Robust Portfolio Optimization Considering Investor Preferences Using Maenhout Preferences and Reinforcement Learning

Abstract

This study develops a robust and adaptive portfolio optimization framework that integrates Maenhout’s ambiguity-averse preference structure with Proximal Policy Optimization (PPO) for dynamic asset allocation in the Tehran Stock Exchange. The model is designed to address two core challenges in emerging markets: parameter estimation error and model misspecification. To do so, the reward function combines expected return with explicit penalties for tail risk and downside risk, including Conditional Value-at-Risk (CVaR) and the Sortino ratio, so that the learned policy is not only return-seeking but also resilient under adverse market conditions. Using daily data from actively traded firms over the 2011–2025 period, the empirical evaluation shows that the proposed framework consistently outperforms classical benchmarks in risk-adjusted performance, downside protection, and stability across changing regimes. In particular, the Maenhout–PPO strategy achieves higher Sharpe and Sortino ratios, lower maximum drawdown, and more controlled tail losses than competing allocation rules. The results indicate that classical mean–variance logic is not sufficiently robust for a market characterized by nonstationarity, inflation pressure, liquidity concentration, and abrupt policy shocks. By contrast, the proposed framework offers a more realistic decision architecture for portfolio management under uncertainty, demonstrating that combining robust control principles with reinforcement learning can materially improve investment performance in emerging-market settings.