Article Preview
Top1. Introduction
In the stock market, for decision makers, the more accurate of prediction, the more favorable it is for future profit acquisition and risk aversion. There are stocks trading at all times, and the changes of stock prices are a comprehensive reflection of political, economic and social factors. Therefore, the stock price prediction has the characteristics of large change range, many change factors, and unstable changes. Hence, how to judge and grasp the level and trend of stock price changes in the stock market is the most concerned issue for stock traders (Cai 2011).
Since the emergence of stocks, stock predictions have received extensive attentions and active researches. So far, many scholars have proposed many prediction methods, and these methods can be divided into two categories, one is traditional statistical methods, and the other is artificial intelligence methods (Guo 1994). Traditional stock prediction methods are all based on linear models. However, due to the relationship among stock prices and political, economic, and social factors are nonlinear, traditional methods cannot fully consider the impact of various factors on stock prices, cannot analyze and fit the highly nonlinear and multi-factor stock market well, and the prediction accuracy is not ideal. In order to overcome this shortcoming, in recent years, many scholars have used various neural network artificial intelligence methods to conduct a large number of stock price prediction studies, and have achieved successful applications (Saad 1998). But the neural network does not consider the selection of input variables. When there are too many input variables, the network structure is complicated, which may increase the training burden of the neural network, and decrease the learning speed. At the same time, subjective selection is likely to select input variables which have little correlation with the output, which will increase the possibility of falling into a local minimum. And this behavior can’t improve the prediction accuracy, but may reduce the performance of neural network prediction (Kohzadi 1996).
In recent years, deep learning has attracted more and more attentions, various deep neural network models have been proposed and have played an important role in natural language processing, image recognition, data analysis or mining and other fields, and the application prospect is extremely broad. Recurrent neural network (RNN) (Elman 1990) can learn the relationship between data at different time points, even the non-linear relationships can also be well expressed and utilized, and the trained model also has high-precision predictive output for different input data. As a result, it has strong adaptive capabilities, and these advantages provide a lot of help for stock price prediction. The stock market is a multi-dimensional dynamic non-linear system, and is affected by many factors, including the historical data, technical indicators, macro factors, etc. We can effectively explain the relationship of changes in the stock market only when a suitable set of inputs is found, and then it becomes possible to make accurate predictions of stock price trends under various changing conditions (Kuan 1994).
Figure 1.
The illustration of the architecture of attention mechanism based long short-term memory (LSTM) model for interval data