Crisis-Aware Regime-Conditioned Diffusion with CVaR Allocation (2510.10807v1)

Abstract: We study whether regime-conditioned generative scenarios, coupled with a convex CVaR allocator, improve portfolio decisions under regime shifts. We introduce Multi-Agent Regime-Conditioned Diffusion (MARCD), which (i) infers latent regimes via a Gaussian HMM, (ii) trains a diffusion model with a tail-weighted objective and a regime-specialized mixture-of-experts (MoE) denoiser to enrich crisis co-movements, and (iii) feeds the generated scenarios into a turnover-aware CVaR epigraph quadratic program with explicit governance. In strict walk-forward tests on liquid multi-asset ETFs (2005-2025), MARCD outperforms standard allocators and improves calibration relative to popular generators. Over 2020-2025 out-of-sample (monthly; 10 bps), MARCD attains Sharpe 1.23 (BL 1.02) and MaxDD 9.3 percent (BL 14.1 percent), a 34 percent reduction, at comparable turnover; stationary block-bootstrap intervals indicate the Sharpe uplift is significant at 5 percent. We provide theory linking tail-weighted diffusion to spectral-risk control of the decision-relevant CVaR gap, oracle/consistency results for the regime-MoE denoiser, and Lipschitz/regret guarantees for the allocator. Together, MARCD offers a reproducible bridge from tail-faithful scenario modeling to governed portfolio decisions with materially improved drawdown control.