RiskBound: Risk-Aware Boundary-Guided Portfolio Optimization via Action Space Reshaping

How can we build investment portfolios that remain robust under shifting market regimes and time-varying risk exposure? Portfolio optimization is a crucial problem in the financial domain, involving periodic asset allocation decisions to maximize long-term returns under risk. Recent advances have highlighted the effectiveness of deep reinforcement learning (RL) for such sequential decision-making problems. However, effective portfolio construction remains challenging due to time-varying, asset-level risk arising from non-stationary market dynamics. Failure to properly account for such risk leads to compounding losses and severe drawdowns over long investment horizons.