Paper: arXiv 2504.14345

Authors: Youngbin Lee, Yejin Kim, Juhyeong Kim, Suin Kim, Yongjae Lee

Abstract

The Black-Litterman model addresses the sensitivity issues of tra- ditional mean-variance optimization by incorporating investor views, but systematically generating these views remains a key challenge. This study proposes and validates a systematic frame- work that translates return forecasts and predictive uncertainty from Large Language Models (LLMs) into the core inputs for the Black-Litterman model: investor views and their confidence lev- els. Through a backtest on S&P 500 constituents, we demonstrate that portfolios driven by top-performing LLMs significantly out- perform traditional baselines in both absolute and risk-adjusted terms. Crucially, our analysis reveals that each LLM exhibits a dis- tinct and consistent investment style which is the primary driver of performance. We found that the selection of an LLM is therefore not a search for a single best forecaster, but a strategic choice of an investment style whose success is contingent on its alignment with the prevailing market regime. The source code and data are available at https://github.com/youngandbin/LLM-BLM.

Complexity vs Empirical Score

  • Math Complexity: 4.0/10
  • Empirical Rigor: 8.5/10
  • Quadrant: Street Traders — empirical and applied, lower math complexity

Why this score: The paper’s core innovation is a systematic framework for translating LLM outputs into Black-Litterman inputs, involving moderate mathematical concepts (Bayesian blending), but the primary focus is on empirical validation. The study demonstrates high empirical rigor through a backtest on S&P 500 constituents, hyperparameter tuning (τ), performance metrics (Sharpe ratio), and the public release of source code and data.

Research Flowchart

  flowchart TD
  subgraph S1 ["Research Goal"]
      A["Goal: Systematically generate investor views for Black-Litterman model using LLMs"]
  end

  subgraph S2 ["Methodology"]
      B["1. Generate forecasts & uncertainty from multiple LLMs<br>2. Convert into BL views & confidence levels<br>3. Construct portfolios via Black-Litterman<br>4. Backtest on S&P 500 constituents"]
  end

  subgraph S3 ["Data & Inputs"]
      C["S&P 500 constituents historical data + LLM predictions"]
  end

  subgraph S4 ["Key Findings"]
      D["- LLM-driven portfolios outperform traditional baselines<br>- Each LLM exhibits distinct, consistent investment style<br>- Performance depends on regime-alignment of LLM style"]
  end

  A --> B
  C --> B
  B --> D

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 systematic framework that translates LLM return forecasts and predictive uncertainty into investor views and confidence levels for the Black‑Litterman model.

Source sentence
This study proposes and validates a systematic framework that translates return forecasts and predictive uncertainty from Large Language Models (LLMs) into the core inputs for the Black-Litterman model: investor views and their confidence levels.

Assumes 5

  • Equilibrium returns and covariance are estimated from historical return data. data
    Source sentence
    When calculating the posterior expected returns μ, we can determine the equilibrium return vector π and the covariance matrix Σ of the stocks from historical return data.
  • The hyperparameter τ balances confidence between market equilibrium prior and LLM views. model
    Source sentence
    The primary hyperparameter in our framework is τ, which, as introduced in the variance-covariance matrix V in Equation (2), balances the confidence between the market equilibrium prior and the LLM-generated views.
  • Transaction costs are 0.1% of portfolio weight changes at each rebalancing. execution
    Source sentence
    Costs are calculated as 0.1% of the portfolio weight changes at each rebalancing and are factored into the final performance.
  • Portfolio is rebalanced every two weeks. execution
    Source sentence
    We run a single backtest over the test period (Sep 2024-Jun 2025), rebalancing the portfolio every two weeks using this fixed τ*.
Show 1 more
  • The dataset consists of the 50 largest S&P 500 constituents by market cap as of March 26, 2025. data
    Source sentence
    We use historical stock price data from the largest 50 S&P 500 constituents by market capitalization as of March 26, 2025, sourced from Yahoo Finance.

Data

Backtest sample: 2024-09 – 2025-06 · S&P 500 constituents · equities

  • Yahoo Finance S&P 500 constituents price data · Yahoo Finance market-prices public
    June 2024 to June 2025 · largest 50 S&P 500 constituents
    Source sentence
    We use historical stock price data from the largest 50 S&P 500 constituents by market capitalization as of March 26, 2025, sourced from Yahoo Finance. The total dataset spans 13 months, from June 2024 to June 2025.

Recognised sources: Yahoo Finance

Costs

Modeled Costs are part of the reported headline results.

0.1% of the portfolio weight changes at each rebalancing

Source sentence
Costs are calculated as 0.1% of the portfolio weight changes at each rebalancing and are factored into the final performance.

Finds 4

Reported performance (BLM-Llama): Sharpe 1.23 (benchmark 0.89) · annual return 26.3% · max drawdown 13.8%

  • BLM‑Qwen achieved the highest CAGR (0.2811) and BLM‑Llama the second‑highest (0.2751), both surpassing EW (0.1907) and MVO (0.0607). performance
    Source sentence
    BLM‑Qwen achieved the highest CAGR of 0.2811, closely followed by BLM‑Llama at 0.2751. Both significantly outperformed the traditional benchmarks, EW (0.1907) and MVO (0.0607).
  • BLM‑Llama delivered the top annualized Sharpe ratio (1.2286), with BLM‑Qwen second (1.0624), far exceeding EW (0.8937) and MVO (0.2793). performance
    Source sentence
    In terms of risk‑adjusted returns, measured by the annualized Sharpe Ratio, Sharpe (ann.), BLM‑Llama delivered the top performance with a ratio of 1.2286. BLM‑Qwen secured the second‑highest position at 1.0624. This indicates that both models generated superior excess returns per unit of risk, far exceeding the benchma
  • Each LLM exhibits a distinct and consistent investment style that drives performance. other
    Source sentence
    Crucially, our analysis reveals that each LLM exhibits a distinct and consistent investment style which is the primary driver of performance.
  • Portfolios using LLM‑generated views significantly outperformed traditional baselines in both absolute and risk‑adjusted terms. performance
    Source sentence
    Through a backtest on S&P 500 constituents, we demonstrate that portfolios driven by top‑performing LLMs significantly out‑perform traditional baselines in both absolute and risk‑adjusted terms.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 24 related papers). Coverage is partial, so this is neither replication nor refutation.

Cited by 4 papers in the archive (1 by independent authors)

Code

https://github.com/youngandbin/LLM-BLM — 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 (36)

  1. Portfolio Selection (1952) — link
  2. Global portfolio optimization (1992)
  3. A step-by-step guide to the Black-Litterman model: Incorporating user-specified confidence levels (2007)
  4. A hybrid approach for generating investor views in Black–Litterman model (2019)
  5. Revolutionizing finance with llms: An overview of applications and insights (2024) — link
  6. Mg-conv: A spatiotemporal multi-graph convolutional neural network for stock market index trend prediction (2022)
  7. Temporal relational ranking for stock prediction (2019)
  8. GPT-4o (2024)
and 28 more
  1. The llama 3 herd of models (2024) — link
  2. Gemma: Open models based on gemini research and technology (2024) — link
  3. Qwen2 Technical Report (2024) — link
  4. Efficient memory management for large language model serving with pagedattention (2023)
  5. The factor tau in the Black-Litterman model (2013)
  6. Capital asset prices: A theory of market equilibrium under conditions of risk (1964)
  7. An Overview of Optimization Models for Portfolio Management. (2024)
  8. Decision-informed Neural Networks with Large Language Model Integration for Portfolio Optimization — → 2502.00828
  9. Medec: A benchmark for medical error detection and correction in clinical notes (2024) — link
  10. Time-llm: Time series forecasting by reprogramming large language models (2023) — link
  11. Your AI, Not Your View: The Bias of LLMs in Investment Analysis — → 2507.20957
  12. THEME: Enhancing Thematic Investing with Semantic Stock Representations and Temporal Dynamics — → 2508.16936
  13. Enhancing portfolio optimization with transformer-GAN integration: A novel approach in the Black-Litterman framework (2024) — link
  14. Portfolio management transformed: An enhanced black–litterman approach integrating asset pricing theory and machine learning (2025)
  15. Empirical Asset Pricing with Large Language Model Agents (2024) — link
  16. Retrieval-augmented large language models for financial time series forecasting (2025) — link
  17. TradingAgents: Multi-agents LLM financial trading framework — → 2412.20138
  18. Finmem: A performance-enhanced llm trading agent with layered memory and character design — → 2311.13743
  19. Dynamic Black-Litterman — → 2404.18822
  20. The intuition behind Black-Litterman model portfolios (2002)
  21. The black-litterman approach: Original model and extensions (2010)
  22. A survey of behavioral finance (2003)
  23. Empirical asset pricing via machine learning (2020)
  24. A deep reinforcement learning framework for the financial portfolio management problem (2017) — link
  25. Bloomberggpt: A large language model for finance (2023) — link
  26. Fingpt: Democratizing internet-scale data for financial large language models (2023) — link
  27. Can ChatGPT forecast stock price movements? (2024)
  28. A survey of large language models in finance (finllms) (2024) — link

6 of 36 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.