CEAM: A Novel Approach Using Cycle Embeddings with Attention Mechanism for Stock Price Prediction

This paper presents a novel deep learning approach for the stock price prediction using a cycle embeddings with attention mechanism (CEAM) applying on Dual-Stage Attention-Based RNN (DA-RNN) model. The cycle characteristic is an important factor in time series prediction problem since it affects the trend of stock price. Thus, an effective cycle information can improve the prediction performance of stock price. In past years, many researches use the cycle feature with other features together as equally important, which might dilute the weight of cycle information since the cycle information should be paid more attention when making prediction on periodic data. As the result, we use CEAM making prediction with cycle information hidden in periodic data. The deep learning-based method has been developed in many fields and is a powerful prediction system. In addition, many researches use the embeddings feature and the attention mechanism to improve the prediction performance. In this paper, we propose a novel approach to capture the cycle information and use it to predict stock prices in U.S. stock market. The cycle information can be formed as a distributed vector as embeddings, called cycle embeddings. The CEAM approach use cycle embeddings to pay attention on periodically historical time series data by learning the cycle semantic relations between cycle characteristics and historical stock prices to optimize the prediction model. Therefore, the CEAM approach can improve the prediction performance for stock price. The experiments in this paper show that our proposed CEAM approach outperforms the another model which combines cycle feature with other features together as equally important.