Reinforcement learning with temporal and variable dependency-aware transformer for stock trading optimization.

Area of Science:

• Quantitative Finance
• Machine Learning
• Financial Engineering

Background:

• Stock trading optimization is vital for financial decision-making in dynamic markets.
• Existing Transformer-Reinforcement Learning (RL) methods often overlook multivariate interactions, limiting policy learning.
• There is a need for models that capture both temporal and cross-variable dependencies in market data.

Purpose of the Study:

• To propose a novel RL model integrating a Temporal and Variable Dependency-aware Transformer for stock trading optimization.
• To enhance the representation of market data by capturing diverse dependency relationships.
• To improve the decision-making information available for RL policy learning.

Main Methods:

• Developed a Temporal and Variable Dependency-aware Transformer integrated into a Reinforcement Learning (RL) framework.
• Implemented short-term and long-term prediction modules utilizing the Transformer for dependency exploration.
• Incorporated a relation representation module for stock asset correlations and a policy decision module for fusing representations.

Main Results:

• The proposed RL model achieved a Sharpe ratio of 1.48 and a portfolio return of 2.65.
• Demonstrated superior performance compared to state-of-the-art methods on CSI-300, S&P-100, and NASDAQ-100 datasets.
• The model effectively captured complex temporal and variable dependencies in market data.

Conclusions:

• The Temporal and Variable Dependency-aware Transformer significantly enhances stock trading optimization within an RL framework.
• The model's ability to learn diverse dependency relationships leads to more comprehensive decision-making information.
• This approach offers a promising direction for improving automated trading strategies and portfolio management.