关键词: 股票价格预测, 多向延迟嵌入（MDT）, 卷积神经网络（CNN）, 长短时记忆网络（LSTM）

Abstract: Stock price prediction has always been the focus of investors’ attention in the stock market. In recent years, deep learning technology has been widely used in this field. Based on the fusion of convolutional neural network（CNN） and long-short-term memory network（LSTM） to build a CNN-LSTM model, it introduces multi-directional delay embedding tensor processing technology MDT（mutiway-delay-embedding）, daily stocks factor vector is subjected to factor reconstruction to generate the Hankel matrix, and the Hankel tensor is generated side by side in time as the input of the CNN-LSTM model. The convolution and pooling of CNN are used to extract features from the input data containing factor correlation information, and then the output feature matrix is input to the LSTM model for correlation prediction, thereby constructing the MDT-CNN-LSTM hybrid model. 48 companies and 12 stock factors involved in 22 industries are selected for stock price forecasting. Through comparative experiments in terms of forecasting accuracy and timeliness, it is shown that the proposs method performs better than other models. Finally, four types of stock indexes are selected for forecasting. The model effect is still at a relatively good level, which verifies the effectiveness and feasibility of the introduction of MDT technology.

Key words: stock price prediction, multi-directional delayed embedding（MDT）, convolutional neural network（CNN）, long-short-term memory network（LSTM）