LAUSR

In deep learning based stock trading strategy models, most of the research just use simple convolutional neural networks (CNN) to process stock data. But task-specific neural network structures have been proposed extensively, and their effectiveness has been demonstrated in computer vision (CV) and natural language processing (NLP) tasks. In this paper, we proposed a multi-scale convolutional neural feature extraction network (MS-CNN) for stock data, which can better extract stock trend features and thus make better decisions. The network structure inspired by the human stock trading model: in human behavior, we do not only look at a single set of stock data, but rather combine all the stock data, such as opening, closing, and trading volume, to make a comprehensive judgment. And humans will consider the current stock trend on different time scales, such as 3-Day Line and 5-Day Line. This is consistent with the two-dimensional convolution kernels commonly used in CV tasks, so we used convolution kernels of $3\times 3$ and $5\times 5$ in the network with two-dimensional convolution size and constructed a novel network structure for stock data. With double deep Q networks (DDQN) algorithm, we get the best performance for our network. The experimental results show that we can obtain high yield on the datasets of Dow Jones (DJI), AAPLE (AAPL), and General Electric (GE).