Deep Reinforcement Learning for Reliability-Based Bi-Objective Portfolio Optimization

The paper introduces MORP-DRL, a deep reinforcement learning framework for reliability-based bi-objective portfolio optimization. It uses Proximal Policy Optimization to jointly optimize expected return and downside risk, represented through variance, Conditional Value-at-Risk, and Entropic Value-at-Risk. Market uncertainty and heavy tails are modeled with GARCH(1,1), Extreme Value Theory, and a t-copula dependence structure, with quasi-Monte Carlo scenario generation. Experiments cover ten global equity indices across pre-COVID, COVID, and post-COVID regimes, and compare the method with NSGA-II under transaction costs and portfolio bounds.