LLM Investment Arena: Innovative Practice of Multi-Agent Simulated Trading System

The llm_invest project builds a multi-agent investment arena, allowing top large language models such as GPT, Gemini, and Claude to compete and collaborate in a simulated trading environment, exploring a new paradigm of AI-driven investment research, and analyzing the application value of multi-agent architecture in financial decision-making.

Quantitative investment has gone through stages of technical analysis, statistical arbitrage, and machine learning, but traditional methods are essentially deterministic algorithms. LLMs have extensive knowledge reserves, reasoning abilities, and text comprehension skills, making them capable of being real "investment analysts". The innovation of llm_invest lies in adopting a multi-agent system, where more complex and robust behavioral patterns emerge through agent interactions.

Agent Design

Each agent is based on a specific LLM, with investment philosophy, toolset, memory mechanism, and risk management rules, forming a unique decision-making style.

Tool Ecosystem

Agents can call tools such as market data, technical analysis, fundamentals, and news sentiment to obtain information.

Bulletin Board Mechanism

A shared communication space where agents publish analyses, research notes, portfolio status, and reflections, simulating team information sharing.

Simulated Trading Execution

Supports virtual account trading, considering market constraints (slippage, costs, etc.) to verify the effectiveness of strategies.

Take the Claude agent as an example:

1. Market Environment Analysis: Evaluate panic index, VIX, and sector performance;
2. Portfolio Status Check: Review position returns, predictions, ratings, and technical indicators;
3. Rebalancing Decision: Combine HRP model recommendations and defensive logic, choose not to trade;
4. Decision Reflection: Emphasize the value of "inaction" and reflect metacognitive ability.

Multi-Provider Support

Configure OpenAI, Google, and Anthropic APIs via environment variables to flexibly select agents.

Data Infrastructure

Use GCP's BigQuery to store historical data, and Firestore to store real-time agent status and bulletin board information.

Predictive Model Integration

Integrate specialized predictive models to provide Ensemble Forecast for decision support.

Management Interface

Build a web interface with Streamlit, supporting functions such as agent configuration, portfolio monitoring, and transaction analysis.

Application Value

• Research Value: Study LLM capability differences, multi-agent collaboration effects, etc., in a controlled environment;
• Strategy Verification: Quickly verify strategy ideas and conduct preliminary feasibility assessments;
• Educational Significance: Decision logs help understand investment logic and market analysis methods.

Limitations

• LLM hallucinations may lead to incorrect analysis;
• Simulated trading cannot fully replicate the real market;
• High API costs and lack of extreme market testing.

Future expectations include:

• More professional agent role division (economists, industry analysts, risk control officers, etc.);
• Complex collaboration mechanisms (debate, voting, investment committee-style decision-making);
• Integration with real trading systems to apply and verify strategies under risk control frameworks.

llm_invest demonstrates the emergent behavior of multiple AI agents under a common goal. Although it is still far from replacing human experts, it opens up a new path for AI-assisted investment decision-making, reflecting humanity's continuous exploration of the essence of intelligence.