【作者】 董坚；

【导师】 姜远； 詹德川；

【作者基本信息】 南京大学 ， 计算机科学与技术， 2019， 硕士

【摘要】 量化投资有着纪律性、系统性和分散化投资等诸多优点,因此其越来越受到学术界和投资界的关注。近些年随着人工智能和深度学习技术的发展,学者与投资实践者把更多的机器学习算法应用到投资领域的各个方面并取得了不错的投资收益。与成熟的海外市场相比,量化投资在国内市场具有更广阔的发展空间和研究价值,本文基于国内市场上对数据挖掘和深度学习在量化选股和统计套利中的应用进行了研究,提出了基于深度模型和谱方法的组合选股模型。文章的主要工作分为如下三个部分:1.将深度学习算法应用到量化选股的问题中。本文用IRGAN(Information Retrieval in Generative Adversarial Networks)作为选股模型,然后设计了一个卷积网络结构并嵌套在IRGAN的生成器和判别器组件中,根据模型的输出得到选股结果。实验结果表明,本文方法在选择高收益率股票上的精度优于其他对比方法。2.将谱方法应用到统计套利选择价格曲线相似的股票组合的问题中。本文用动态时间规整距离替换欧氏距离作为股票的距离度量,通过谱聚类算法聚类产生相似股票组合。实验结果表明,在中低频时间尺度上,本文方法分析出的相似股票组合在价格曲线走势的相似程度上高于其他对比算法。3.提出基于深度模型和谱方法的组合选股模型。本文将深度模型选择出的股票作为核心股票集,用基于谱方法的相似股模型得到与核心股票的价格曲线走势相似的相似股票集,将核心股票集和相似股票集作为组合选股模型的选股结果。实验结果表明,本文提出的组合模型能够大幅提高模型选股的精度并且该模型的回测收益好于其他对比算法。更多还原

【Abstract】 Quantitative investment has many advantages,such as discipline,systematization and decentralization.It has attracted more and more attention from academia and investment circles.In recent years,with the development of artificial intelligence and deep learning technology,scholars and investment practitioners have applied more machine learning algorithms to various aspects of the financial field and achieved good investment returns.Compared with mature overseas markets,quantitative investment has a broader development space and unlimited research value in the domestic market.In this paper,the application of data mining and deep learning in quantitative stock selection and statistical arbitrage is studied,and a combination stock selection model based on deep model and spectral method is proposed.The main work of this paper is divided into the following three parts:1.Deep learning is applied to the problem of quantitative stock selection.Using IRGAN(Information Retrieval in Generative Adversarial Networks)as the stock selection model,a convolutional network structure is designed and embedded in the generator and discriminator components of IRGAN.The stock selection results are obtained according to the output of the model.The experimental results show that the accuracy of this method in choosing high yield stocks is better than that of other comparative methods.2.Spectral method is applied to the problem of selecting similar stock portfolios for statistical arbitrage.The Euclidean distance is replaced by the dynamic time warping distance as the distance measure of stocks.Similar stock portfolios are generated by clustering using spectral clustering algorithm.The experimental results show that on the medium and low frequency time scale,the similarity degree of the similar stock portfolio mined by this method is higher than that of other comparative algorithms.3.Combining the above two models,a combination stock selection model based on deep model and spectral method is proposed.The stock selected by the deep model is taken as the core stock set,and the similar stock model based on spectral method is used to get the stock set similar to the price trend of the core stock.The core stock set and the similar stock set are taken as the result of the combination stock selection model.The experimental results show that the combined model proposed in this paper can greatly improve the accuracy of stock selection results and the return on simulated investment of the model is better than other comparative algorithms.更多还原