Explainable post hoc portfolio management financial policy of a Deep Reinforcement Learning agent

Area of Science:

• Quantitative Finance
• Artificial Intelligence
• Machine Learning

Background:

• Modern portfolio theory models fail in high-volatility markets due to unmet assumptions.
• Deep Reinforcement Learning (DRL) shows promise for portfolio management but lacks interpretability.
• Current DRL methods use complex neural networks that obscure decision-making processes.

Purpose of the Study:

• To develop an Explainable Deep Reinforcement Learning (XDRL) approach for financial portfolio management (PM).
• To enhance the transparency and interpretability of DRL agent actions in investment decision-making.
• To provide a method for assessing the behavior and risk associated with DRL-driven investment policies.

Main Methods:

• Integration of Proximal Policy Optimization (PPO) DRL algorithm with explainable AI techniques.
• Application of model-agnostic methods including feature importance, SHAP, and LIME for interpretability.
• Development of a novel XDRL framework for post hoc financial policy explanation.

Main Results:

• Successfully demonstrated the ability to interpret DRL agent actions in real-time.
• Identified key features influencing investment decisions, validating the model's explanatory power.
• Empirically illustrated the XDRL approach's effectiveness in enhancing prediction transparency.

Conclusions:

• The developed XDRL approach offers the first explainable post hoc financial policy for DRL agents in PM.
• This method allows for the assessment of agent behavior against investment policy requirements and risk evaluation.
• XDRL enhances trust and accountability in AI-driven financial decision-making.