【作者】 王伟；

【导师】 石玉峰；

【作者基本信息】 山东大学 ， 应用统计（专业学位）， 2022， 硕士

【摘要】 近年来,越来越多的学者将机器学习理论运用在金融行业的各个方面,近几年备受关注的生成对抗网络也不例外。目前大部分学者将生成对抗网络及其改进模型用于数据生成或者是模拟资产价格变动的路径上,也用于进行价格预测和欺诈检测等方面;也有部分学者将生成对抗网络用于资产管理方面,但他们往往是先将资产价格变动路径或者分布生成出来,再利用生成的结果按照目标函数进行优化,进而得到资产组合策略。但在这过程中,生成对抗网络的生成器依然只是起着数据生成的作用,并没有直接得到资产组合策略。与其他学者不同的是,本文是利用Portfolio-GAN模型的生成器直接生成出资产的权重向量,结合资产的收益率得到相应的组合收益率,再进行对抗训练。同时,为能够实现多期调整和智能选股功能,我们在输入数据集进行了滑动窗口的处理以及在生成器的输出层采用ReLU激活函数。为了能够更好的实现,我们在生成器和判别器的隐藏层中分别加入了LSTM和一维卷积神经网络。为能够得到表现更好的资产组合,我们在选取多只相同类型的基金的基础上,对数据以每5个交易日为期限进行划分,选择每一个相同时间段中收益表现最好的数据按时间进行重新组合,得到一个新的组合收益序列,并称其为目标基金。我们将这个目标基金作为真实分布放入模型进行生成并做对比分析,同时将生成的投资组合策略与等权配置模型和风险平价模型进行比较。最终实证结果表明,生成的资产组合策略的收益率分布与目标基金的收益分布接近,说明Portfolio-GAN的学习能力较为优异。总体上来看,Portfolio-GAN生成的投资组合的评价指标表现优于基金的表现,累积收益率曲线优于或接近真实基金的累计收益率曲线,表明Portfolio-GAN模型生成投资组合优于真实基金,达到我们的预期目标;在与等权配置模型和风险平价模型的对比中,Portfolio-GAN生成的资产组合表现优异,其夏普比率和收益回撤比基本上都是最高的。因此我们认为Portfolio-GAN生成的资产组合是有实际意义的,能够成为智能投顾的一种新的方法,也可以对投资者提供实际参考价值。更多还原

【Abstract】 In recent years,more and more scholars have applied machine learning theory to various aspects of the financial field,and generative adversarial networks,which have received great attention in recent years,have also been applied to the financial field.Most scholars currently use generative adversarial networks and their improved models in data generation or simulating the path of asset price movements,as well as in price prediction and fraud detection,etc.Some scholars have also used generative adversarial networks in asset management,but they tend to generate asset price movement paths or distributions first,and then use the generated results to optimize according to an objective function,which leads to asset portfolio strategies.However,in this process,the generators of generative adversarial networks still only play the role of data generation and do not directly get the asset portfolio.Unlike other scholars,this paper uses the generator of the Portfolio-GAN model to directly generate the weight vector of the assets,combine the returns of the assets to obtain the corresponding portfolio returns,and then proceed to adversarial training.Also,to enable multi-period adjustment and intelligent stock selection,we apply a sliding window on the input dataset and a ReLU activation function on the output layer of the generator.For better implementation,we use LSTM and 1D convolutional neural networks in the hidden layers of the generator and discriminator,respectively.To obtain a better performance of asset portfolio,we select multiple funds in the same type,divide the data with a period of every 5 trading days,select the best return performance in each same time period and recombine the data by time to obtain a new portfolio return series,and call it the target fund.We put this target fund into the model as a real distribution for generation and do comparative analysis,and compare the generated portfolio strategy with the equal-weighted allocation model and the risk parity model.The final empirical results show that the return distribution of the generated asset portfolio strategies is close to that of the target fund,indicating that Portfolio-GAN has a superior learning ability.Overall,the evaluation metrics of the portfolios generated by Portfolio-GAN outperform the performance of the funds,and the cumulative return curves are better than or close to the cumulative return curves of the real funds,indicating that the Portfolio-GAN model generates portfolios that outperform the real funds and meet our expected objectives.In comparison with the equal-weighted allocation model and the risk parity model,the portfolio generated by Portfolio-GAN has performed well,with essentially the highest Sharpe ratio and Calmar ratio.Therefore,we believe that the asset portfolio generated by Portfolio-GAN is practically meaningful and can be a new approach to smart investment,and can also provide practical reference value to investors.更多还原