Research And Implementation Of Stock Trend Analysis System Based On Deep Learning

With the rapid development of Chinese economy,the domestic stock market has become one of the most concerned stock markets in the world.The field of stock analysis and stock prediction are the focus of scholars.The research shows that traditional statistics and machine learning can not dig out the deep logic behind the stock,which leads to inaccurate prediction,nevertheless the deep learning is more respected,which could give more reliable help to the majority of stock investors.The main works of the thesis are as follows:Firstly,we investigate the current situation of the domestic stock market environment,compare the stock fundamentals analysis method with the technical analysis method,and analyze the potential characteristics that affect the stock price volatility.Then we use TuShare to obtain and preprocess the stock data.The thesis investigates the mainstream clustering algorithms and the pain points of stock clustering task,uses the nearest neighbor propagation algorithm to cluster the daily rise and fall trend,and puts forward a preliminary noise reduction clustering precursor stock trend analysis method.The thesis investigates the shortcomings of mainstream stock forecasting models,discusses the pain points of stock forecasting tasks,improves its network structure and training methods,and proposes a more advantageous neural network model for stock trend forecasting.TensorFlow and Keras are used to construct,train and explore the model,Finally,through the comparative experiment analysis and verification,the model has better generalization effect in the task of stock trend prediction.The thesis designs and implements a stock trend analysis system based on deep learning,including regular automatic prediction of stock future trend and other functions.The main innovations of the thesis are as follows:A method for selecting stock data from multiple feature sources is proposed to strengthen the expression of stock characteristics.The incremental data storage algorithm is designed to solve the problem of too slow stock data collection.A clustering precursor stock is proposed trend analysis method to solve the problem of over-fitting in mainstream stock forecasting models.A tree clustering idea of interval fluctuations is proposed to reduce noise interference.A SA-LSTM-CNN stock trend prediction model is proposed,which solves the problem of distortion of the mainstream neural network fitting effect and inaccurate predictions during sharp rises or falls.For the design and analysis part of the model experiment,considering that the experimental cycle is limited every 24 hours on the stock trading day,the thesis proposes a two-dimensional experimental stock selection strategy;which lays the foundation for each round of experimental parameter adjustment to develop in a positive direction.A configuration strategy of hierarchical hyperparameter is proposed to balance the contradiction between the clustering magnitude of stocks and model fitting.For the stock trend analysis system,it can maintain the robustness of the system function through configuration fine-tuning model.The research results of the thesis have a certain reference value for the research of stock trend based on deep learning.