| In recent years,supervised learning and reinforcement learning have been applied to the research of quantitative investment in stock market successively.By using this kind of deep learning technology to realize massive data analysis,quantitative investment has gradually transited from simple CTA / alpha strategies to trading strategies realized by artificial intelligence.At the same time,there are many problems in the application of artificial intelligence in quantitative investment,which are included as follows: 1.The state-of-art methods mainly focus on model structures and algorithm,which have the weakness of insufficient analysis of investment behavior factors;2.These methods overstate the ability of feature extraction of network,and neglect the importance of prior knowledge of market law;3.Most of the strategic models lack risk control for trading behavior,and fail to realize that it is easy to lead to serious risks by simple trading operations.In view of these existing problems,this article follows the logic formed by financial investment strategies and uses a multi-source-driven approach to analysis.In this way,the big task of stock index buying and selling operations can be decomposed into interrelated and progressive layered strategy problems.First,we divided it into three tasks: trend prediction,feature extraction,and operation decision.Then,we build three modules separately through in-depth exploration and algorithm design for each task.Finally,a multi-source-driven quantitative investment model under a deep layering strategy was established.Among them,the trend prediction module is based on the WGAN(Wasserstein Generative Adversarial Networks)principle and the distribution characteristics of price time series.In this part,a RC(Regression constraint)-WGAN network for predicting the index’s ups and downs is established.As for the feature extraction module,it is based on financial related theories such as time series,technical indicators,and behavior.In this part,we propose a multi-source feature combination based on prior knowledge to describe the stock market.As for the operation decision module,the Deep Deterministic Policy Gradient(DDPG)algorithm based on reinforcement learning is used.And we have designed a risk control network to implement specific trading position strategies.Finally,this paper first uses Dow Jones industrial index and Shanghai Stock Exchange Index as data sources for training and testing in the virtual marketenvironment,and verifies the effectiveness and feasibility of each module and model as a whole.Secondly,as a trader,I refers to the auxiliary investment strategies provided by this model,and uses Guotai Junan securities account to carry out a one-year real stock market experiment to evaluate the application of the model in the actual scenario.The results of simulation and real market experiments show that the method in this paper performs well in terms of return on investment and robustness,and can provide valuable reference for investors. |
