Research And Implementation Of A-share Portrait And Rotation Based On Deep Learning

With the rapid development of my country’s economy,the status of the A-share market in the world financial market become more and more important,which also causes more and more foreign investors to pay more attention on the Chinese stock market.The quantitative investment that has been quite mature abroad has also taken root in the Chinese financial market and grow in recent years.As more and more investors conduct quantitative trading,the research enthusiasm in the domestic quantitative trading field also increases.The stock market rotation research based on deep learning is one of the most popular research topics in the current quantitative investment field.And the construction of effective data sets and the selection of deep learning models are the two major factors affecting the results of stock market rotation research.Although traditional data sets cover all aspects of the stock market as much as possible,the invalid and redundant data in them will also affect the output results.In terms of model selection,traditional deep learning-based stock market research generally selects from RNN and CNN models and their variants.However,whether RNN or CNN is selected,the problem of model homogeneity caused by extensive research and application makes the market effectiveness of quantitative stock selection strategies based on traditional neural networks continue to decrease.Therefore,developing new quantitative stock selection strategies to explore more market trading opportunities has become an urgent problem to be solved.For the above problems,in terms of data,this paper uses machine learning models to dynamically establish A-share portraits for the stock market in each period to ensure the validity of the data;for the model,this paper conducts model research on two routes,RNN and CNN.In the direction of RNN,this paper proposes a new model,Attention-IndLSTM,by decoupling the structure of LSTM,which is currently studied more,and combining the self-attention mechanism algorithm.In the direction of CNN,this paper combines the self-attention algorithm with the TCN model propose the new Attention-TCN model.Through comparative experiments in terms of error and back-testing of return,it is found that new models perform better than the traditional deep learning model in the research of stock market rotation.In addition,this paper integrates the two models by using the Bagging ensemble algorithm,and performs error testing backtesting of return on the ensemble model.The final experimental result shows that the ensemble model can improve the model performance on stock market rotation investment by "learning from others’ strengths",which proves the effectiveness of the model in this research direction.The research results obtained in this paper provide a new idea for the research of stock market rotation in the field of quantitative investment,and have certain reference value.