统计与数据科学学院知识库

如今全球金融市场迅速发展，智慧金融作为新兴的金融服务模式，能够提高决策的精确性和效率。随着经济的快速发展，中国式现代化金融服务正在形成，提供更为高效、智能、个性化的金融服务。在此背景下，量化投资和多因子选股策略的研究与应用获得广泛关注。本研究尝试通过引入图神经网络，捕捉股票间的复杂关系以及股票收益与各因子之间的动态交互，进一步提升多因子选股策略的效果，为量化投资领域提供新的思路和工具。

本文的研究区间为2016年1月至2023年6月共90个月，股票池为沪深300成分股，剔除缺失值较多的股票后，股票池中共229支股票。选择了5大类反映股票基础属性的共计36个因子作为构建多因子交易策略的备选因子，通过IC检验和MIC检验来评估因子有效性，研究了这些因子与股票收益之间的非线性关系，确保了选入模型的因子具有较强的预测能力。最后绘制热力图，计算方差膨胀系数剔除因子共线性，最终因子池中共保留15个因子。在建立图神经网络模型的过程中，本文考虑了股票间的价格关联性和行业关系等多维度数据，构建了复杂的图结构来捕捉市场中的微观结构和动态变化。在此基础上，创新性地引入了Huber损失函数来优化图神经网络的训练过程，并通过交叉验证调整参数。通过对比模型优化前后的误差以及损失曲线，本文发现，与传统的损失函数相比，Huber损失函数在处理股票收益率的尖峰厚尾分布特征时更为有效，能够减少极端值的影响，提高模型在复杂市场条件下的稳定性和鲁棒性。

该策略的回测结果表明，动态图神经网络多因子选股策略在测试周期内表现出色，实现了高达72.31%的总收益，年化收益率为25.31%。这一成果不仅证明了图神经网络在量化投资领域的应用潜力，也显示了基于先进算法优化的多因子选股策略在实际市场中的有效性。通过构建模型与回测分析，本文为量化投资策略的进一步发展和创新提供了有力的理论支持和实践指导。

Nowadays, the global financial market is developing rapidly, and smart finance, as a new financial service model, can improve the accuracy and efficiency of decision-making. With the rapid development of the economy, Chinese-style modern financial services are taking shape, providing more efficient, intelligent and personalized financial services. In this context, the research and application of quantitative investment and multi-factor stock selection strategy have been widely concerned. This study attempts to capture the complex relationship between stocks and the dynamic interaction between stock returns and various factors through the introduction of graph neural network, further improve the effect of multi-factor stock selection strategy, and provide new ideas and tools for quantitative investment.

The research period of this paper is 90 months from January 2016 to June 2023. The stock pool consists of 300 component stocks of Shanghai and Shenzhen. After excluding stocks with more missing values, the stock pool consists of 229 stocks. A total of 36 factors reflecting the basic attributes of 5 categories of stocks are selected as alternative factors for constructing multi-factor trading strategies. IC test and MIC test are used to evaluate the effectiveness of factors, and the nonlinear relationship between these factors and stock returns is studied to ensure that the factors selected in the model have strong predictive ability. Finally, the thermal map is drawn, the variance expansion coefficient is calculated to eliminate the collinearity of factors, and the final factor pool retains 15 factors. In the process of establishing the graph neural network model, this paper considers the multi-dimensional data such as price correlation and industry relationship among stocks, and constructs a complex graph structure to capture the microstructure and dynamic changes in the market. On this basis, the Huber loss function is innovatively introduced to optimize the training process of the graph neural network, and the parameters are adjusted by cross-validation. Finally, by comparing the error and loss curves of the model before and after optimization, this paper finds that compared with the traditional loss function, the Huber loss function is more effective in dealing with the peak and thick tail distribution characteristics of stock returns, which can reduce the influence of extreme values and improve the stability and robustness of the model under complex market conditions.

The backtest results of the strategy show that the dynamic graph neural network multi-factor stock selection strategy has performed well during the test period, achieving a total return of up to 72.31% and an annualized return of 25.31%. This achievement not only proves the application potential of graph neural network in the field of quantitative investment, but also shows the effectiveness of multi-factor stock selection strategy based on advanced algorithm optimization in the actual market. By constructing model and backtesting analysis, this paper provides strong theoretical support and practical guidance for the further development and innovation of quantitative investment strategy.