Study On Stock Index Forecasting And Trading Strategy Based On Improved BP Neural Network

Today,with the increasing participation of securities market,an effective prediction method has great practical significance for regulators and investors.Because the change of stock price or price index is very complicated,compared with the traditional non machine learning or simple technical index analysis,it is more convenient and effective to establish multivariable nonlinear system application model for prediction.Because of its wide adaptability,learning ability and mapping ability,neural network has made amazing achievements in modeling multivariable nonlinear systems,and has become a new method of forecasting time series.Based on the contributions made by predecessors in the prediction model construction and application,this paper further optimizes the performance evaluation.At the same time,in view of the different application conditions of the strategic performance evaluation,I hope to have a certain reference to the investment behavior.This paper is committed to building and using BP neural network to predict CSI 300 closing index data recently,improving the traditional mean line strategy and simulating the transaction performance evaluation.This paper takes the CSI 300 index related data from January 1,2010 to August 3,2017 as the research object and carries on the empirical analysis by using Eviews,S-PLUS and R software.The main part of the paper first carries out the exploration of the predictability of the CSI 300 closing index data,carrying out a long memory test of this time series.After this,the paper focuses on the construction and optimization of BP neural network.The optimization model includes the sample interval selected by Bai-Perron test,using principal component analysis method for determining the input variables,the number of hidden layer neurons determined by using ten fold cross validation and custom AIC standards,and the improvement on neural network algorithm by increasing the momentum.Finally,the strategy of improving the average transaction is made and the simulated transaction and performance evaluation are carried out.The analysis shows that there is long memory of the CSI 300 closing index data,thus,the historical data is feasible prediction.This is the premise to verify the forecast work;By combining previous experience and optimizing BP neural network step by step,we find that it can accurately predict today's closing price index through its own multivariable nonlinear system,and the good transaction results can be achieved through the development of trading strategies;Since the construction of BP neural network prediction before has never considered the time effect of the interception,the empirical part uses comparative method to explore the importance of interception time prediction results.The empirical results show that before the construction of the BP neural network is the need for structure mutation test of time series data,because of the possibility of the strategy will be suitable for the different risk preferences of investors.