理想等挑战，本论文引入了市场外部信息，通过利用分析师对股票的群体预测能够产生相关公司的收益可预测性这一事实，本论文构建了分析师-股票交互网络来聚合分析师的预测结果。为了应对聚合过程中面临的动态异质性和不确定性问题，本论文在变分框架下使用重参数化的方式构建了面向分析师-股票交互网络的概率图神经堆叠机制。在本论文自主构建的一个大规模分析师-股票评论数据集上的预测实验验证了所提方法的有效性。

Since the inception of the security market, the stock prediction has been a crucial research topic in financial investment and forecasting science. Developing effective prediction methods holds immense value for understanding the operational mechanics of financial markets, augmenting asset management efficiency, and preventing systemic risks. Artificial intelligence technology has been widely used in stock market prediction in recent years, prominently represented by deep learning methods. These techniques are primarily targeted toward individual stocks and rely on mining volatility patterns within stock price time series and semantics in financial texts to predict market trends. However, the internal and external interdependencies within the financial system are increasingly strengthened with the persistent evolution of the financial market. Under the interweaving of multiple explicit and implicit factors such as political, economic, commercial, and investment activities, stock prices gradually exhibit complex phenomena such as strong correlation and tight coupling, which makes it increasingly difficult to construct an effective deep learning-based market prediction model for an individual stock.

To tackle the above difficulties, the latest graph representation learning-based stock prediction method utilizes networks to capture the interdependencies between stocks and transforms stock prediction problems into prediction problems on graphs. By leveraging graph representation learning methods, these techniques extract valuable signals from the resulting stock representations to model the stock market as a holistic system, thereby improving stock market prediction performance to some degree. However, the existing prediction methods are based on superficial stock relationships and fail to fully consider the complexity of internal correlations between stocks and external investor predictions that are valuable to describe the fine structure of the stock market, which limits the predictive ability of graph representation learning. This dissertation analyzes the complexity of the stock market from the perspective of a network and constructs dynamic co-movement networks, multilayer connection networks, hidden interdependency networks, and analyst-stock interaction networks to fully represent two types of market structures (stock-stock connections and investor-stock connections). Based on the characteristics of these four financial networks, this dissertation proposes a series of graph representation learning methods to extract valuable information encoded within financial networks and improve the prediction effect of the stock market.

The main research contributions of this dissertation are as follows:

1. To address the challenge of modeling the dynamic co-movement between stock prices, this dissertation proposes an inductive graph representation learning method for dynamic networks. By measuring the similarity between stock price movements, the method constructs the dynamic stock co-movement network that can depict the phenomenon of price co-movement. The method devises a hybrid-attention dynamic graph neural network to model the dynamic relationships between stocks and the complex temporal dependencies within stock time series data. A batch gradient descent training method is proposed to reduce the training cost for dynamic networks. The effectiveness of the proposed method is verified through comparative experiments on stock market prediction tasks in multiple markets.

2. To address the challenge that previous graph representation learning-based stock prediction methods cannot effectively model the multi-dimensional connections between stocks, this dissertation proposes a multilayer-network graph representation learning method for stock prediction. Specifically, to consider the complex interdependencies between focal stock and its connected stocks' features, the method develops a spatial-temporal aggregation mechanism to generate intra-layer stock representations containing the network's corresponding financial meaning. Then the method introduces a deep cross network to fuse the stock representations from multiple networks to obtain comprehensive inter-layer representations. Moreover, to improve the value of the model's prediction results in investment decision-making, the method proposes a multi-task prediction module that combines both trends and returns prediction tasks. Prediction experiments on the Chinese A-share market demonstrate the effectiveness of the proposed method.

3. To address the challenge that previous graph representation learning-based stock prediction methods cannot mine the hidden interdependencies between stock prices produced by the internal activities of the stock market, this dissertation proposes a graph representation learning-based stock prediction method by learning the hidden interdependencies between stocks from historical data. The method utilizes the attention weight matrix learned from the sparse attention mechanism in an end-to-end way as the hidden interdependencies between stocks. Then these hidden interdependencies networks and stock features are fed into a recurrent graph neural network to model the evolution of the market. Prediction experiments and simulated trading on two stock markets in China and the United States have verified the effectiveness of the proposed method.

4. To address the challenges faced by the existing graph representation learning-based prediction methods, such as short data time spans, high-dimensional sparsity, and insufficient out-of-sample generalization, this dissertation proposes a novel approach that introduces external market information. Based on the fact that stock analysts' prediction behaviors can produce return predictability for related companies, the method constructs an analyst-stock interaction network to aggregate analysts' predictions. To address the dynamic heterogeneity and uncertainty inherent in the aggregation process, the dissertation proposes a probabilistic graph neural stacking mechanism for analyst-stock interaction networks, which utilizes a reparameterization approach under a variational framework. The effectiveness of the proposed method is demonstrated through a prediction experiment conducted on a large-scale dataset of analyst stock reviews, which was independently constructed in this dissertation.