Paper: arXiv 2510.07099

Authors: Himanshu Choudhary, Arishi Orra, Manoj Thakur

Abstract

In the ever-changing and intricate landscape of financial markets, portfolio optimisation remains a formidable challenge for investors and asset managers. Conventional methods often struggle to capture the complex dynamics of market behaviour and align with diverse investor preferences. To address this, we propose an innovative framework, termed Diffusion-Augmented Reinforcement Learning (DARL), which synergistically integrates Denoising Diffusion Probabilistic Models (DDPMs) with Deep Reinforcement Learning (DRL) for portfolio management. By leveraging DDPMs to generate synthetic market crash scenarios conditioned on varying stress intensities, our approach significantly enhances the robustness of training data. Empirical evaluations demonstrate that DARL outperforms traditional baselines, delivering superior risk-adjusted returns and resilience against unforeseen crises, such as the 2025 Tariff Crisis. This work offers a robust and practical methodology to bolster stress resilience in DRL-driven financial applications.

Complexity vs Empirical Score

  • Math Complexity: 7.5/10
  • Empirical Rigor: 6.5/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper involves advanced mathematics including Diffusion Probabilistic Models and deep reinforcement learning (PPO) with formal loss functions and conditioning variables, but provides empirical validation on real-world financial data (Dow 30 stocks) with transaction costs and clear out-of-sample testing, indicating a balance of theoretical and practical rigor.

Research Flowchart

  flowchart TD
  G["Research Goal:<br>Enhance Portfolio Robustness<br>under Stress Scenarios"] --> D["Inputs: Historical Market Data<br>Investor Preferences"]
  D --> M1["Method: Diffusion Model (DDPM)<br>Generate Synthetic Crash Scenarios"]
  D --> M2["Method: Deep Reinforcement Learning<br>Train Policy Agent"]
  M1 --> C["Integration: Condition DRL<br>on Stress Scenarios"]
  M2 --> C
  C --> F["Outcome: DARL Framework<br>Superior Risk-Adjusted Returns<br>Resilience to 2025 Tariff Crisis"]

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 the Diffusion-Augmented Reinforcement Learning (DARL) framework that integrates Denoising Diffusion Probabilistic Models with Deep Reinforcement Learning for portfolio management.

Source sentence
we propose an innovative framework, termed Diffusion-Augmented Reinforcement Learning (DARL), which synergistically integrates Denoising Diffusion Probabilistic Models (DDPMs) with Deep Reinforcement Learning (DRL) for portfolio management.

Assumes 4

  • Classical portfolio methods assume normally distributed returns and stationarity of data. statistical
    Source sentence
    However, these classical methods rely on strong assumptions such as normally distributed returns and stationarity of data.
  • Synthetic crash scenarios can be generated by conditioning DDPMs on stress intensity. model
    Source sentence
    By leveraging DDPMs to generate synthetic market crash scenarios conditioned on varying stress intensities
  • A transaction cost of 0.05% is applied to every trade. execution
    Source sentence
    a transaction cost of 0.05% is charged on every trade to mirror market frictions.
  • The training data consists of daily closing prices for the Dow 30 stocks from 2011 to 2023. data
    Source sentence
    Daily closing prices for all DJI stocks were retrieved from Yahoo Finance, covering 1 January 2011 to 31 July 2025. For model development, the training window spans 1 January 2011 to 31 December 2023

Data

Backtest sample: 2024-01 – 2025-07 · Dow Jones Industrial Average stocks · equities · daily

  • Dow Jones Industrial Average (DJI) 30 constituent stocks · Yahoo Finance market-prices public
    2011-01-01 to 2025-07-31 · 30 constituent stocks
    Source sentence
    We used data from the Dow Jones Industrial Average (DJI), comprising its 30 constituent stocks. Daily closing prices for all DJI stocks were retrieved from Yahoo Finance, covering 1 January 2011 to 31 July 2025.

Recognised sources: Yahoo Finance

Costs

Modeled Costs are part of the reported headline results.

0.05% per trade

Source sentence
a transaction cost of 0.05% is charged on every trade to mirror market frictions.

Finds 4

Reported performance (Diffusion-Augmented Reinforcement Learning (DARL)): Sharpe 1.91 (benchmark 0.77) · annual return 34.7% · total return 59.5% · max drawdown 15.8%

  • DARL achieved the highest cumulative return of 59.52% and annualized return of 34.71%, outperforming the market index (17.88% cumulative, 11.07% annualized). performance
    Source sentence
    It achieved the highest cumulative return (59.52%) and annualized return (34.71%), substantially surpassing the market index (17.88% cumulative, 11.07% annualized).
  • DARL attained a Sharpe ratio of 1.91 and a Calmar ratio of 2.20, significantly higher than competing models. performance
    Source sentence
    In terms of risk-adjusted performance, DARL demonstrated a Sharpe ratio of 1.91 and a Calmar ratio of 2.20, which are significantly higher than those of competing models.
  • The maximum drawdown of DARL was -15.76%, lower than Hybrid GA (-16.79%) and the Without Augmentation setup (-20.30%). performance
    Source sentence
    the proposed model recorded a maximum drawdown of −15.76%, which is relatively lower than Hybrid GA (−16.79%) and Without Augmentation (−20.30%)
  • During volatile market phases (Tariff Announcement), DARL showed faster recovery and more stable growth compared to traditional approaches. performance
    Source sentence
    During volatile market phases (Tariff Announcement), the proposed displayed faster recovery and more stable growth, whereas traditional approaches like OLMAR, Hybrid GA, and the index suffered sharp declines and slower recoveries.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 65 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 (17)

  1. Enhancing deep reinforcement learning for stock trading: a reward shaping approach via expert feedback (2025) — link
  2. FinXplore: An Adaptive Deep Reinforcement Learning Framework for Balancing and Discovering Investment Opportunities (2025) — link
  3. Risk-Adjusted Deep Reinforcement Learning for Portfolio Optimization: A Multi-reward Approach (2025) — link
  4. Deep Reinforcement Learning for Investor-Specific Portfolio Optimization: A Volatility-Guided Asset Selection Approach — → 2505.03760
  5. Dynamic Reinforced Ensemble using Bayesian Optimization for Stock Trading (2024) — link
  6. New Paradigm for Economic and Financial Research With Generative AI: Impact and Perspective (2024) — link
  7. FinRL-Meta: Market Environments and Benchmarks for Data-Driven Financial Reinforcement Learning (2022) — link
  8. Scenario generation for market risk models using generative neural networks (2021) — link
and 9 more
  1. Denoising Diffusion Probabilistic Models (2020) — link
  2. Quant GANs: deep generation of financial time series (2019) — link
  3. Proximal Policy Optimization Algorithms (2017) — link
  4. A Deep Reinforcement Learning Framework for the Financial Portfolio Management Problem (2017) — link
  5. A portfolio optimization model with three objectives and discrete variables (2010) — link
  6. Portfolio Selection with Transaction Costs (1990) — link
  7. Portfolio Selection (1971) — link
  8. Reinforcement learning: An introduction, Cambridge (1998)
  9. International Diversification of Investment Portfolios (1970)

1 of 17 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.