Indicator-Specific Recurrent Neural Networks with Co-teaching for Stock Trend Prediction.

Stock trend prediction is a challenging problem due to the complexity of stock data. Recently, many works applied deep learning methods for stock trend prediction and achieve impressive results. However, these methods still suffer from two limitations: 1) Various types of technical indicators are input into a single model, making it difficult for the model to learn differentiated features. 2) Noisy data in the stocks is not handled effectively. Therefore, in this paper, we propose a stock trend prediction framework using indicator-specific recurrent neural networks with co-teaching. Specifically, we first collect data from Chinese stock market and divide them into fourteen categories. Then we apply multiple RNNs to extract features separately from different technical indicator categories which can learn comprehensive features. In addition, we leverage multi-head attention for effective feature interaction and fusion. At last, we utilize co-teaching method during the training process to reduce the impact of noisy data. Experimental results show both the effectiveness and superiority of our method.