Stock Price Prediction Based On Stock Industry Classification And Multi-factor Model

The Chinese stock market is a vital component of the market economy system that gradually formed with the continuous progress of reform and opening up,originating in the early 1990 s.Since the successful pilot operation of the Shenzhen Stock Exchange and the Shanghai Stock Exchange,the Chinese stock market has continuously developed and improved,contributing to the overall economic construction and social development of China.Although compared to Western stock markets,the Chinese stock market is still relatively young,it is not immune to various speculations and speculative behavior.Against this background,stock price prediction has become an important basis for investors to make correct investment decisions and has significant implications for securities market regulation.In this thesis,we study the short-term and long-term stock trend prediction based on deep learning based on the time-series characteristics of stocks,proposing a technical factor graph attention network based on stock interrelationships and a wide and deep asymmetric bidirectional LSTM memory unit model based on long-term feature extraction.We also design a quantitative stock selection and trading strategy that combines both models.The main work of this thesis includes:(1)Short-term stock trend prediction based on a technical factor graph attention network.We analyze the characteristics of mutual influence of short-term stock trends and construct multiple stock relationship graphs and technical factor data tables based on basic data such as industry classification data and concept data of corresponding listed companies,as well as daily trading data.We construct a technical factor graph attention network model based on these characteristics and demonstrate its effectiveness through the ablation experiments of different relationship data,compare its prediction effect with several commonly used deep learning methods under the same period data,and prove the stability of the model through training with different period data.(2)Long-term stock trend prediction based on Wide & Deep Asymmetric Bidirectional Legendre Memory Unit.In this thesis,the long-term trend of stock is studied,and the longterm trend characteristics in the implicit and time-series data are analyzed.By combining and improving the common recurrent neural networks for time-series data,the asymmetric bidirectional Legendre memory unit is proposed and combined with the framework of Wide& Deep network to build a Wide & Deep asymmetric bidirectional Legendre memory unit network model for long-term trend prediction of stock.Finally,the effectiveness of the model for long-term stock trend prediction is tested by the ablation experiment of its own module and the comparison experiment and analysis with the existing commonly used recurrent class neural networks.(3)Trading strategy construction and simulated trading tests.We combine the research in the first two parts and design a final trading strategy by combining the results of the two trend prediction models.We conduct detailed experiments to verify its effectiveness and profitability for practical use.