Takeaway:
A multi-agent system of fine-tuned LLMs coordinating across crypto-specific data, news, technicals, and risk factors can beat the market, generating higher risk-adjusted returns than traditional benchmarks and single-agent models.
Key Idea: What Is This Paper About?
This paper presents an LLM-powered multi-agent framework for managing crypto portfolios. Unlike single models, the system uses specialized agents (e.g., market expert, technical analyst, news interpreter) that collaborate to analyze multi-modal data—charts, news, on-chain metrics—and make optimized investment decisions. The result is superior performance and more explainable decisions across top 30 cryptocurrencies.
Economic Rationale: Why Should This Work?
Crypto markets are noisy, volatile, and span many data types. A multi-agent LLM system can break the problem into parts and integrate context better than a single monolithic model.
Relevant Economic Theories and Justifications:
- Behavioral Finance: Modular agents reduce overfitting and bias via specialization.
- Efficient Market Hypothesis (semi-strong form): Public info is not fully priced in, especially in crypto.
- Ensemble Learning: Aggregating expert predictions improves robustness and generalization.
Why It Matters:
This system provides a blueprint for interpretable, LLM-driven asset management—critical in high-volatility, low-transparency markets like crypto.
How to Do It: Data, Model, and Strategy Implementation
Data Used
- Data Sources: Coingecko, Cointelegraph, Blockchain.info, Coin Metrics, academic literature via Google Scholar
- Time Period: Training: Jun–Oct 2023; Testing: Nov 2023–Sep 2024
- Asset Universe: Top 30 cryptocurrencies by market cap
Model / Methodology
- Type of Model: Multi-Agent System using GPT-4o-based agents
- Key Features: News sentiment, on-chain & off-chain factors, candlestick charts, academic knowledge
- Training Approach: Prompt-based fine-tuning with expert agents for each modality (news, technicals, fundamentals)
- Collaboration Mechanism:
- Intrateam: Ensemble predictions using confidence-weighted averaging
- Interteam: Shared memory between market and crypto-specific agents for cross-context reasoning
Trading Strategy
- Signal Generation: Weekly rise/fall classification + rise probability from each agent
- Portfolio Construction:
- Full crypto or 50/50 cash allocation based on market prediction
- Top 20% of cryptos selected using consensus signals (Very High rise probability)
- Rebalancing Frequency: Weekly (event-driven via prediction updates)
Key Table or Figure from the Paper
Explanation:
Compares Sharpe, mean return, and volatility for the LLM multi-agent model vs market index, Bitcoin, and equal-weighted crypto portfolio. The LLM system performs best across all periods (full, boom, bust), especially during down markets.
Final Thought
🤖 AI teamwork beats solo agents—and even beats Bitcoin—in crypto investing.
Paper Details (For Further Reading)
- Title: LLM-Powered Multi-Agent System for Automated Crypto Portfolio Management
- Authors: Yichen Luo, Yebo Feng, Jiahua Xu, Paolo Tasca, Yang Liu
- Publication Year: 2025
- Journal/Source: arXiv Preprint
- Link: https://arxiv.org/abs/2501.00826
🛠️ How to Build the LLM-Powered Multi-Agent Crypto Portfolio System
This guide shows how to recreate the system from the paper LLM-Powered Multi-Agent System for Automated Crypto Portfolio Management. It includes agents, data, prompts, collaboration logic, and API usage.
📦 1. System Components Overview
- Expert Training Module: Fine-tunes LLM agents using historical data + academic papers.
- Multi-Agent Investment Module: Uses real-time data to allocate assets and select cryptos.
- Agents:
- Data Fetcher
- News Expert
- Market Factor Expert
- Crypto Factor Expert
- Technical Expert
- Trader
- LLM Used: GPT-4o (via OpenAI API)
- Crypto Exchange API: Used by Trader agent to execute trades (e.g., Binance API).
📊 2. Data Collection & Preprocessing
-
Sources:
- Coingecko API: OHLC prices, volume, market cap
- Blockchain.info, CoinMetrics: on-chain metrics
- Cointelegraph: weekly news headlines (web scraping)
- Google Scholar: academic papers
-
Modalities Collected:
- Candlestick charts (30 days)
- Binary future trend labels: "Rise"/"Fall"
- Risk factors: momentum, volatility, size
- News headlines
- Literature relevant to crypto pricing
🧾 3. Prompt Templates for Training
📌 System Instruction for All Agents
"You are a professional cryptocurrency analyst, specializing in explaining the predicted target based on the provided knowledge and information."
📌 Prompt 1: Market Factor Explanation Agent
"Learn the following cryptocurrency investment knowledge. Using this knowledge, explain the predicted target for the upcoming week based on the provided information.
Market factors have been categorized into Very High, High, Medium, Low, and Very Low.
The predicted market return is categorized into Rise or Fall.
Knowledge: {literature} (End of knowledge)
Information: {factors/news} (End of information)
Market trend: {future market trend} (End of market trend)"
📌 Prompt 2: News Explanation Agent
"Learn the following cryptocurrency investment knowledge. Using this knowledge, explain the predicted target for the upcoming week on the provided news headlines.
Knowledge: {literature} (End of knowledge)
Information: {news} (End of information)
Market trend: {future market trend} (End of market trend)"
📌 Prompt 3: Crypto Factor Explanation Agent
"Learn the following cryptocurrency investment knowledge. Using this knowledge, explain the predicted price trend of {target crypto} for the upcoming week based on the provided information.
Crypto data are categorized into Very High, High, Medium, Low, and Very Low.
Price trend: {future price trend} (End of price trend)
Knowledge: {literature} (End of knowledge)
Information: {factors} (End of information)"
📌 Prompt 4: Vision (Candlestick Chart) Agent
"Learn the following cryptocurrency investment knowledge. Using this knowledge, explain the predicted price trend of {target crypto} for the upcoming week based on the provided candlestick chart.
The chart includes OHLC prices, a 30-day MA line, and volume bars.
Price trend: {future price trend} (End of price trend)
Knowledge: {literature} (End of knowledge)
Image URL: {URL}"
📌 Prompt 5: Final Fine-Tuning Prompt
"Analyze the following information of crypto to determine its target in a week. Please respond with Rise or Fall and provide your reasoning for the prediction.
{info} (End of information)"
Response Format:
"Price trend: {trend}
Explanation: {explanation}"
🧠 4. Agent Fine-Tuning
Each expert LLM is fine-tuned separately using the prompts above.
- Fine-tune one model per agent (News, Market Factor, Crypto Factor, Technical)
- Use OpenAI’s fine-tuning API: https://platform.openai.com/docs/guides/fine-tuning
- Input format: JSONL files with prompt-completion pairs
🔁 5. Multi-Agent Investment Flow (Real-Time)
5.1 Data Fetcher
- Pulls real-time data from APIs
- Sends formatted data to market and crypto teams
5.2 Market Team (Market Factor + News)
-
Each agent predicts:
"Market trend: Rise" or "Market trend: Fall"
"Explanation: {reason}" -
Extract log-prob of token "Rise" from both agents
-
Convert to linear probs, average:
final_rise_prob = average([prob_market_factor, prob_news])
-
If
final_rise_prob > 0.5→ 100% crypto -
Else → 50% crypto / 50% cash
5.3 Crypto Team (Crypto Factor + Technical)
-
Each crypto is evaluated by both experts
-
Use market team outputs as shared memory
-
Predict rise probability per crypto:
final_rise_prob_crypto = average([prob_tech, prob_factor])
-
Rank all cryptos by score
-
Top 20% = final selection for investment
💸 6. Trader Agent
- Uses Crypto Exchange APIs (e.g., Binance, Coinbase)
- Allocates assets:
- Market = Rise → fully invest in selected cryptos
- Market = Fall → 50% crypto, 50% stablecoin
- Executes trades accordingly
📊 7. Evaluation Metrics
- Classification: Accuracy, MCC
- Portfolio: Sharpe Ratio, Cumulative Return, Weekly Mean & Std
- Asset Pricing: Quintile & HML returns
- Explainability (via GPT-4o scoring):
- Professionalism
- Clarity & Coherence
- Objectivity
- Rationale
- Consistency
- Use intrateam collaboration to average agent predictions by confidence
- Use interteam collaboration to pass memory/context between teams
- Fine-tuning + agent specialization → higher performance than single GPT models
- System supports full end-to-end investment decision and execution