Behavioral Proxy Conditioning for Financial Stress Scenario Generation with a Pretrained Diffusion Model

Controllable financial scenario generation is challenging because historical crises are rare and difficult to model directly. To address this, we adapt a pretrained diffusion model for financial time-series generation using interpretable behavioral proxy conditioning. The proxy combines HMM regime probabilities, cross-market correlation, and realized volatility rank, allowing the model to generate calm and stress scenarios for seven global equity indices. To make conditioning more robust, we randomly drop proxy groups during training, allowing the model to use information from all components. These design choices enable the model to generate realistic financial scenarios that remain responsive to regime conditions. The final model passes $20/21$ risk/tail checks plus $7/7$ volatility sanity checks, for a total of 27/28 regime-separation checks across all four seeds with an average stress-to-calm volatility ratio of $(3.625 ± 0.084)\times$ and $0$% exact replay of historical windows. Portfolio-level stress scenarios also show substantially worse CVaR95 than calm scenarios, suggesting that behavioral proxy conditioning can make pretrained generative models effective for downstream financial stress testing under data scarcity.