Abstract
Despite the popularity of trend-following strategies in financial markets, they often lack adaptability to the emerging varied markets. Recently, deep learning (DL) methods demonstrate the effectiveness in stock-market analysis. Thus, the application of DL methods to enhance trend-following strategies has received substantial attention. However, there are two key challenges to be solved before the adoption of DL methods in enhancing trend-following strategies: (1) how to design an effective data selector to include more related data? (2) how to design a profit-based model to enhance strategies? To address these two challenges, this paper contributes to a new framework, namely profit-based deep architecture with the integration of reinforced data selector (PDA-RDS) to improve the effectiveness of DL methods. In particular, profit-based deep architecture (PDA) integrates a dynamic profit weight and a focal loss function to obtain high profits. In addition, reinforced data selector (RDS) is constructed to select high-quality training samples and a training-aware immediate reward is designated to improve the effectiveness of RDS. Extensive experiments on both U.S. and China stock-market datasets demonstrate that PDA-RDS outperforms the state-of-the-art baseline methods in terms of higher cumulative percentage rate and average percentage rate, both of which are crucial to investment strategies.
Original language | English |
---|---|
Pages (from-to) | 1685-1705 |
Number of pages | 21 |
Journal | World Wide Web |
Volume | 26 |
Issue number | 4 |
DOIs | |
Publication status | Published - 14 Oct 2022 |
Scopus Subject Areas
- Software
- Hardware and Architecture
- Computer Networks and Communications
User-Defined Keywords
- Data selection
- Deep learning
- Reinforcement learning
- Transfer learning
- Trend-following strategy