Paper: arXiv 2501.00826

Authors: Yichen Luo, Yebo Feng, Jiahua Xu, Paolo Tasca, Yang Liu

Abstract

Cryptocurrency investment is inherently difficult due to its shorter history compared to traditional assets, the need to integrate vast amounts of data from various modalities, and the requirement for complex reasoning. While deep learning approaches have been applied to address these challenges, their black-box nature raises concerns about trust and explainability. Recently, large language models (LLMs) have shown promise in financial applications due to their ability to understand multi-modal data and generate explainable decisions. However, single LLM faces limitations in complex, comprehensive tasks such as asset investment. These limitations are even more pronounced in cryptocurrency investment, where LLMs have less domain-specific knowledge in their training corpora. To overcome these challenges, we propose an explainable, multi-modal, multi-agent framework for cryptocurrency investment. Our framework uses specialized agents that collaborate within and across teams to handle subtasks such as data analysis, literature integration, and investment decision-making for the top 30 cryptocurrencies by market capitalization. The expert training module fine-tunes agents using multi-modal historical data and professional investment literature, while the multi-agent investment module employs real-time data to make informed cryptocurrency investment decisions. Unique intrateam and interteam collaboration mechanisms enhance prediction accuracy by adjusting final predictions based on confidence levels within agent teams and facilitating information sharing between teams. Empirical evaluation using data from November 2023 to September 2024 demonstrates that our framework outperforms single-agent models and market benchmarks in classification, asset pricing, portfolio, and explainability performance.

Complexity vs Empirical Score

  • Math Complexity: 3.0/10
  • Empirical Rigor: 6.0/10 (capped from 7.0: transaction costs not modelled; see penalty flags)
  • Quadrant: Street Traders — empirical and applied, lower math complexity (effective empirical rigor capped from 7.0)

Why this score: The paper presents an empirical framework with real-time data, backtests against benchmarks, and specific performance metrics, indicating high implementation and data rigor. However, its mathematical complexity is low, relying on LLM architectures and confidence aggregation rather than dense mathematical derivations or advanced statistical theory.

Research Flowchart

  flowchart TD
  A["Research Goal:<br>Explainable Automated<br>Crypto Portfolio Management"] --> B["Methodology:<br>Multi-Agent LLM Framework"]

  B --> C{"Key Components"}

  C --> D["Specialized Agents"]
  C --> E["Expert Training Module<br>Fine-tuning with multi-modal data"]
  C --> F["Multi-Agent Investment Module<br>Real-time decision making"]
  C --> G["Collaboration Mechanisms<br>Intrateam & Interteam coordination"]

  D & E & F & G --> H["Computational Process:<br>Agent collaboration &<br>Confidence-weighted prediction"]

  H --> I["Key Outcomes<br>2023-2024 Evaluation"]
  
  I --> J{"Performance Results"}
  
  J --> K["Outperforms benchmarks<br>in classification & pricing"]
  J --> L["Superior portfolio performance"]
  J --> M["Enhanced explainability"]

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 an explainable, multi‑modal, multi‑agent framework for cryptocurrency investment, comprising expert training and multi‑agent investment modules.

Source sentence
we propose an explainable, multi-modal, multi-agent framework for cryptocurrency investment. Our framework uses specialized agents that collaborate within and across teams to handle subtasks such as data analysis, literature integration, and investment decision-making for the top 30 cryptocurrencies by market capitaliz

Assumes 5

  • Only the top 30 high‑capitalization cryptocurrencies are considered, excluding low‑liquidity assets. market
    Source sentence
    The rationale for including only high‑capitalization cryptocurrencies is that those with low market capitalization often exhibit pricing dynamics that differ significantly from high‑liquidity cryptocurrencies, partly due to risks such as pump-and-dump schemes.
  • Data from November 2023 to September 2024 is used as a test set to avoid information leakage from ChatGPT‑4o’s training data. data
    Source sentence
    to September 2024 as the test set, given that ChatGPT‑4o’s training data extends only up to October 2023.
  • Portfolios are equal‑weighted when evaluating performance. execution
    Source sentence
    All portfolios are equal‑weighted.
  • Risk‑factor quintiles are determined using cross‑sectional data and the initial two years of market‑related factors. statistical
    Source sentence
    Risk factors are categorized into five quintiles: “Very Low”, “Low”, “Medium”, “High”, and “Very High”. The quintile cutoffs are determined using cross‑sectional data for crypto‑related factors and the initial two years of data for market‑related factors.
Show 1 more
  • Real‑time data from Coingecko, Blockchain.info, Coin Metrics, and news from Cointelegraph are sufficient for investment decisions. data
    Source sentence
    are computed using OHLC price data, trading volume, and market capitalization from Coingecko, along with on‑chain data sourced from Blockchain.info and Coin Metrics.

Data

Backtest sample: 2023-11 – 2024-08 · Top 30 cryptocurrencies by market cap · crypto · weekly

  • Cryptocurrency price, volume and market‑cap data · Coingecko market-prices public
    June 2023 to September 2024 · Top 30 cryptocurrencies by market capitalization
    Source sentence
    The 30‑day candlestick charts (Fig. 1a) and binary price trends (Fig. 1b) are derived from open high low and close (OHLC) price data and trading volume provided by Coingecko.
  • On‑chain data · Blockchain.info and Coin Metrics crypto-onchain public
    June 2023 to September 2024 · Top 30 cryptocurrencies by market capitalization
    Source sentence
    are computed using OHLC price data, trading volume, and market capitalization from Coingecko, along with on‑chain data sourced from Blockchain.info and Coin Metrics.
  • Cryptocurrency news headlines · Cointelegraph text-news public
    June 2023 to September 2024 · Top 30 cryptocurrencies by market capitalization
    Source sentence
    news headline data (Fig. 1d) is obtained through web crawling from Cointelegraph.

Recognised sources: CoinGecko

Costs

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

Finds 2

Reported performance (Multi-agent cryptocurrency investment framework): Sharpe 1.54 (benchmark 1.38)

  • The multi‑agent framework achieves the highest classification accuracy for cryptocurrency price‑trend prediction, outperforming a single fine‑tuned GPT‑4o model. predictive-accuracy
    Source sentence
    The multi‑agent framework outperforms a single GPT‑4o model with fine‑tuning, demonstrating that multiple expert agents, each trained with domain‑specific knowledge, perform better than a single agent trained with a comprehensive dataset for cryptocurrency price trend prediction.
  • Overall, the proposed framework outperforms single‑agent models and market benchmarks in both classification accuracy and asset‑pricing performance. performance
    Source sentence
    Our experimental results demonstrate that our model outperforms single‑agent models, both with and without fine‑tuning, in terms of classification accuracy and asset pricing performance. Furthermore, we show that our framework surpasses market benchmarks in portfolio performance.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 20 related papers). Coverage is partial, so this is neither replication nor refutation.

Cited by 3 papers in the archive (2 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 (111)

  1. A Multi-agent System Based On LLM For Trading Financial Assets (2025) — link
  2. A Reflective LLM-based Agent to Guide Zero-shot Cryptocurrency Trading — → 2407.09546
  3. A Survey of Large Language Models (2023) — link
  4. A Survey on Multimodal Large Language Models (2024)
  5. A Time Series is Worth 64 Words: Long-term Forecasting with Transformers (2023) — link
  6. Agent AI: Surveying the Horizons of Multimodal Interaction (2024) — link
  7. AgentCoord: Visually Exploring Coordination Strategy for LLM-based Multi-Agent Collaboration (2024) — link
  8. Analysis of Ethereum Smart Contracts and Opcodes (2020) — link
and 103 more
  1. Are institutions informed about news? (2015) — link
  2. Are More LLM Calls All You Need? Towards Scaling Laws of Compound Inference Systems (2024) — link
  3. Autoformer: Decomposition Transformers with Auto-Correlation for Long-Term Series Forecasting (2021)
  4. AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation (2023) — link
  5. Automate Strategy Finding with LLM in Quant investment — → 2409.06289
  6. Automated cryptocurrency trading approach using ensemble deep reinforcement learning: Learn to understand candlesticks (2024) — link
  7. Bitcoin’s limited adoption problem (2022) — link
  8. Blockchain Disruption and Smart Contracts (2019) — link
  9. Blockchain without Waste: Proof-of-Stake (2021) — link
  10. Blockchain-Based Settlement for Asset Trading (2019) — link
  11. CFP Board Issues Crypto Guidelines (2022) — link
  12. Chain of Hindsight Aligns Language Models with Feedback (2023)
  13. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models (2022)
  14. Committee-Based Blockchains as Games between Opportunistic Players and Adversaries (2024) — link
  15. Common Risk Factors in Cryptocurrency (2022) — link
  16. Common risk factors in the returns on stocks and bonds (1993) — link
  17. Communicative Agents for Software Development (2023) — link
  18. Consumers as Financiers: Consumer Surplus, Crowdfunding, and Initial Coin Offerings (2022) — link
  19. Corporate Capture of Blockchain Governance (2023) — link
  20. Crypto Risk Premia (2022) — link
  21. Cryptocurrencies as a financial asset: A systematic analysis (2019) — link
  22. Cryptocurrency forecasting with deep learning chaotic neural networks (2019) — link
  23. Cryptocurrency price fluctuation and time series analysis through candlestick pattern of bitcoin and ethereum using machine learning (2024) — link
  24. Cryptocurrency trading: a comprehensive survey (2022) — link
  25. Decentralization through Tokenization (2023) — link
  26. Decentralized Mining in Centralized Pools (2021) — link
  27. Decentralizing Money: Bitcoin Prices and Blockchain Security (2022) — link
  28. Deep learning and technical analysis in cryptocurrency market (2023) — link
  29. DeFi lending: intermediation without information? (2022) — [link](https://ideas.repec.org/p/bis/bisblt/57.html https://ideas.repec.org//p/bis/bisblt/57.html)
  30. Do news headlines matter in the cryptocurrency market? (2022) — link
  31. Empirical Asset Pricing via Machine Learning (2020) — link
  32. Encouraging Divergent Thinking in Large Language Models through Multi-Agent Debate (2023) — link
  33. Enhancing LLM Trading Performance with Fact-Subjectivity Aware Reasoning (2024) — link
  34. Ethereum smart contracts: Analysis and statistics of their source code and opcodes (2020) — link
  35. EtherSolve: Computing an Accurate Control-Flow Graph from Ethereum Bytecode (2021) — link
  36. Evaluating the Text-to-SQL Capabilities of Large Language Models (2022) — link
  37. Exchange-of-Thought: Enhancing Large Language Model Capabilities through Cross-Model Communication (2023) — link
  38. Explainable Machine Learning Exploiting News and Domain-Specific Lexicon for Stock Market Forecasting (2021) — link
  39. Financial technology: Efficiency evaluation of challenger banks (2019) — link
  40. FinCon: A Synthesized LLM Multi-Agent System with Conceptual Verbal Reinforcement for Enhanced Financial Decision Making (2024) — link
  41. Fine-tuning (2024) — link
  42. FinGPT: Democratizing Internet-scale Data for Financial Large Language Models (2023) — link
  43. FinNLP-FNP-LLMFinLegal @ COLING 2025 Shared Task: Agent-Based Single Cryptocurrency Trading Challenge (2025) — link
  44. FinVision: A Multi-Agent Framework for Stock Market Prediction — → 2411.08899
  45. Forecasting cryptocurrency returns with machine learning (2023) — link
  46. From mining to markets: The evolution of bitcoin transaction fees (2019) — link
  47. Generative Agents: Interactive Simulacra of Human Behavior (2023) — link
  48. Harnessing the Power of LLMs in Practice: A Survey on ChatGPT and Beyond (2024)
  49. How Valuable Is FinTech Innovation? (2019) — link
  50. Human-centric justification of machine learning predictions (2017) — link
  51. Improving Factuality and Reasoning in Language Models through Multiagent Debate (2023) — link
  52. In Search of Attention (2011) — link
  53. Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting (2021) — link
  54. Initial Coin Offerings: Financing Growth with Cryptocurrency Token Sales (2020) — link
  55. Is Bitcoin Really Untethered? (2020) — link
  56. Kenneth R. French - Description of Fama/French Factors — link
  57. Large Language Models in Finance: A Survey — → 2311.10723
  58. Learning to Generate Explainable Stock Predictions using Self-Reflective Large Language Models — → 2402.03659
  59. Long Short-Term Memory (1997) — link
  60. Machine learning and the cross-section of cryptocurrency returns (2024) — link
  61. MARG: Multi-Agent Review Generation for Scientific Papers — link
  62. Measuring Economic Policy Uncertainty (2016) — link
  63. MedAgents: Large Language Models as Collaborators for Zero-shot Medical Reasoning (2023) — link
  64. MetaGPT: Meta Programming for A Multi-Agent Collaborative Framework (2023) — link
  65. Models (2024) — link
  66. Multimodal Gen-AI for Fundamental Investment Research — → 2401.06164
  67. Nothing but noise? Price discovery across cryptocurrency exchanges (2021) — link
  68. Optimal financing with tokens (2021) — link
  69. Optimal Versus Naive Diversification: How Inefficient is the 1/N Portfolio Strategy? (2009) — link
  70. PEN: Prediction-Explanation Network to Forecast Stock Price Movement with Better Explainability (2023) — link
  71. PIXIU: A Large Language Model, Instruction Data and Evaluation Benchmark for Finance (2023) — link
  72. Portfolio Management with Cryptocurrencies: The Role of Estimation Risk (2019) — link
  73. Portfolio optimization in the era of digital financialization using cryptocurrencies (2020) — link
  74. Product market competition with crypto tokens and smart contracts (2023) — link
  75. Ransomware activity and blockchain congestion (2021) — link
  76. ReAct: Synergizing Reasoning and Acting in Language Models (2023) — link
  77. Reflexion: language agents with verbal reinforcement learning (2023) — link
  78. Reinforcement Learning with Maskable Stock Representation for Portfolio Management in Customizable Stock Pools — → 2311.10801
  79. Resisting Manipulative Bots in Meme Coin Copy Trading: A Multi-Agent Approach with Chain-of-Thought Reasoning — → 2601.08641
  80. Rethinking the Bounds of LLM Reasoning: Are Multi-Agent Discussions the Key? (2024) — link
  81. Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (2020) — link
  82. Risk-Based Portfolio Optimization in the Cryptocurrency World * (2019) — link
  83. Risks and Returns of Cryptocurrency (2021) — link
  84. Scaling Large-Language-Model-based Multi-Agent Collaboration (2024) — link
  85. Sex, Drugs, and Bitcoin: How Much Illegal Activity Is Financed through Cryptocurrencies? (2019) — link
  86. Stock market 101 : from bull and bear markets to dividends, shares, and margins - your essential guide to the stock market (2024)
  87. Technical trading and cryptocurrencies (2021) — link
  88. The advantages of the Matthews correlation coefficient (MCC) over F1 score and accuracy in binary classification evaluation (2020) — link
  89. The Blockchain Folk Theorem (2019) — link
  90. The Characteristics and Portfolio Behavior of Bitcoin Investors: Evidence from Indirect Cryptocurrency Investments (2022) — link
  91. The matthews correlation coefficient (Mcc) is more reliable than balanced accuracy, bookmaker informedness, and markedness in two-class confusion matrix evaluation (2021) — link
  92. The Sharpe Ratio (1994) — link
  93. The Treynor Capital Asset Pricing Model — link
  94. TimesNet: Temporal 2D-Variation Modeling for General Time Series Analysis (2023) — link
  95. To FinTech and Beyond (2019) — link
  96. Token-based platform finance (2022) — link
  97. Tokenomics: Dynamic Adoption and Valuation (2021) — link
  98. TradExpert: Revolutionizing Trading with Mixture of Expert LLMs — → 2411.00782
  99. Trading and arbitrage in cryptocurrency markets (2020) — link
  100. Trend-based forecast of cryptocurrency returns (2023) — link
  101. Value and Momentum Everywhere (2013) — link
  102. Vision-Language Models for Vision Tasks: A Survey (2024)
  103. When AI Meets Finance (StockAgent): Large Language Model-based Stock Trading in Simulated Real-world Environments Additional Key Words and Phrases: Multi-Agent AI System, Simulated Investor Behavior, Systematic Biases, Stock trading simulation ACM Reference Format (2024) — link

9 of 111 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.