Paper: arXiv 2508.11152

Authors: Tianjiao Zhao, Jingrao Lyu, Stokes Jones, Harrison Garber, Stefano Pasquali, Dhagash Mehta

Abstract

The field of artificial intelligence (AI) agents is evolving rapidly, driven by the capabilities of Large Language Models (LLMs) to autonomously perform and refine tasks with human-like efficiency and adaptability. In this context, multi-agent collaboration has emerged as a promising approach, enabling multiple AI agents to work together to solve complex challenges. This study investigates the application of role-based multi-agent systems to support stock selection in equity research and portfolio management. We present a comprehensive analysis performed by a team of specialized agents and evaluate their stock-picking performance against established benchmarks under varying levels of risk tolerance. Furthermore, we examine the advantages and limitations of employing multi-agent frameworks in equity analysis, offering critical insights into their practical efficacy and implementation challenges.

Complexity vs Empirical Score

  • Math Complexity: 2.5/10
  • Empirical Rigor: 6.0/10
  • Quadrant: Street Traders — empirical and applied, lower math complexity

Why this score: The paper relies primarily on LLM prompt engineering and system design rather than advanced mathematical derivations, resulting in low math complexity; however, it includes a systematic backtest against benchmarks with specific performance metrics and risk profiles, demonstrating substantial empirical rigor.

Research Flowchart

  flowchart TD
  A["Research Goal: LLM Multi-Agents for Equity Portfolio Construction"] --> B["Methodology: Role-Based Multi-Agent System"]
  B --> C["Inputs: Financial Data & Market News"]
  C --> D["Process: Collaborative Stock Selection"]
  D --> E["Process: Portfolio Construction vs Benchmarks"]
  E --> F["Outcomes: Stock-Picking Performance & Risk Analysis"]
  F --> G["Conclusion: Efficacy & Challenges of Multi-Agent Frameworks"]

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 novel multi‑agent investment framework that uses LLM‑powered collaboration and debate to collect and synthesize equity research inputs.

Source sentence
In this work, we propose a novel multi-agent investment framework that leverages LLM-powered collaboration and debate to collect and synthesize equity research inputs.

Assumes 2

  • Each agent only accesses data relevant to its specific role and assigned task. data
    Source sentence
    Each agent is granted access only to the data relevant to its specific role and assigned task.
  • Risk‑averse and risk‑neutral agent behavior is modeled via prompt engineering that embeds investor traits into agent instructions. model
    Source sentence
    We design agents for both risk‑averse and risk‑neutral scenarios using prompt engineering to embed investor traits directly into agent instructions.

Data

Backtest sample: 2024-02 – 2024-06 · equities

  • 10‑K and 10‑Q financial disclosures fundamentals public
    Source sentence
    10K/10Q Fundamental Agent Financial disclosure data (includes discussion of major company prospectus and financial statements).
  • Bloomberg financial news · Bloomberg text-news commercial
    Source sentence
    Bloomberg ID Sentiment Agent Ticker’s corresponding unique identifier for Bloomberg News
  • Historical stock price and volume data · Yahoo Finance market-prices public
    Source sentence
    Price (Open, High, Low, Close) Valuation Agent The initial, highest, lowest, and final trading price of a stock within a specific trading period

Recognised sources: Bloomberg Yahoo Finance

Costs

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

Finds 3

  • In the risk‑neutral setting, the multi‑agent framework outperforms both the benchmark and single‑agent frameworks during the testing window. performance
    Source sentence
    In the risk-neutral setting, the proposed multi-agent outperforms the benchmark as well as the single agent framework during the testing window.
  • In the risk‑averse setting, all agent‑selected portfolios underperform the benchmark, partly due to exclusion of volatile technology stocks. performance
    Source sentence
    In the risk-averse setting, all agent-selected portfolios followed a conservative investment strategy, leading to underperformance relative to the benchmark.
  • The multi‑agent portfolio shows slightly lower volatility and reduced drawdowns compared to individual valuation and fundamental portfolios. performance
    Source sentence
    the multi-agent portfolio exhibited slightly lower volatility and reduced drawdowns, indicating an ability to mitigate losses more effectively than the individual strategies.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 22 related papers). Coverage is partial, so this is neither replication nor refutation.

Cited by 8 papers in the archive (7 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 (125)

  1. MAPS: Multi-agent reinforcement learning-based portfolio management system (2020) — link
  2. Designing a multi-agent portfolio management system (1996)
  3. Evaluating multi-agent coordination abilities in large language models (2023) — link
  4. Multi-agent deep reinforcement learning algorithm with trend consistency regularization for portfolio management (2023)
  5. FinMem: A Performance-Enhanced LLM Trading Agent with Layered Memory and Character Design — → 2311.13743
  6. A Multimodal Foundation Agent for Financial Trading: Tool-Augmented, Diversified, and Generalist — → 2402.18485
  7. Self-Reflection in LLM Agents: Effects on Problem-Solving Performance (2024) — link
  8. Emotional finance: The role of the unconscious in financial decisions (2010)
and 117 more
  1. Behavioral finance and portfolio management: Review of theory and literature (2020)
  2. Applications of Client Behavior: A Practitioner’s Perspective (2017)
  3. Behavioral finance-theory and practical application: Systematic analysis of departures from the homo oeconomicus paradigm are essential for realistic financial research and analysis (2001)
  4. Multi-agent collaboration: Harnessing the power of intelligent llm agents (2023) — link
  5. Herding in the Italian stock market: a case of behavioral finance (2004)
  6. Financial decision-making in markets and firms: A behavioral perspective (1995)
  7. Practical Challenges of Implementing Behavioral Finance: Reflections from the Field (2017)
  8. Behavioral finance: investors, corporations, and markets (2010)
  9. Prospect theory: An analysis of decision under risk (2013)
  10. Autogen: Enabling next-gen llm applications via multi-agent conversation framework (2023) — link
  11. More agents is all you need (2024) — link
  12. Marketsenseai 2.0: Enhancing stock analysis through llm agents — → 2502.00415
  13. Llm-planner: Few-shot grounded planning for embodied agents with large language models (2023)
  14. Api-bank: A comprehensive benchmark for tool-augmented llms (2023) — link
  15. Towards Efficient LLM Grounding for Embodied Multi-Agent Collaboration (2024) — link
  16. Medagents: Large language models as collaborators for zero-shot medical reasoning (2023) — link
  17. Communicative agents for software development (2023) — link
  18. Metagpt: Meta programming for multi-agent collaborative framework (2023) — link
  19. A survey on in-context learning (2022) — link
  20. Impact of risk tolerance and demographic factors on financial investment decision (2018)
  21. FinRobot: An Open-Source AI Agent Platform for Financial Applications using Large Language Models — → 2405.14767
  22. Llm-based nlg evaluation: Current status and challenges (2024) — link
  23. Retrieval-augmented generation for large language models: A survey (2023) — link
  24. Videoagent: Long-form video understanding with large language model as agent (2024) — link
  25. Bloomberggpt: A large language model for finance (2023) — link
  26. Stock ranking prediction using list-wise approach and node embedding technique (2021)
  27. Minilm: Deep self-attention distillation for task-agnostic compression of pre-trained transformers (2020)
  28. Universal language model fine-tuning for text classification (2018) — link
  29. Improving factuality and reasoning in language models through multiagent debate (2023) — link
  30. Attention is all you need (2017)
  31. Bert: Pre-training of deep bidirectional transformers for language understanding (2018) — link
  32. Recommender systems (2016)
  33. Categorizing mutual funds using clusters (1999)
  34. Scikit-learn: Machine learning in Python (2011)
  35. “Morningstar Categorization.” (2018) — link
  36. Recommender systems (1997)
  37. Clustering financial time series: an application to mutual funds style analysis (2004)
  38. Mutual funds apart from the crowd (2018)
  39. Mutual fund objective misclassification (2000)
  40. Time series clustering and classification by the autoregressive metric (2008)
  41. Soft Clustering for Funds Management Style Analysis: Out-of-Sample Predictability (2008)
  42. Fund2Vec: Mutual Funds Similarity using Graph Learning (2021) — link
  43. Machine learning fund categorizations (2020)
  44. On Robustness of Mutual Funds Categorization and Distance Metric Learning (2021)
  45. node2vec: Scalable feature learning for networks (2016)
  46. Does Mutual Fund Management in India correspond to its investment objective classification? (2005)
  47. Compensation incentives and risk taking behavior: evidence from mutual funds (1996)
  48. Mutual fund misclassification: Evidence based on style analysis (1997)
  49. Incentive fees and mutual funds (2003)
  50. The Consistency of Self-Declared Hedge Fund Styles—A Return-Based Analysis with Self-Organizing Maps (2005)
  51. Mutual fund styles (1997)
  52. Tweets and peers: defining industry groups and strategic peers based on investor perceptions of stocks on Twitter (2011)
  53. Patent citations and stock performance: Constructing a dynamic industry classification (2014)
  54. Can we design an industry classification system that reflects industry architecture? (2018)
  55. Search-based peer firms: Aggregating investor perceptions through internet co-searches (2015)
  56. Text-based network industries and endogenous product differentiation (2016)
  57. Introducing the North American industry classification system (1998)
  58. Global industry classification standard (GICS \textregistered ) methodology: guiding principles and methodology for GICS (2020)
  59. Industry costs of equity (1997)
  60. Social learning and corporate peer effects (2015)
  61. Differences between COMPUSTAT and CRSP SIC codes and related effects on research (1994)
  62. Spanish Mutual Fund Misclassification: Empirical Evidence (2005)
  63. Stylistic Differences across Hedge Funds as Revealed by Historical Monthly Returns (2010)
  64. hedge Fund classification using K-means clustering Method (2003)
  65. Ultrametricity in fund of funds diversification (2004)
  66. Cluster analysis and manager selection (1986)
  67. Stable Return-Based Fund Classification (2017)
  68. Rifinitive. “Lipper Fund Research.” (2016) — link
  69. Classifying companies by industry using word embeddings (2018)
  70. Clustering mutual funds by return and risk levels (2010)
  71. Variance-based criteria for clustering and their application to the analysis of management styles of mutual funds based on time series of daily returns (2008)
  72. Using Risk Characteristics to Classify Funds (2018)
  73. Objectives and performance of mutual funds, 1960–1969 (1974)
  74. Asset allocation: Management style and performance measurement (1992)
  75. Visualizing stock-mutual fund relationships through social network analysis (2009)
  76. Clustering approaches for financial data analysis: a survey (2016) — link
  77. Do fund objectives affect diversification policies? (1982)
  78. Classifying mutual funds in India: Some results from clustering (2007)
  79. Related stocks selection with data collaboration using text mining (2019)
  80. The style consistency of hedge funds (2007)
  81. A Data-Driven Categorization of Investable Assets (2015)
  82. Clustering mutual funds based on investment similarity (2015)
  83. Self-organizing maps could improve the classification of Spanish mutual funds (2006)
  84. Morningstar’s classification of large-cap mutual funds (2001)
  85. Double clustering for rating mutual funds (2015)
  86. Distributed Representations of Sentences and Documents (2014) — link
  87. Learning Mutual Fund Categorization using Natural Language Processing (2022)
  88. Don’t Take Their Word for It: The Misclassification of Bond Mutual Funds (2021)
  89. Changing names with style: Mutual fund name changes and their effects on fund flows (2005)
  90. Investment style misclassification and mutual fund performance (2017)
  91. Efficient Estimation of Word Representations in Vector Space (2013) — link
  92. Longformer: The long-document transformer (2020) — link
  93. Palm: Scaling language modeling with pathways (2022) — link
  94. Overcoming catastrophic forgetting in neural networks (2017)
  95. Language models are few-shot learners (2020)
  96. Palm: Scaling language modeling with pathways (2022) — link
  97. Mpnet: Masked and permuted pre-training for language understanding (2020)
  98. Language models are unsupervised multitask learners (2019)
  99. Pathways: Asynchronous distributed dataflow for ml (2022)
  100. Roberta: A robustly optimized bert pretraining approach (2019) — link
  101. Instance-based Explanations for Gradient Boosting Machine Predictions with AXIL Weights (2023) — link
  102. How Trading Analytics and Data Science Can Improve Investment Outcomes (2022)
  103. Liquidity—how to capture a multidimensional beast (2016)
  104. Geometry- and Accuracy-Preserving Random Forest Proximities (2023) — link
  105. Quantifying Outlierness of Funds from their Categories using Supervised Similarity — → 2308.06882
  106. Supervised similarity learning for corporate bonds using Random Forest proximities (2022)
  107. Random Forests (2001)
  108. A unified approach to interpreting model predictions (2017)
  109. Discriminatory analysis. Nonparametric discrimination: Consistency properties (1989)
  110. The Handbook of Fixed Income Securities, Eighth Edition (2011) — link
  111. A Survey of Large Language Models for Financial Applications: Progress, Prospects and Challenges — → 2406.11903
  112. Reinforcement learning: An introduction (2018)
  113. A comprehensive overview of large language models (2023) — link
  114. A survey on large language model based autonomous agents (2024)
  115. Large language model based multi-agents: A survey of progress and challenges (2024) — link
  116. A systematic survey of prompt engineering in large language models: Techniques and applications (2024) — link
  117. FinVerse: An Autonomous Agent System for Versatile Financial Analysis (2024) — link

6 of 125 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.