本文主要的研究工作是建立基于深度信念网络的量化择时模型，通过实证分析来检验深度信念网络对我国股票市场交易数据的学习能力。大量实践表明股票市场上各量化交易指标之间存在某种非线性的相互作用关系，利用股票的历史交易数据和市场的其他诸多信息可以在一定程度上预测未来股价变动趋势。而神经网络模型能够有效地处理一些复杂的非线性问题，尤其是近年来在诸多领域的技术和应用都取得巨大突破的深度学习神经网络模型，可以运用计算机程序从海量数据中进行抽象学习，挖掘各类数据内部隐藏的规律。 本文研究的深度信念网络就是近十年来取得很多研究成果的深度学习网络常用模型之一。本文首先梳理了神经网络和深度学习的研究思想，并详细讨论了受限玻尔兹曼机和深度信念网络模型和算法结构以及模型搭建过程中各参数指标的选取；其次在TensorFlow计算框架下实现了深度信念网络模型的构建，并利用沪深300指数日度交易数据和分钟交易数据进行了股价涨跌分类预测；最后根据预测结果确定交易信号，并设定交易准则进行投资交易。 本文搭建的股价预测模型的最优结构包含两个隐层，对收盘价的整体预测正确率最高可达到约60%。本文的实证研究表明，基于深度信念网络构建的量化择时交易策略可以获得一定超额收益。对于沪深300指数股票日度交易数据，采用20天预测结构可以比5天预测和1天预测获得更好的投资收益，并且训练集时间跨度越长，策略收益效果越好；对于分钟交易数据，采用10分钟预测结构获得的策略收益效果最好，10分钟、20分钟和30分钟预测可以比1分钟和5分钟预测获得更好的收益。本文的实证分析结果证明了基于深度信念网络构建的量化择时策略在我国股票市场上具有一定的可行性和有效性。

The main research in this paper is to build a quantitative time selection model based on deep belief network and then test the learning ability of the deep belief network to Chinese stock market transaction data through empirical analysis. We know there is a nonlinear interaction between quantitative trading data and the historical data of trading and other information of the market can be used to predict the trend of future stock price. Lots of research showed that neural network models can effectively deal with some complex nonlinear problems. In recent years, deep learning neural network has made great breakthroughs in many fields, and can be used to use computer programs to learn and mine the hidden rules from all kinds of mass data. The deep belief network studied in this paper is one of the most commonly used models of deep learning network in recent ten years. This paper first combed the research ideas of neural network and deep learning, and discussed the model and algorithm structure of restricted Boltzmann machine and deep belief network and the selection of parameters in the process of model building in detail. Under the TensorFlow framework, the deep belief network model is built. We used the CSI300 daily and minutely transaction data to carry on the stock price classification prediction, then according to the prediction result we determined the transaction signal and carried on the transaction according to the transaction criterion. The optimal structure of stock price prediction model built in this paper contains two hidden layers, and the best prediction accuracy of closing price is about 60%. The empirical research in this paper showed that quantified timing trading strategy based on deep belief network can achieve certain excess returns. For the CSI300 daily transaction data, the 20 days prediction structure can get better results than the 5 days or 1 day prediction structure, and the longer the time span of the training set is, the better the strategy benefit is. For minutely trading data, 10 minutes prediction structure can get best benefits, and the 10 minutes, 20 minutes and 30 minutes prediction structure can get better results than 1 minute or 5 minutes prediction structure. The empirical analysis results in this paper showed that the quantitative timing strategy based on deep belief network is feasible and effective in Chinese stock market.