An improved LSTM stock price prediction model based on multiple basis function expansion and two-layer Bagging algorithm

Abstract

Using neural network technology to predict the time series data of stock prices is a hot issue in the field of computational science. For example, Long Short-Term memory network technology (LSTM) has been widely used in stock price prediction. However, if the original stock opening price series information is not enough to predict the future trend, the effect of LSTM model may not be ideal. In addition, improper selection of the step size will also make the prediction effect of the LSTM model poor. This paper proposes an improved LSTM model based on multi-view basis function expansion and Bagging algorithm. On the one hand, under a fixed step size, the intra-day price information of the stock is extracted by using a variety of basis function expansion forms as prior information, and then the Bagging algorithm is used to establish the relationship between the basis expansion coefficients and the prediction residual of the original LSTM model to predict the residual, thereby increasing the prediction accuracy. On the other hand, the problem of LSTM step size selection is adaptively solved by using the idea of model average instead of the model selection method, and the specific method is to combine the weighted average of LSTM models with different time lengths. The actual data analysis shows that the proposed method improves the prediction accuracy of the original LSTM model, and the t-test shows that the proposed method has significant robustness.