Dilated Sequential Models with Dual-Stage Attention for Stock Prediction
Item type:Conference Paper
Author:Nguyen Tan Trung, Luu, G.-N.
Citation:Luu, G.-N., & Nguyen, T. T. (2024). Dilated Sequential Models with Dual-Stage Attention for Stock Prediction. In T. K. Dang, J. Küng, & T. M. Chung (Eds.), Future Data and Security Engineering. Big Data, Security and Privacy, Smart City and Industry 4.0 Applications (pp. 210–222). Springer Nature.
Abstract:This paper introduces the Dual-stage Attention Dilated architecture for sequential models to address the challenges of stock price prediction. Traditional sequential models face three major issues: complex dependencies, vanishing and exploding gradients, and efficient parallelization. The dilated recurrent skip connections effectively address these challenges. Additionally, the dual-stage attention mechanism handles long-term temporal dependencies and selects relevant driving series by employing input attention in the encoder and temporal attention in the decoder. By combining dual-stage attention with dilated recurrent architectures, the DAD models more efficiently select pertinent input features and capture long-term dependencies in time series data. Experimental results on the S&P 500, HSI, and DJIA datasets demonstrate that the proposed models outperform previous methods and are effective for any sequential model. The trade-off between computation time and accuracy will also be analyzed.
