Abstract
We develop a portfolio allocation framework that leverages deep learningtechniques to address challenges arising from high-dimensional, non-stationary,and low-signal-to-noise market information. Our approach includes a dynamicembedding method that reduces the non-stationary, high-dimensional state spaceinto a lower-dimensional representation. We design a reinforcement learning(RL) framework that integrates generative autoencoders and online meta-learningto dynamically embed market information, enabling the RL agent to focus on themost impactful parts of the state space for portfolio allocation decisions.Empirical analysis based on the top 500 U.S. stocks demonstrates that ourframework outperforms common portfolio benchmarks and the predict-then-optimize(PTO) approach using machine learning, particularly during periods of marketstress. Traditional factor models do not fully explain this superiorperformance. The framework's ability to time volatility reduces its marketexposure during turbulent times. Ablation studies confirm the robustness ofthis performance across various reinforcement learning algorithms.Additionally, the embedding and meta-learning techniques effectively manage thecomplexities of high-dimensional, noisy, and non-stationary financial data,enhancing both portfolio performance and risk management.