深度学习驱动的股票回报预测研究
摘
要
公司特征是预测股票收益的重要变量，对资产定价风险因子的构建和投资组合的管
理具有重要意义。受维度限制，早期研究往往聚集于少数几个变量的线性关系构造股票
投资组合，对股票截面预期收益差异的解释能力有限。近年来，基于神经网络的深度学
习方法快速崛起，该方法不仅可以探索高维输入变量与资本资产价格之间的复杂非线性
关系及动态联系，而且不容易受输入变量之间的相关性影响。目前，深度学习方法应用
于 A股市场进行股票预测方面的研究虽然处于起步阶段，但极大促进了研究者对 A股
资产定价和组合管理的理解。
本文以 90个公司特征为输入变量，使用线性回归模型、全连接神经网络模型、长
短期记忆神经网络模型对股票回报进行预测。基于预测构建组合的表现来评估不同模型
在 A股市场上对股票回报的预测能力，探讨深度学习方法相比传统线性方法在股票回
报预测方面是否更优。然后将预测对象换为同一时期股票收益率高低排序的相对回报，
探究不同模型预测能力的变化。最后对基于不同模型预测构建组合进行定价分析。
以股票回报为预测对象，基于传统线性回归模型预测构建组合可以获得一定的超额
收益，但受市值影响较大。基于全连接神经网络模型预测构建组合可以获取远超线性模
型的超额收益，说明公司特征和股票预期回报之间存在不可忽视的非线性关系和交互作
用。神经网络模型能更好地刻画出公司特征与预期收益之间的关系，证明了该关系并不
是简单的线性关系。考虑公司特征时序变化信息后，基于长短期记忆神经网络模型预测
构建组合获取的超额收益高于线性模型，略差于全连接神经网络模型，但夏普比率更优，
说明公司特征时序变化为股票回报预测提供了不可忽视的信息。
以股票相对回报为预测对象，基于线性回归模型和神经网络模型的预测构建组合获
取的超额收益都得到了显著提高。基于全连接神经网络预测构建投资组合的超额收益最
大，长短期记忆神经网络次之，线性回归模型最小。该研究结果与预测股票回报所得结
论基本一致，同时也说明以股票相对回报为预测对象受深度学习模型存在的过拟合问题
的影响更小。
在对基于模型预测构建组合进行定价分析中，以股票回报为预测对象，基于 OLS模
型预测构建组合的阿尔法不显著，而基于神经网络模型预测构建组合的阿尔法均非常显
著，说明线性多因子模型可以解释组合收益，线性回归模型从公司特征中提取的信息隐
含在 5个定价因子中。神经网络模型构建组合所依赖的信息不同于线性回归模型构建组
合所依赖的信息，非线性关系和交互效应会导致线性定价模型失效。以股票相对回报为
预测对象，基于神经网络模型预测构建组合的阿尔法优于线性模型，所有模型构建组合
I

摘要
的阿尔法均非常显著，说明降低过拟合不利影响后，模型能够挖掘更多的市场不完美非
有效产生的信息。
关键词：股票回报；预测；深度学习；公司特征
II

深度学习驱动的股票回报预测研究
Abstract
Corporate characteristics are important variables for predicting stock returns, and are of
great significance to the construction of asset pricing risk factors and the management of
investment portfolios. Due to the limitation of dimensions, early scholars often concentrated on
the linear relationship of a few variables to construct stock portfolios, and their ability to explain
the expected return differences of stock cross-sections was limited. In recent years, deep
learning methods based on neural networks have emerged rapidly. This method can not only
explore the complex nonlinear relationship and dynamic connection between high-dimensional
input variables and capital asset prices, but also is not easily affected by the correlation between
input variables. At present, the application of deep learning methods to the A-share market for
stock forecasting research is in its infancy, but it has greatly promoted researchers'
understanding of A-share asset pricing and portfolio management.
This paper uses 90 company characteristics as input variables, and uses linear regression
model, fully connected neural network model, and long short-term memory neural network
model to predict stock returns. Based on the performance of the forecast construction portfolio,
we evaluate the predictive ability of different models for stock returns in the A-share market,
and explore whether the deep learning method is better than the traditional linear method in
predicting stock returns. Then, the prediction object is changed to the relative return of stock
returns in the same period, and the changes in the prediction ability of different models are
explored. Finally, the pricing analysis is carried out on the forecast construction portfolio based
on different models.
Taking stock returns as the forecasting object, building a portfolio based on the traditional
linear regression model forecasting can obtain a certain excess return, but it is greatly affected
by the market value. Based on the prediction of the fully connected neural network model, the
construction of the portfolio can obtain excess returns far exceeding the linear model, indicating
that there is a non-negligible nonlinear relationship and interaction between company
characteristics and stock expected returns. The neural network model can better describe the
relationship between company characteristics and expected returns, which proves that the
relationship is not a simple linear relationship. After considering the time series change
information of company characteristics, the excess return obtained by building a portfolio based
on long short-term memory neural network model prediction is higher than the linear model,
slightly worse than the fully connected neural network model, but the Sharpe ratio is better,
III
。。。以下略