Stress-Aware Scenario Generation for Reliable Portfolio Inference under Regime Shifts

Financial markets face shocks and regime shifts that destabilize portfolio strategies. Classical optimizers average across states and underprice tail risk, while RL agents such as FinRL often overfit to noise and fail in crises. Even Bayesian regime models and entropy-regularized RL (e.g.\ SAC) either work offline or use fixed bandwidths, leaving them brittle when stress levels change. We propose a trust-aware belief update that anchors regime posteriors to the prior with a KL term and adapts entropy dynamically from residual stress. This regulates the inference bandwidth, contracting in stable periods and widening in crises, and doubles as a generative scenario engine that reproduces the persistence and recovery of the crisis. Across synthetic regimes, noisy bandits, and long-horizon portfolios, the method cuts drawdowns, speeds recovery, and improves calibration while sustaining strong Sharpe and Sortino ratios. Statistical tests confirm predictive content in regimes, and utility valuations show signals align with real investor trade-offs. Stress-aware belief updates thus offer a lightweight, online mechanism for robust and interpretable portfolio inference in non-stationary markets.