Paper: arXiv 2510.10807

Authors: Ali Atiah Alzahrani

Abstract

We examine whether regime-conditioned generative scenarios combined with a convex CVaR allocator improve portfolio decisions under regime shifts. We present MARCD, a generative-to-decision framework with: (i) a Gaussian HMM to infer latent regimes; (ii) a diffusion generator that produces regime-conditioned scenarios; (iii) signal extraction via blended, shrunk moments; and (iv) a governed CVaR epigraph quadratic program. Contributions: Within the Scenario stage we introduce a tail-weighted diffusion objective that up-weights low-quantile outcomes relevant for drawdowns and a regime-expert (MoE) denoiser whose gate increases with crisis posteriors; both are evaluated end-to-end through the allocator. Under strict walk-forward on liquid multi-asset ETFs (2005-2025), MARCD exhibits stronger scenario calibration and materially smaller drawdowns: MaxDD 9.3% versus 14.1% for BL (a 34% reduction) over 2020-2025 out-of-sample. The framework provides an auditable pipeline with explicit budget, box, and turnover constraints, demonstrating the value of decision-aware generative modeling in finance.

Complexity vs Empirical Score

  • Math Complexity: 7.5/10
  • Empirical Rigor: 8.5/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper employs advanced mathematics including diffusion models, Gaussian HMMs, mixture-of-experts denoisers, and convex optimization with CVaR epigraph QP, supported by theoretical theorems and proofs. Empirically, it demonstrates rigorous walk-forward backtesting on real ETF data (2005-2025) with explicit constraints, transaction costs, and detailed out-of-sample metrics, showing a 34% reduction in max drawdown.

Research Flowchart

  flowchart TD
  A["Research Goal: Assess Regime-Conditioned Generative Models for CVaR-Constrained Portfolios under Shifts"] --> B["Data: Liquid Multi-Asset ETFs (2005-2025)"]

  subgraph C ["MARCD Methodology"]
      C1["Gaussian HMM<br>Latent Regime Inference"]
      C2["Diffusion Generator<br>Tail-Weighted &<br>Regime-Conditioned Scenarios"]
      C3["Signal Extraction<br>Blended & Shrunk Moments"]
      C4["CVaR Epigraph QP<br>Explicit Constraints"]
  end

  B --> C1
  C1 --> C2
  C2 --> C3
  C3 --> C4

  C4 --> D["Walk-Forward Validation<br>Out-of-Sample: 2020-2025"]

  D --> E["Key Findings"]
  E --> E1["Superior Scenario Calibration"]
  E --> E2["Max Drawdown: 9.3% (vs. 14.1% for BL)"]
  E --> E3["34% Reduction in Max Drawdown"]
  E --> E4["Validated Decision-Aware Generative Modeling"]

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes MARCD, a regime‑conditioned generative‑to‑decision pipeline that combines a Gaussian HMM, a diffusion scenario generator, blended moment extraction, and a CVaR epigraph quadratic program for portfolio allocation.

Source sentence
We present MARCD, a generative-to-decision framework with: (i) a Gaussian HMM to infer latent regimes; (ii) a diffusion generator that produces regime-conditioned scenarios; (iii) signal extraction via blended, shrunk moments; and (iv) a governed CVaR epigraph QP.

Assumes 6

  • Monthly rebalancing on the last trading day with a fixed 10 bps transaction cost per trade. execution
    Source sentence
    All strategies (EW, RP, BL, MARCD) rebalance monthly on the last trading day, incur identical transaction costs of 10 bps per trade, and are subject to the same turnover controls
  • Turnover is limited by an ℓ1 cap of 0.20 on weight changes. execution
    Source sentence
    an ℓ1 cap ∥wt−wt−1∥1 ≤0.20; no penalty, κ=0, unless stated otherwise
  • The latent regime process follows a Gaussian Hidden Markov Model with a 3‑year rolling estimation window. model
    Source sentence
    Protocol. Strict walk-forward; HMM 3y rolling; scenarios conditioned on zt
  • The asset universe consists of ten liquid ETFs with daily total returns from Yahoo Finance Adjusted Close (including dividends). data
    Source sentence
    Data. Daily total returns computed from Yahoo Finance Adjusted Close (dividends included)
Show 2 more
  • The CVaR allocator uses an epigraph quadratic program with Lipschitz solution map and decision‑regret bounds. model
    Source sentence
    The CVaR epigraph QP (Sec. 3.4) has a Lipschitz solution map in (µ, Σ) and admits a decision-regret bound under moment and CVaR perturbations
  • Scenario diagnostics (KS, ES, VS, VaR, CVaR) are evaluated on the same monthly‑rebalanced, cost‑adjusted returns. statistical
    Source sentence
    Diagnostics. KS, ES, VS; LB p(|r|); VaR0.95 unconditional coverage (Kupiec UC) p; CVaR0.95 error ( bps).

Data

Backtest sample: 2020 – 2025 · liquid multi-asset ETFs · multi-asset · monthly

  • Ten liquid multi‑asset ETFs · Yahoo Finance market-prices public
    2005‑2025 · 10 liquid ETFs
    Source sentence
    Data. Daily total returns computed from Yahoo Finance Adjusted Close (dividends included)

Recognised sources: Yahoo Finance

Costs

Modeled Costs are part of the reported headline results.

A flat 10 bps cost per trade is applied to all strategies and included in reported performance.

Source sentence
All strategies (EW, RP, BL, MARCD) rebalance monthly on the last trading day, incur identical transaction costs of 10 bps per trade

Finds 5

Reported performance (MARCD): Sharpe 1.23 (benchmark 1.02) · annual return 10.3% · max drawdown 9.3%

  • MARCD achieves a MaxDD of 9.3 % versus 14.1 % for the benchmark (BL), a 34 % reduction in drawdown over the 2020‑2025 out‑of‑sample period. performance
    Source sentence
    Over 2020–2025 OOS, MaxDD is 9.3% for MARCD vs BL 14.1% and EW 21.2% (34%/56% lower; absolute 4.8%/11.9%)
  • MARCD’s Sharpe ratio is 1.23, compared with 1.02 for BL and 0.72 for EW, indicating higher risk‑adjusted returns. performance
    Source sentence
    Performance— MARCD attains higher risk-adjusted performance (Sharpe 1.23 vs. 1.02 for BL)
  • The Calmar ratio improves to 1.11 for MARCD versus 0.70 for BL, reflecting better peak‑to‑trough behavior. performance
    Source sentence
    Consistent with this, Calmar improves to 1.11 (BL 0.70; EW 0.38)
  • In ablation studies, removing regime conditioning or the CVaR term raises MaxDD to 11.3 % and 14.6 % respectively, and reduces Sharpe to around 1.12‑1.13. performance
    Source sentence
    Removing either regime conditioning or the CVaR term weakens tail control
  • Scenario diagnostics show MARCD has the lowest KS, ES, VS and CVaR error among baselines (KS 0.154, ES 0.289, VS 0.247, CVaR0.95 err 15 bps). statistical
    Source sentence
    MARCD 0.154 0.289 0.247 0.50 0.58 15

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 87 related papers). Coverage is partial, so this is neither replication nor refutation.

Code

No repository by the authors found in the paper.

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.

Cited References (44)

  1. Portfolio selection (1952)
  2. Global portfolio optimization (1992)
  3. A new approach to the economic analysis of nonstationary time series and the business cycle (1989)
  4. Dynamic linear models with Markov-switching (1994)
  5. Regime shifts in financial markets (2012)
  6. Regime switching with time-varying transition probabilities (1994)
  7. A well-conditioned estimator for large-dimensional covariance matrices (2004)
  8. Modeling the conditional distribution of interest rates as a regime-switching process (1996)
and 36 more
  1. Bayesian Econometrics (2010)
  2. Time-series generative adversarial networks (2019)
  3. Quant GANs: Deep generation of financial time series (2019)
  4. Deep learning volatility: A deep generative model approach (2021)
  5. Denoising diffusion probabilistic models (2020)
  6. Score-based generative modeling through stochastic differential equations (2021)
  7. Autoregressive denoising diffusion models for multivariate probabilistic time series forecasting (2021)
  8. CSDI: Conditional score-based diffusion models for probabilistic time series imputation (2021)
  9. Informer: Beyond efficient transformer for long sequence time-series forecasting (2023)
  10. A transformer-based framework for multivariate time-series representation learning (2021)
  11. Deep learning with long short-term memory networks for financial market predictions (2018)
  12. Distributionally robust optimization under moment uncertainty with application to data-driven problems (2010)
  13. Optimization of conditional value-at-risk (2000)
  14. Stress tests and systemic risk (2015)
  15. Robust assessment of stress tests (2014)
  16. Principles for stress testing (2018)
  17. Comprehensive Capital Analysis and Review (CCAR) (2020)
  18. Multi-agent actor-critic for mixed cooperative-competitive environments (2017)
  19. Agent-based computational finance (2006)
  20. Econophysics: Empirical facts and agent-based models (2011)
  21. Deep reinforcement learning for automated trading: A survey (2020)
  22. Generative agents: Interactive simulacra of human behavior (2023)
  23. Tail-GAN: Generating realistic extreme risk scenarios (2023)
  24. Modeling tabular data using conditional GANs (2020)
  25. Generative AI in Finance: Potential and Pitfalls (2023)
  26. Long Short-Term Memory (1997) — link
  27. An Empirical Evaluation of Generic Convolutional and Recurrent Networks for Sequence Modeling (2018) — link
  28. Temporal Fusion Transformers for Interpretable Multi-horizon Time Series Forecasting (2021) — link
  29. Predict, Refine, Synthesize: Self-Guiding Diffusion Models for Probabilistic Time Series (2023)
  30. Multi-Resolution Diffusion Models for Time Series Forecasting (2024)
  31. Generative Learning for Financial Time Series with Irregular and Scale-Invariant Patterns (2024)
  32. A Financial Time Series Denoiser Based on Diffusion Models (2024) — link
  33. Simulating Asset Prices using Conditional Time-Series GAN (2024) — link
  34. Can GANs Learn the Stylized Facts of Financial Time Series? — → 2410.09850
  35. Time-series Generative Adversarial Networks (2019) — link
  36. TimeVAE: A Variational Auto-Encoder for Multivariate Time Series Generation (2021) — link

1 of 44 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.