股票收益率预测是金融研究和量化投资共同关注的重点话题，近年来机器学习和深度学习技术在该问题上的应用是一个研究热点。本文从实务的角度出发，基于分钟级别时序市场数据，应用深度学习方法，预测中国A股市场股票日内收益率。本研究的主要工作从四个方面展开：第一，深入分析股票日内收益率预测的特点和机器学习方法应用时需关注的问题，复现和搭建回归、分类两种基础深度学习模型，验证深度学习方法的可行性与有效性。第二，借鉴资产定价研究中截面收益率预测方式，在本问题中应用排序学习方法，将绝对收益率预测转化为截面收益率相对排序的预测，相比于基础模型，预测效果有显著提升。第三，考虑股票市场的时间动态性对模型训练的负面影响，尝试应用“无验证集训练方法”，优化训练流程，使得模型能够充分利用历史数据，进一步提升模型预测效果。第四，从量化投资角度出发，正确评估深度学习模型的预测能力与盈利能力，对模拟量化策略进行历史回测。回测结果显示，排序算法和无验证集训练方法有效地提升了深度学习模型对股票日内收益率的预测效果。本研究的主要结论是：其一，排序学习模型能通过更好地捕捉截面相对排序信息，显著提升预测效果。其二，使用“无验证集训练”方式，能够缓解股票市场模式时间动态性和深度学习中常使用的“训练-验证”训练方式之间的矛盾，有助于改善模型预测性能。

Stock yield prediction is a common concern in financial research and quantitative investment. The application of machine learning and deep learning technology in the problem is a hotspot in recent years. From a practical point of view, this paper uses deep learning method to predict the intraday return of stocks on Chinese A-share market based on the minute-level market data.The main work consists of four parts:1.To study the characteristics of stock intraday yield prediction problem and the application of machine learning methods, we build regression and classification models as basics, and check the feasibility and efficiency of deep learning methods.2.Based on the concern of cross-sectional return prediction in asset pricing research, Learning to Rank method is applied in this question to transform absolute yield prediction into cross-sectional yield relative ordering prediction, which has significantly improved compared with the basic models.3.Considering the negative effect of the time dynamics of stock market on deep learning models, a method called ‘non-validation set training’ is attempted to improve the training process, so that the model can make full use of historical data and further improve the model prediction.4.From quantitative investment aspect, the deep learning model are correctly evaluated, based on the simulation backtest. The backtest results show that Learning to Rank algorithm and the non-validation set training method effectively improve the prediction outcome.The conclusion is as follows: Firstly, Learning to Rank model can significantly improve prediction capability by better capturing cross-sectional yield sorting information. Secondly, non-validation set training relieves the confliction of market time dynamics and train-validation training process, which contributes to a better prediction.