Paper: arXiv 2609.35086

Authors: Kelvin J. L. Koa, Xinyang Li, Ke-Wei Huang

Abstract

In this work, we study portfolio optimization under the stochastic discount factor (SDF) framework by learning market state representations that capture the underlying risk structures of financial data. This is challenging due to several factors: financial markets exhibit non-stationary dynamics with shifting regimes, multimodal inputs such as price and news data often contain stochastic noise, and existing diffusion-based approaches, while effective for modeling stochastic dynamics, rely on assumptions such as isotropic Gaussian noise that fail to capture the state-dependent nature of financial uncertainty. To address these challenges, we introduce RADAR, a retrieval-augmented diffusion framework that learns market representations by conditioning on similar historical regimes. RADAR leverages retrieval to construct context-dependent noise distributions, applies conditional diffusion to denoise multimodal representations, and initializes the diffusion process using empirical statistics to reflect state-dependent uncertainty. Experiments show that RADAR achieves state-of-the-art performance on key risk-adjusted metrics while producing economically meaningful signals on asset returns and correlations.

Complexity vs Empirical Score

  • Math Complexity: 7.5/10
  • Empirical Rigor: 6.0/10 (capped from 8.0: transaction costs not modelled; see penalty flags)
  • Quadrant: Holy Grail — high math complexity, high empirical rigor (effective empirical rigor capped from 8.0)

Why this score: This paper presents a highly novel approach by integrating retrieval-augmented diffusion models for SDF-based portfolio optimization, addressing key challenges in financial modeling. It demonstrates strong empirical results with state-of-the-art performance and provides code for reproducibility, making it a significant contribution.

Research Flowchart

  flowchart TD
    A["Research Goal: Portfolio Optimization under SDF"] --> B{"Challenges: Non-stationary, Multimodal Noise, Isotropic Diffusion Limits"};

    B --> C["Methodology: RADAR Framework"];

    C --> D{"Data/Inputs: Price Data, News Data"};

    D --> E["Computational Processes: Retrieval-Augmented Diffusion, Conditional Diffusion, Empirical Statistics Init."];

    E --> F["Key Findings/Outcomes: SOTA Risk-Adjusted Metrics, Economically Meaningful Signals"];

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 RADAR, a retrieval-augmented diffusion framework that learns context‑aware market representations for SDF‑based portfolio optimization.

Source sentence
RADAR, a retrieval-augmented diffusion framework that learns context-aware market representations by conditioning on similar historical regimes to denoise multimodal features.

Assumes 5

  • Financial markets exhibit non‑stationary dynamics with shifting regimes. market
    Source sentence
    financial markets exhibit non-stationary dynamics with shifting regimes
  • Multimodal inputs such as price and news data contain stochastic noise. data
    Source sentence
    multimodal inputs such as price and news data often contain stochastic noise
  • Standard diffusion models assume isotropic Gaussian noise, which does not capture state‑dependent uncertainty. model
    Source sentence
    existing diffusion‑based approaches, while effective for modeling stochastic dynamics, rely on assumptions such as isotropic Gaussian noise that fail to capture the state‑dependent nature of financial uncertainty
  • A rolling‑window protocol with 4‑year training windows avoids look‑ahead bias. execution
    Source sentence
    we adopt a rolling‑window protocol: each window uses 4 years of data for training (with a 90:10 train/validation split) and the subsequent 1 year for out‑of‑sample testing, rolling forward by 1 year
Show 1 more
  • The diffusion noise distribution can be conditioned on retrieved historical contexts. model
    Source sentence
    We introduce a state‑dependent diffusion formulation where the noise distribution is conditioned on retrieved historical contexts

Data

Backtest sample: 2011-06 – 2020-06 · S&P 500 constituent stocks · equities · daily

  • S&P 500 constituent stocks dataset · Yahoo Finance and FNSPID market-prices public
    June 2011 to June 2020 · 377 U.S. equities that appear in both the S&P 500 index and the news corpus
    Source sentence
    We collect daily adjusted close prices from Yahoo Finance and financial news articles from FNSPID [10], and the final dataset includes 377 U.S. equities that appear in both the S&P 500 index and the news corpus. The duration spans from June 2011 to June 2020, covering approximately 2,270 trading days.

Recognised sources: FNSPID Yahoo Finance

Costs

Not modeled A trading or execution result is reported without costs.

Finds 4

Reported performance (RADAR): Sharpe 1.06 (benchmark 0.51) · annual return 28.5% · total return 250.2% · max drawdown 30.8%

  • RADAR achieves state‑of‑the‑art performance on key risk‑adjusted metrics such as Sharpe, Sortino, and Calmar ratios. performance
    Source sentence
    RADAR consistently achieves state‑of‑the‑art performance across key risk‑adjusted metrics, including Sharpe, Sortino, and Calmar ratios.
  • Quintile portfolios formed from higher‑scoring assets consistently outperform lower‑scoring ones, indicating economically meaningful representations. performance
    Source sentence
    As shown, the NAV curves exhibit a clear separation, with Q1 consistently outperforming the lower quintiles. This indicates that the RADAR embeddings capture economically meaningful cross‑sectional asset returns signals that translate into improved portfolio performance.
  • RADAR consistently attains the highest Sharpe ratio and cumulative return across various input lengths and rebalancing horizons. performance
    Source sentence
    RADAR consistently achieves the highest Sharpe ratio and cumulative return across all six configurations.
  • The method generalizes well to multiple non‑financial domains, often matching or exceeding baselines. performance
    Source sentence
    Overall, we find that RADAR is largely generalizable across multiple domains.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 48 related papers). Coverage is partial, so this is neither replication nor refutation.

Cited by 1 paper in the archive

Code

https://github.com/xinyangli5579-star/RADAR — the authors' own repository

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

Cited References (59)

  1. The Utility of Wealth (1952) — link
  2. Portfolio Selection (1952) — link
  3. Large (and deep) factor models — → 2402.06635
  4. Artificial intelligence asset pricing models (2025)
  5. Deep attentive learning for stock movement prediction from social media text and company correlations (2020)
  6. Listening to chaotic whispers: A deep learning framework for news-oriented stock trend prediction (2018)
  7. Diffusion variational autoencoder for tackling stochasticity in multi-step regression stock price prediction — → 2309.00073
  8. A new approach to the economic analysis of nonstationary time series and the business cycle (1989)
and 51 more
  1. Giving content to investor sentiment: The role of media in the stock market (2007)
  2. Diffsformer: A diffusion transformer on stock factor augmentation — → 2402.06656
  3. Autoregressive denoising diffusion models for multivariate probabilistic time series forecasting (2021)
  4. Deep learning for event-driven stock prediction. (2015)
  5. A dual-stage attention-based recurrent neural network for time series prediction (2017) — link
  6. Efficient capital markets: A review of theory and empirical work (1970)
  7. Learning to generate explainable stock predictions using self-reflective large language models — → 2402.03659
  8. Stock movement prediction from tweets and historical prices (2018)
  9. A Random Walk Down Wall Street (1973)
  10. Enhancing stock movement prediction with adversarial training (2018) — link
  11. Building diversified portfolios that outperform out-of-sample (2016)
  12. A deep reinforcement learning framework for the financial portfolio management problem (2017) — link
  13. Alphastock: A buying-winners-and-selling-losers investment strategy using interpretable deep reinforcement attention networks (2019)
  14. Learning embedded representation of the stock correlation matrix using graph machine learning (2024)
  15. Clustering and Similarity Learning in Financial Markets: A Tutorial for the Practitioners (2025)
  16. Asset embeddings (2025)
  17. The role of conditioning information in deducing testable restrictions implied by dynamic asset pricing models (1987)
  18. Honey, I shrunk the sample covariance matrix (2003)
  19. Finbert: Financial sentiment analysis with pre-trained language models (2019) — link
  20. NGAT: A Node-level Graph Attention Network for Long-term Stock Prediction — → 2507.02018
  21. itransformer: Inverted transformers are effective for time series forecasting (2023) — link
  22. Parametric portfolio policies: Exploiting characteristics in the cross-section of equity returns (2009)
  23. Apt or “aipt”? the surprising dominance of large factor models (2024)
  24. Digesting anomalies: An investment approach (2015)
  25. Mispricing factors (2017)
  26. Short-and long-horizon behavioral factors (2020)
  27. Fnspid: A comprehensive financial news dataset in time series — → 2402.06698
  28. Optimal versus naive diversification: How inefficient is the 1/N portfolio strategy? (2009)
  29. Fin-r1: A large language model for financial reasoning through reinforcement learning (2025) — link
  30. Time-mmd: Multi-domain multimodal dataset for time series analysis (2024)
  31. Prompt leakage effect and mitigation strategies for multi-turn LLM applications (2024)
  32. A unified approach to interpreting model predictions (2017)
  33. " Why should i trust you?" Explaining the predictions of any classifier (2016)
  34. A simple, positive semi-definite, heteroskedasticity and autocorrelationconsistent covariance matrix (1986)
  35. Comparing predictive accuracy (2002)
  36. Robust performance hypothesis testing with the Sharpe ratio (2008)
  37. The stationary bootstrap (1994)
  38. Statistical methods for research workers. (1934)
  39. Retrieval-augmented diffusion models for time series forecasting (2024)
  40. Pattern recognition and machine learning (2006)
  41. Machine learning: a probabilistic perspective (2012)
  42. Non-stationary transformers: Exploring the stationarity in time series forecasting (2022)
  43. Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting (2021)
  44. Denoising diffusion probabilistic models (2020)
  45. Blurring Diffusion Models (2023)
  46. Diffusion models with learned adaptive noise (2024)
  47. Neural flow diffusion models: Learnable forward process for improved diffusion modelling (2024)
  48. Neural diffusion models (2023) — link
  49. Conditional Diffusion Model with Nonlinear Data Transformation for Time Series Forecasting (2025)
  50. Non-stationary Diffusion For Probabilistic Time Series Forecasting (2025)
  51. Transformer-modulated diffusion models for probabilistic multivariate time series forecasting (2024)

6 of 59 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.