Paper: arXiv 2502.00828

Authors: Yoontae Hwang, Yaxuan Kong, Stefan Zohren, Yongjae Lee

Abstract

This paper addresses the critical disconnect between prediction and decision quality in portfolio optimization by integrating Large Language Models (LLMs) with decision-focused learning. We demonstrate both theoretically and empirically that minimizing the prediction error alone leads to suboptimal portfolio decisions. We aim to exploit the representational power of LLMs for investment decisions. An attention mechanism processes asset relationships, temporal dependencies, and macro variables, which are then directly integrated into a portfolio optimization layer. This enables the model to capture complex market dynamics and align predictions with the decision objectives. Extensive experiments on S&P100 and DOW30 datasets show that our model consistently outperforms state-of-the-art deep learning models. In addition, gradient-based analyses show that our model prioritizes the assets most crucial to decision making, thus mitigating the effects of prediction errors on portfolio performance. These findings underscore the value of integrating decision objectives into predictions for more robust and context-aware portfolio management.

Complexity vs Empirical Score

  • Math Complexity: 8.0/10
  • Empirical Rigor: 6.0/10 (capped from 7.0: transaction costs not modelled; see penalty flags)
  • Quadrant: Holy Grail — high math complexity, high empirical rigor (effective empirical rigor capped from 7.0)

Why this score: The paper employs advanced mathematics including attention mechanisms, gradient-based analysis for differentiating through convex optimization, and hybrid loss functions combining statistical and decision objectives. It demonstrates empirical rigor through extensive experiments on S&P100 and DOW30 datasets, comparing against state-of-the-art models, and includes performance metrics and gradient-based interpretability analyses.

Research Flowchart

  flowchart TD
  A["Research Goal: Improve Portfolio Optimization by aligning predictions with decisions"] --> B{"Data & Inputs"};
  B --> B1["S&P100 & DOW30<br>Price Data"];
  B --> B2["Asset Relationships<br>Macro Variables"];
  B --> B3["Historical Outcomes"];
  
  B --> C["Methodology: LLM-Decision Focused Learning"];
  C --> D["Attention Mechanism<br>Processes temporal & relational data"];
  D --> E["Decision Layer<br>Portfolio Optimization Layer"];
  
  E --> F{"Computational Process"};
  F --> F1["Predict Asset Returns"];
  F --> F2["Optimize Portfolio Allocation"];
  F --> F3["Compute Decision Loss"];
  
  F --> G["Key Findings & Outcomes"];
  G --> G1["Outperforms SOTA Models"];
  G --> G2["Gradient Analysis shows focus on crucial assets"];
  G --> G3["Minimizes prediction error impact on decisions"];
  
  style A fill:#e1f5fe
  style C fill:#fff3e0
  style G fill:#e8f5e8

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 a decision-informed neural network framework that integrates large language models for portfolio optimization, using attention mechanisms and a hybrid loss.

Source sentence
We develop a decision-informed neural network framework that integrates LLMs for portfolio optimization.

Assumes 2

  • Only companies that maintained consistent index membership throughout the study period are included to avoid survivorship bias. data
    Source sentence
    To address potential survivorship bias, we include only companies that maintained consistent index membership throughout the study period.
  • Macroeconomic indicators from FRED are predictive of asset returns. market
    Source sentence
    selected based on their documented predictive power in asset pricing: weekly initial jobless claims (ICSA), consumer sentiment (UMCSENT), new home sales (HSN1F), unemployment rate (UNRATE), and high-yield bond spread (HYBS).

Data

Backtest sample: S&P 100 and DOW 30 stocks · equities · daily

  • DOW 30 equity returns · WRDS market-prices commercial
    January 2010 to December 2023 · DOW 30 constituents
    Source sentence
    Our primary data consists of equity returns from two major indices: the DOW 30 and a market-cap-weighted subset of 50 constituents from the S&P 100.
  • S&P 100 subset equity returns · WRDS market-prices commercial
    January 2010 to December 2023 · 50 constituents from the S&P 100
    Source sentence
    Our primary data consists of equity returns from two major indices: the DOW 30 and a market-cap-weighted subset of 50 constituents from the S&P 100.
  • Five macroeconomic indicators · FRED macro public
    January 2010 to December 2023
    Source sentence
    The financial data, obtained from WRDS, is complemented by five macroeconomic indicators from FRED, selected based on their documented predictive power in asset pricing.

Recognised sources: FRED WRDS

Costs

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

Finds 4

Reported performance (Decision-Informed Neural Network (DINN)): Sharpe 1.29 · annual return 63.2% · max drawdown 39.5%

  • DINN consistently outperforms state-of-the-art deep learning models on both S&P 100 and DOW 30 datasets. performance
    Source sentence
    Extensive experiments on S&P100 and DOW30 datasets show that our model consistently outperforms state-of-the-art deep learning models.
  • During the COVID-19 period, DINN-selected portfolios achieve Sharpe ratios of 1.14 (S&P 100) and 0.87 (DOW 30), higher than non-selected portfolios (0.59 and 0.58). performance
    Source sentence
    During the COVID-19 period, DINN-selected portfolios achieve an SR of 1.14 for the S&P 100 and 0.87 for the DOW 30, exceeding the SR of portfolios formed from non-selected stocks (0.59 and 0.58, respectively).
  • High‑sensitivity assets (large gradient |∂ŵ/∂µ|) exhibit lower mean‑squared error than low‑sensitivity assets, indicating the model learns what matters. predictive-accuracy
    Source sentence
    high-sensitivity assets actually show lower MSE. This occurs because DINN allocates greater learning capacity to stocks where prediction errors would result in higher decision-related costs.
  • The prob‑sparse attention mechanism helps the model focus on a smaller subset of assets, yielding lower drawdowns and higher risk‑adjusted performance in stress regimes. performance
    Source sentence
    The inclusion of a prob-sparse attention mechanism may helps the model identify and emphasize a smaller subset of assets critical to replicating market dynamics under a variety of macroeconomic conditions.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 57 related papers). Coverage is partial, so this is neither replication nor refutation.

Cited by 9 papers in the archive (3 by independent authors)

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 (69)

  1. Portfolio Selection (1952) — link
  2. The Markowitz optimization enigma: Is ‘optimized’optimal? (1989)
  3. Optimal versus naive diversification: How inefficient is the 1/N portfolio strategy? (2009)
  4. The effect of errors in means, variances, and covariances on optimal portfolio choice (1993)
  5. A well-conditioned estimator for large-dimensional covariance matrices (2004)
  6. Improved estimation of the covariance matrix of stock returns with an application to portfolio selection (2003)
  7. Empirical asset pricing via machine learning (2020)
  8. Smart “predict, then optimize” (2022)
and 61 more
  1. The dynamics of emerging market equity flows (2002)
  2. Asset allocation under multivariate regime switching (2007)
  3. Optnet: Differentiable optimization as a layer in neural networks (2017)
  4. Differentiable Convex Optimization Layers (2019)
  5. Task-based end-to-end model learning in stochastic optimization (2017)
  6. Characteristics are covariances: A unified model of risk and return (2019)
  7. Temporal Representation Learning for Stock Similarities and Its Applications in Investment Management — → 2407.13751
  8. Time-LLM : Time series forecasting by reprogramming large language models (2024)
  9. Chronos: Learning the Language of Time Series (2024) — link
  10. Integrating prediction in mean-variance portfolio optimization (2023)
  11. Chatgpt-based investment portfolio selection — → 2308.06260
  12. Attention is all you need (2017)
  13. Decision-focused learning: Foundations, state of the art, benchmark and future opportunities (2024)
  14. Gpt-4 technical report (2023) — link
  15. An Overview of Optimization Models for Portfolio Management (2024)
  16. An Overview of Machine Learning for Portfolio Optimization. (2024)
  17. End-to-end, decision-based, cardinality-constrained portfolio optimization (2025)
  18. Mean-variance optimization for asset allocation (2021)
  19. The effects of errors in means, variances, and correlations on the mean-variance framework (2022)
  20. A Survey of Large Language Models for Financial Applications: Progress, Prospects and Challenges — → 2406.11903
  21. Reversible instance normalization for accurate time-series forecasting against distribution shift (2021)
  22. FEDformer : Frequency enhanced decomposed transformer for long-term series forecasting (2022)
  23. Autoformer: Decomposition Transformers with Auto-Correlation for Long-Term Series Forecasting (2021)
  24. Cost of Asset Allocation in Equity Market: How Much Do Investors Lose Due to Bad Asset Class Design? (2014)
  25. One Fits All : Power General Time Series Analysis by Pretrained LM (2023)
  26. Estimation of Large Financial Covariances: A Cross-Validation Approach (2020) — link
  27. Do return prediction models add economic value? (2012)
  28. Canonical portfolios: Optimal asset and signal combination (2023)
  29. TEMPO : Prompt-based Generative Pre-trained Transformer for Time Series Forecasting (2024) — link
  30. Language models are unsupervised multitask learners (2019)
  31. The llama 3 herd of models (2024) — link
  32. On the sensitivity of mean-variance-efficient portfolios to changes in asset means: some analytical and computational results (1991)
  33. On portfolio optimization: Forecasting covariances and choosing the risk model (1999)
  34. Risk reduction in large portfolios: Why imposing the wrong constraints helps (2003)
  35. Portfolio selection with parameter and model uncertainty: A multi-prior approach (2007)
  36. Machine learning and portfolio optimization (2018)
  37. Nonlinear shrinkage of the covariance matrix for portfolio selection: Markowitz meets Goldilocks (2017)
  38. The Gerber Statistic: A Robust Co-Movement Measure for Portfolio Optimization. (2022)
  39. Application of robust statistics to asset allocation models (2007)
  40. Taming the factor zoo: A test of new factors (2020)
  41. Capital asset prices: A theory of market equilibrium under conditions of risk (1964)
  42. The valuation of risk assets and the selection of risky investments in stock portfolios and capital budgets (1975)
  43. Common risk factors in the returns on stocks and bonds (1993)
  44. A five-factor asset pricing model (2015)
  45. A fast algorithm for the minimum covariance determinant estimator (1999)
  46. Minimum volume ellipsoid (2009)
  47. Distributionally robust end-to-end portfolio construction (2023)
  48. The matrix cookbook (2008)
  49. Factor models, machine learning, and asset pricing (2022)
  50. A universal end-to-end approach to portfolio optimization via deep learning (2021) — link
  51. Deep learning in asset pricing (2024)
  52. iTransformer: Inverted Transformers Are Effective for Time Series Forecasting (2024)
  53. A Time Series is Worth 64 Words: Long-term Forecasting with Transformers (2023)
  54. TimesNet: Temporal 2D-Variation Modeling for General Time Series Analysis (2023)
  55. Crossformer: Transformer Utilizing Cross-Dimension Dependency for Multivariate Time Series Forecasting (2023)
  56. Are language models actually useful for time series forecasting? (2024)
  57. One fits all: Power general time series analysis by pretrained lm (2023)
  58. Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting (2021)
  59. 60 years of portfolio optimization: Practical challenges and current trends (2014)
  60. The effects of estimation error on measures of portfolio credit risk (2003)
  61. Parameter uncertainty in portfolio selection: Shrinking the inverse covariance matrix (2012)

3 of 69 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.