Research On Price Forcasting Of Shanghai Composite Index: Based On LASSO Dimensionality Reduction,LSTM And Mixing Model

As an important part of the financial market,the stock market can provide enterprises with important financing channels and is also regarded as a barometer of a country's economy.If the trend and range of stock prices can be predicted reasonably,it can guide investors to make reasonable investment decisions and asset allocation,increase investment returns and reduce risks.Traditional stock price forecasting is mainly based on linear models,while ignoring the nonlinear characteristics of stock price series.Machine learning methods for various non-stationary,stochastic and other nonlinear complex problems provide new ideas for stock price prediction.Among them,the long-term short-term memory neural network model(LSTM)has long-term and short-term memory capabilities,and empirical studies have shown that it can effectively predict stock prices.Therefore,this paper studies the problem of stock price prediction based on LSTM neural network in the case of high-dimensional variables and mixed data.According to previous studies,information on the volatility of stock returns can improve the accuracy of stock price forecasts.Therefore,this article uses the volatility sequence of returns as a feature of the LSTM model to predict stock prices.In this way,it is particularly important to accurately build the model of volatility.Traditional research is mainly based on GARCH model to model volatility,but ignores the predictive effect of macroeconomic variables on volatility.There are also some studies that consider the impact of macroeconomic variables on the volatility forecasts,but these studies mainly use traditional stepwise regression methods and principal component analysis to select macroeconomic variables to predict volatility.There are two problems in those research:one is that the model can only process data of the same frequency,so it can only get low-frequency volatility series.The other is that the use of the variable screening methodscan only effectively handle a small number of variables and the calculation is complicated,which requires manual screening of predictors in advance.However,the manual screening of predictors is subjective,easily leading omission.In order to solve these problems,this article makes two improvements.First of all,fully considering the high dimensionality of the macroeconomic variables actually available,this article adopts the more cutting-edge LASSO variable selection method to select variables,and further reduces the dimensionality according to factor analysis.Secondly,considering that the actual stock price prediction is mainly for high-frequency data,and the frequency of macroeconomic variables is low,this paper adopts the MIDAS method to estimate high-frequency volatility based on low-frequency macroeconomic factors to solve the problem of frequency inconsistency.Specifically,this article uses the LSTM neural network model to predict the stock index,and proposes the LSTM-LASSO-GARCH-MIDAS prediction model.In order to determine the input variables of the model,firstly,the article uses the LASSO method to screen high-dimensional macroeconomic variables,and perform factor analysis on the initially selected variables to obtain a small number of macroscopic factors.Next,the article uses the mixed frequency GARCH-MIDAS model to estimate the high frequency volatility through the low frequency macro factor.Finally,the highfrequency volatility estimated by factor analysis and the technical indicator factors after dimensionality reduction are used as features,and the LSTM neural network model is input to predict the price of the Shanghai stock index.The advantages of this model are:(1)LASSO and factor analysis methods can retain sufficient macroeconomic information while ensuring the simplicity of the model;(2)GARCH-MIDAS model can reveal the impact of low-frequency macroeconomic factors on high-frequency fluctuations;(3)The LSTM neural network model can be used to solve the nonstationary and non-linear prediction problems of stock price series.In the empirical research,this paper selects the daily closing price of the Shanghai Composite Index from January 5,2009 to September 30,2019 and monthly data of 102 macroeconomic variables over the same period,and uses the model proposed in this paper to predict stock prices.The model has achieved good prediction results.At the same time,the prediction effects of the LSTM model and the LSTM-GARCH model without macroeconomic variable information are compared.Mean square prediction error,DA test and DM test results show that the model has the best predictive ability.