Multi-Agent Regime-Conditioned Diffusion (MARCD) for CVaR-Constrained Portfolio Decisions

Battle Count: 69

Directly addresses core quantitative trading challenges: regime-aware scenario generation, tail-risk management via CVaR, turnover-constrained portfolio allocation, and walk-forward backtesting. The 34% MaxDD reduction vs. Black-Litterman and improved Sharpe (1.23 vs 1.02) are practically significant. The auditable pipeline with explicit constraints aligns with institutional compliance requirements. Monthly rebalancing and 10 ETF universe limit high-frequency applicability but are relevant for systematic asset allocation strategies.

Requires integrating four distinct components: (1) Gaussian HMM fitting with rolling-window updates, (2) Training a conditional DDPM with UNet architecture, tail-weighted loss, and MoE denoiser (~1-2M parameters, 250k training steps), (3) Signal extraction with Ledoit-Wolf shrinkage and moment blending, (4) Solving a convex CVaR epigraph QP with O(N) constraints at each rebalance. Walk-forward protocol adds orchestration complexity. The MoE gating mechanism and tail-weighted loss require careful implementation to avoid target leakage. QP solves are fast (O(d^3 + Nd^2)) but the full pipeline is non-trivial to productionize.

The paper provides detailed hyperparameters (UNet 8 stages, base ch=64, cosine schedule, EMA=0.999, AdamW lr=1e-4, batch=256, 250k steps, seed=2020), algorithm pseudocode (Algorithm 1), data source (Yahoo Finance Adjusted Close), and explicit constraint formulations. However, no GitHub repository or code release is mentioned. The HMM fitting details, exact ETF tickers, and full training code are not provided. Theoretical proofs are in the appendix but some are sketches.