Deep learning with multiple scale attention and direction regularization for asset price prediction. (30th December 2021)

• Record Type: Journal Article
• Authors: Xu, Fucui; Tan, Shan
• Is Part Of: Expert systems with applications. Volume 186(2021)
• Publication Date: 2021-12-30
• Subjects: Asset price prediction -- Deep representations -- Scale attention mechanism -- Direction priori information; Expert systems (Computer science) -- Periodicals; Systèmes experts (Informatique) -- Périodiques; Electronic journals; 006.33
• DOI: 10.1016/j.eswa.2021.115796
• Languages: English
• ISSNs: 0957-4174

Abstract

Abstract: Forecasting the stock price is a challenging task due to its complex dynamic behaviors, affected by long-term trends, seasonal changes, cyclical changes, and irregular changes. Although many deep learning techniques have been applied to stock price forecasting, few of them have a deep insight into these complex behaviors. In this work, we propose a four-step hybrid model, named ESTA-Net, to adaptively extract these behavior patterns for stock price forecasting. Firstly, the empirical mode decomposition is applied to decompose a closing price sequence into intrinsic mode functions (IMFs). The goal of this step is to extract multiple quasi-stationary features of different time scales from the historical closing price sequence. Secondly, each IMF is modeled and forecasted by a temporal attention long short term memory (TALSTM) network. The TALSTM network is designed to capture the long term dependency of each IMF. Thirdly, the learned deep representations of IMFs are fed into a scale attention network (SANet), which adaptively selects relevant deep representations of multiple time scale features extracted from the historical price sequence. Finally, these learned deep features are fed into a fully connected layer to predict the future closing price. In addition, to make the proposed model learn the movement direction of the closing price, we propose a novel regularization term, i.e. the direction regularization term, to train the proposed model. This regularization term measures the inconsistency between the predicted movement direction and the actual movement direction of the closing price. Experiments show that the proposed model significantly outperforms benchmark models. Particularly, on seven financial market indices, the proposed model with the direction regularization term achieves the highest POCID (32.04% higher than that of CNN) and the lowest MAPE (37.36% lower than that of DA-RNN). Highlights: We design a scale attention deep learning model for asset price prediction. Our model learns the movement direction during forecasting the future stock price. Our model exceeds benchmarks in forecasting future price and movement direction.