Reinforcement Learning-Based Portfolio Optimization Under Dynamic Market Conditions and Multi-Objective Financial Constraints

Abstract

Portfolio management requires decisions under evolving return distributions, liquidity limitations, transaction costs, and drawdown constraints. This paper proposes a constrained reinforcement learning architecture combining regime-sensitive observations, a recurrent policy, risk-aware rewards, and explicit feasible-set projection. The framework jointly addresses return, downside exposure, turnover, concentration, and liquidity while separating hard mandates from soft preferences. A walk-forward evaluation design, reproducible baselines, stress testing, and sensitivity analysis are specified. Numerical tables and graphs illustrate internally consistent hypothetical outcomes rather than measured performance: no historical data were downloaded, model trained, or real backtest executed for this manuscript. The illustrative comparison shows how a constrained agent might trade some gross return for improved drawdown and lower turnover. The paper contributes a practical specification for implementable and auditable portfolio learning and identifies the empirical tests required before any claims of superiority.