Forecasting Of CSI 300 Index With BiLSTM Model Based On AM

The stock market,as an important place for enterprises to raise capital and the public to invest profitably,has attracted the attention of people from all walks of life,and stock price prediction methods have sprung up.However,the complexity and uncertainty of stock data,which are non-stationary,highly noisy and volatile series,make classical statistical models diminish in stock price prediction.With the development of machine learning,it is found that deep learning can better capture stock data information,so this thesis starts from deep learning and based on LSTM,a bi-directional long and short-term memory network(AM-BiLSTM)model based on attention mechanism is constructed to predict the price of CSI 300 index.To address the shortcoming of non-differentiated learning of all features in general model training,the attention mechanism AM,which strengthens important features by redistributing the weights of each data,is introduced;and the optimization model of LSTM,bi-directional LSTM,is selected,which makes it possible to learn data information bi-directionally through cis-temporal and inverse time series to ensure the data utilization adequacy.In this way,the combined AM-BiLSTM model of this thesis is established.At the same time,Nadam is used to optimize the initial learning rate to improve the convergence speed and quality of the model.Secondly,In order to expand the data dimension,this thesis uses basic indicators based on a priori knowledge of finance,constructs a multi-category feature system for stocks consisting of technical indicators and key turning points,and takes the data after eliminating redundant information among indicators by PCA as the initial data.Finally,the AM-BiLSTM model is used to predict the closing prices of the representative CSI 300 index for 1,3 and 5 days in the future,and the prediction error is small from the evaluation index;and the prediction accuracy of the AM-BiLSTM model is higher compared with that of single models.It shows that the AM-BiLSTM model constructed in this thesis based on multicategory feature system works well for short-term prediction of CSI 300 index,further indicating that the algorithm design idea adopted in this thesis has certain reference value for accurate prediction of CSI 300 stock prices.