Price Trailing for Financial Trading Using Deep Reinforcement Learning.

Area of Science:

• Quantitative Finance
• Computational Finance
• Artificial Intelligence in Finance

Background:

• Machine learning is increasingly applied to financial trading for automated pattern extraction.
• Developing effective machine learning trading strategies is complex due to reward design and hyperparameter tuning.
• Existing methods often fail to leverage information across multiple financial instruments.

Purpose of the Study:

• To propose a deep reinforcement learning (RL) approach for financial trading.
• To address the limitations of noisy profit-and-loss rewards and information silos across instruments.
• To enhance trading agent performance using novel reward shaping and data preprocessing.

Main Methods:

• Implemented a deep reinforcement learning framework with a novel price trailing-based reward shaping approach.
• Developed a data preprocessing method enabling training across diverse FOREX currency pairs.
• Utilized recurrent deep learning models to mitigate overfitting risks in market-wide RL agents.

Main Results:

• The proposed approach significantly improved agent performance metrics, including profit, Sharpe ratio, and maximum drawdown.
• Demonstrated effectiveness on a large-scale dataset with 28 instruments.
• Enabled the development of market-wide RL agents capable of exploiting cross-instrument information.

Conclusions:

• The novel deep reinforcement learning strategy with price trailing reward shaping offers a robust improvement over traditional methods.
• The data preprocessing technique facilitates the creation of more powerful, generalizable RL trading agents.
• This research advances the application of AI in quantitative trading, particularly for multi-instrument scenarios.