A Study On Stock Price Prediction And Stock Portfolio Based On Improved LSTM Neural Network

Stock price forecasting and stock portfolio management based on financial time series are two important research topics in the field of investment.In the face of traditional time series forecasting models for non-linear and volatile financial asset price research,time series changes without rules to follow can no longer meet the practical application of stock price and return forecasting.In terms of stock portfolio,it is necessary to meet the asset allocation needs of different investors by weighing the scale of expected return and risk.This study seeks to achieve quantitative investment through intelligent algorithms.This dissertation uses stock technical indicators as the research benchmark in the context of the Chinese stock market without the intervention of politics,policies and business conditions,and with no major unexpected events and no information of individual stock explosion as the research premise.For individual stock price prediction,a stock price prediction model based on MDT-Bi LSTM is proposed.The tensor processing technique MDT with multi-way delay embedding is used to reconstruct the stock factor vector with daily frequency data,and the stock price prediction is completed by the long-term dependence of LSTM.The results show that the model is operable in terms of prediction accuracy and timeliness.At the same time,the accuracy of the stock price prediction is enhanced to confirm the effective selection of individual stock return predictors in stock portfolio construction.On the problem of constructing stock portfolios,this thesis constructs dynamic stock portfolios from two perspectives: dynamic optimal asset allocation with high-frequency yield forecasting and systematic risk hedging to diversify investment risk and guarantee return acquisition.On the one hand,ELSTM-BL stock portfolio model is constructed based on yield forecasting.By forecasting the next day’s individual stock returns to form the subjective view parameters of the BL model and inputting them into the BL model to obtain the allocation weights of the stock portfolio,the flexible learning rate setting method using exponential decay avoids the parameters swinging back and forth on both sides of the extreme optimal value,thus making the model converge better and enhancing the stability of the model while ensuring the speed of optimization.Compared with other models,it has higher investment return and stability.On the other hand,ELSTM-ALPHA stock portfolio model is proposed based on risk hedging theory.The stock index futures are used to hedge the systematic risk of the stock portfolio and the ELSTM neural network is used to regress the objective function to obtain the absolute alpha return.The experiments show that the return is stable with ELSTM-ALPHA investment strategy and the retracement risk is significantly reduced.This dissertation is based on deep learning methods and improved LSTM neural network model from the stock price prediction and construction of stock portfolio two aspects of the subject of analysis and research,combined with the model and data comparison and analysis of the test method to propose a scientific and effective quantitative investment tools for the securities market,to help investors judge the trend of individual stocks,accurate control of the timing of entry and exit investment,optimize the asset structure,to achieve efficient and safe investment The aim is to help investors judge individual stock trends,accurately control the timing of entry and exit,optimize asset structure,and achieve efficient and safe investment.This provides a new way of thinking for solving quantitative investment problems in the stock market using deep learning methods.