How to Verify Reward Alignment in Portfolio Reinforcement Learning

Addressing the problem in portfolio reinforcement learning where agents may precisely optimize rewards yet deviate from true financial preferences, this study develops a terminal mean-variance equilibrium decision audit method that can distinguish reward specification bias from learning error. The method constrains the payoff of single-action changes in a finite exogenous state market, supports independent sample expansion, and does not rely on equilibrium strategies. Validated through backtesting across 9 controlled markets, 20 random seeds, and 432 months of historical data, the method can quantify whether learned decisions match established financial objectives without affecting the backtest performance of the objectives themselves.