FinAgents

A multi-agent system that simulates a debate between famous investors analyzing stock investment opportunities.

Overview

FinAgents creates a virtual investment committee where AI agents with personas of famous investors like Warren Buffett and Cathie Wood debate whether stocks are good investments. The system:

1. Gathers real financial data using yfinance
2. Has specialized analyst agents evaluate different aspects of the stock
3. Has investor agents with unique investment philosophies debate the merits
4. Synthesizes a comprehensive investment recommendation

Current Status

Prototype: This system currently uses pre-generated static responses instead of live LLM calls. The investor and analyst agents return canned text from local_claude_responses.py rather than calling the Anthropic API. Financial data retrieval via yfinance is real, but the AI debate is simulated. See Making It Operational below for instructions on wiring up real LLM calls.

Key Features

• Parallel Processing: Agents run concurrently for faster analysis
• Diverse Perspectives: 5 famous investor personas with unique philosophies
• Specialized Analysis: 5 types of analyst agents focus on different aspects
• Real Market Data: Uses yfinance to pull actual stock information
• Comprehensive Output: Generates detailed reports saved as Markdown files

Investor Personas

• Warren Buffett: Value investing, long-term perspective
• Ray Dalio: Macro investing, economic cycles
• Cathie Wood: Growth and disruptive innovation
• Peter Lynch: Growth at a reasonable price
• Michael Burry: Contrarian investing, spotting market inefficiencies

Setup

1. Create a virtual environment:

python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

2. Install dependencies:

pip install -r requirements.txt

3. Set up environment variables: Create a .env file with:

ANTHROPIC_API_KEY=your_api_key_here
DEFAULT_MODEL=claude-3-7-sonnet-20250219
STOCK_DATA_SOURCE=yfinance

Usage

Run the main script:

python main.py

The system will analyze Microsoft (MSFT) by default, using all agents in parallel, and save detailed reports to the results directory.

Output Files

For each analyzed stock (e.g., MSFT):

• results/MSFT_decision.md: Final synthesized investment recommendation
• results/MSFT_detailed_analysis.md: Detailed reports from all analysts and investors

Project Structure

finagents/
├── main.py                      # Main script to run the system
├── src/
│   ├── agents/
│   │   ├── investor_agents.py   # Famous investor personas
│   │   ├── analyst_agents.py    # Financial analyst agents
│   │   ├── debate_manager.py    # Orchestrates the analysis process
│   ├── utils/
│   │   ├── stock_data.py        # Functions for retrieving stock data
├── tests/                       # Comprehensive test suite
│   ├── unit/                    # Unit tests for individual components
│   ├── integration/             # Integration tests for the whole system
├── results/                     # Output directory for analysis reports
├── requirements.txt             # Project dependencies
├── .env                         # Environment variables (API keys)
├── run_tests.py                 # Script to run all tests
├── README.md                    # This file
├── ARCHITECTURE.md              # Detailed system architecture
├── QUICKSTART.md                # Quick start guide
├── PERSONAS.md                  # Detailed guide to agent personas
├── EXTENDING.md                 # Guide for extending the system

Requirements

• Python 3.8+
• Anthropic API key (for Claude 3.7)
• Internet connection (for financial data retrieval)

Running Tests

The project includes a test suite focusing on core data utilities and file operations:

python run_tests.py

This will run the stock data utility tests and output file tests, providing a detailed report.

Making It Operational

To convert this prototype into a fully operational system with real LLM calls:

1. Wire up the Anthropic API

Replace the LocalInvestor and LocalAnalyst classes with real LangChain LLM calls. In src/agents/investor_agents.py and src/agents/analyst_agents.py, the create_investor_agent() and create_analyst_agent() functions currently return LocalInvestor/LocalAnalyst objects that serve canned responses. Replace them with LangChain's ChatAnthropic:

from langchain_anthropic import ChatAnthropic

def create_investor_agent(persona_name: str):
    llm = ChatAnthropic(model=os.getenv("DEFAULT_MODEL", "claude-sonnet-4-20250514"))
    prompt = create_investor_prompt(persona_name)  # Already defined in the file
    return prompt | llm

2. Remove mock infrastructure

• Delete local_claude_responses.py (1,556 lines of static responses)
• Remove from unittest.mock import MagicMock from debate_manager.py
• Replace self.llm = MagicMock() and self.synthesis_chain = MagicMock() with real LangChain chain instances

3. Add live market context

Replace debate_manager.py's get_market_context() static placeholder with real data. Options:

• Use yfinance to fetch S&P 500, VIX, and Treasury yields (already a dependency)
• Use a financial news API for current market sentiment

4. Make the portfolio path configurable

In main.py, the portfolio path is hardcoded. Replace with:

portfolio_path = os.getenv(
    "PORTFOLIO_PATH",
    os.path.expanduser("~/SourceCode/etorotrade/yahoofinance/output/portfolio.csv")
)

5. Install additional dependency

pip install langchain-anthropic

6. Sync setup.py with requirements.txt

Update setup.py version constraints to match requirements.txt (langchain >=0.2.0, langchain-community >=0.2.0).

License

MIT

DISCLAIMER:

This project is purely educational. The AI-generated personas are fictional and do not represent real statements or endorsements from Warren Buffett, Ray Dalio, Cathie Wood, Peter Lynch, or Michael Burry.