Reinforcement Learning for Continuous-Time Mean-Variance Portfolios

Summary

This paper summary describes a reinforcement-learning approach to continuous-time mean-variance portfolio selection. It frames the problem as entropy-regularized relaxed stochastic control, aiming to balance theoretical tractability with practical implementation. The authors establish that an optimal feedback policy must be Gaussian, with variance that changes over time, and derive a policy-improvement result used to construct a realizable learning algorithm.

The summary reports that simulations and empirical studies found the proposed algorithm and variants outperformed traditional methods and approaches based on deep neural networks. It does not provide the experimental setup, data, evaluation metrics, or numerical results, so the comparisons cannot be assessed from this account alone. The findings are specific to the paper’s modeled portfolio-selection problem; the summary does not establish how the method performs under different assumptions, asset universes, or live trading conditions.

Key ideas

• The method applies reinforcement learning to continuous-time mean-variance portfolio selection.
• The problem is formulated as entropy-regularized relaxed stochastic control.
• The optimal feedback policy is characterized as Gaussian with time-varying variance.
• A policy-improvement result supports construction of a practical learning algorithm.
• The summary reports favorable simulation and empirical comparisons but omits their details.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.