DRL Optimizes Bi-Objective Portfolios with Reliability and Downside Risk

This paper presents a novel deep reinforcement learning (DRL) framework, called MORP-DRL, designed for bi-objective portfolio optimization. The framework addresses the inherent complexities of investment decisions under uncertainty, aiming to simultaneously optimize expected returns and mitigate downside risks. Unlike traditional static optimization methods, MORP-DRL is capable of handling sequential decision-making, tail risks, and market frictions such as transaction costs. The proposed system integrates three complementary risk measures: variance, Conditional Value-at-Risk (CVaR), and Entropic Value-at-Risk (EVaR). To accurately model market dynamics and heavy-tailed behavior, asset returns are represented using GARCH(1,1), Extreme Value Theory, and a t-copula dependence structure, with realistic scenarios generated via quasi-Monte Carlo simulation. A Proximal Policy Optimization (PPO) strategy is employed, incorporating practical constraints like transaction costs and portfolio bounds. Experimental results across various market regimes on global equity indices demonstrate that MORP-DRL achieves superior risk-return performance and enhanced downside risk reduction compared to the NSGA-II benchmark, while also showing scalability for high-dimensional portfolios.