AI Hedge Fund Crypto

A next-generation algorithmic trading framework that leverages graph-based workflow architecture, ensemble technical analysis, and AI language models to make data-driven cryptocurrency trading decisions. This system employs a directed acyclic graph (DAG) of specialized nodes for multi-timeframe analysis, enabling sophisticated signal generation through weighted combinations of diverse trading strategies.

At its core, the system builds upon LangGraph's computational graph architecture to process market data through a pipeline of technical analysis nodes. Each strategy implements a BaseNode interface that processes multi-interval data for multiple assets simultaneously. The framework then aggregates these signals using adaptive weighting mechanisms, evaluates risk parameters, and formulates position management decisions through large language model (LLM) analysis.

The system stands out through its:

• AI-Enhanced Decision Making: Integration of large language models (LLMs) for portfolio management decisions, combining technical signals with sophisticated reasoning
• Compositional Architecture: Distinct nodes for data fetching, strategy execution, risk management, and portfolio management
• Signal Ensemble Approach: Weighted aggregation of multiple technical strategies (trend following, mean reversion, momentum, volatility, and statistical arbitrage)
• Multi-Timeframe Analysis: Simultaneous processing across various time intervals for more robust signal generation
• Dynamic Strategy Visualization: Automatic generation of computational graph visualizations to better understand the decision flow
• Comprehensive Backtesting: Robust historical performance evaluation with detailed metrics and visualizations

Backtest Results

Sample backtesting results showcasing how well-designed trading strategies can generate high-quality signals when implemented in this framework. The system's performance is driven by the quality of your strategies, with the framework enabling efficient implementation, LLM-based decision refinement, and multi-timeframe analysis:

These results demonstrate the system's ability to:

• Generate profitable trading signals across different market conditions
• Manage risk effectively with position sizing rules
• Adapt to changing market dynamics with multi-timeframe analysis
• Provide detailed performance metrics for strategy evaluation

Architecture

The system is built around a highly configurable directed acyclic graph (DAG) architecture, where data flows through various processing nodes. This architecture provides exceptional flexibility, allowing users to customize both the available strategies and the timeframes for analysis without modifying the core code.

Node-Based Workflow System

At its heart, the system uses LangGraph to create a dynamic computational graph where:

1. Start Node: Initializes the workflow and prepares the state for processing
2. Data Nodes: Process market data for specific intervals (e.g., 30m, 1h, 4h)
3. Strategy Nodes: Apply technical analysis algorithms to the processed data
4. Risk Management Node: Evaluates position limits and exposure
5. Portfolio Management Node: Makes final trading decisions using LLM reasoning

The key innovation is that both the data nodes and strategy nodes are fully configurable via the config.yaml file, allowing you to add, remove, or modify these nodes without changing the core architecture.

Configurable Timeframes

You can specify multiple timeframes for analysis in your configuration:

signals:
  intervals: ["5m", "15m", "30m", "1h", "4h", "1d"]

The system will dynamically create separate data processing nodes for each timeframe, allowing strategies to analyze market behavior across multiple time horizons simultaneously. This multi-timeframe approach provides more robust signals by capturing both short-term and long-term market trends.

Configurable Strategies

Similarly, you can specify which trading strategies to include:

signals:
  strategies: ['MacdStrategy', 'RSIStrategy', 'BollingerStrategy']

The system will dynamically load and integrate only the specified strategies into the workflow graph. Each strategy is implemented as an independent node that:

1. Receives the aggregated multi-timeframe data
2. Applies its specific technical analysis algorithms
3. Generates trading signals with confidence levels
4. Passes these signals to the risk management node

This modular approach allows you to easily experiment with different combinations of strategies without rewriting any code. You can also create your own custom strategy modules and add them to the configuration.

Data Flow Architecture

The complete data flow works as follows:

1. The start node initializes the workflow state
2. Multiple timeframe nodes fetch and process data in parallel for different intervals
3. The merge node combines the multi-timeframe data into a unified state
4. Multiple strategy nodes analyze this unified data and generate signals
5. The risk management node applies position sizing and exposure limits
6. The portfolio management node makes final trading decisions using all available signals and LLM reasoning

This architecture provides several advantages:

• Flexibility: Change strategies or timeframes without code modifications
• Parallelization: Process multiple timeframes concurrently for improved performance
• Isolation: Maintain separation of concerns between different components
• Extensibility: Add new strategy nodes without affecting existing functionality
• Visualization: Generate visual representations of the workflow for better understanding

Features

• Strategy-Based Architecture: Implement and backtest multiple trading strategies
• Multi-Timeframe Analysis: Analyze multiple intervals (1h, 4h, 1d, etc.) simultaneously
• Multiple Technical Indicators: MACD, RSI, and custom indicator support
• Comprehensive Backtesting: Test strategies against historical data
• Portfolio Management: Manage positions with support for both long and short trades
• Performance Metrics: Detailed performance statistics and visualization
• Strategy Visualization: Automatically generate strategy workflow graphs

Table of Contents

• Setup
• Configuration
• Usage
  • Running Backtest Mode
  • Running Live Mode
• Creating Custom Strategies
• Project Structure
• Contributing
• Star History
• License
• Disclaimer
• Referral Links

Setup

Prerequisites

• Python 3.9 or higher (Python 3.12 recommended — used during development)
• Binance account (required for accessing market data)

⚠️ While Python 3.9+ should work, we recommend using Python 3.12 for full compatibility with the development environment.

Installation

1. Clone the repository:

git clone https://github.com/51bitquant/ai-hedge-fund-crypto.git
cd ai-hedge-fund-crypto

2. Set up using uv (recommended):

# Install uv if you don't have it
curl -fsSL https://install.lunarvim.org/uv.sh | sh

# create an venv with python 3.12
uv venv --python 3.12

# Activate the virtual environment
# For macOS/Linux:
source .venv/bin/activate
# For Windows (PowerShell):
.venv\Scripts\Activate.ps1

# For Windows (CMD):
.venv\Scripts\activate.bat
# Install dependencies using pyproject.toml and lockfile
uv pip sync

If you have more questions, feel free to check out the uv documentation

3. Copy the example configuration:

cp config.example.yaml config.yaml

4. Set up environment variables:

cp .env.example .env

5. Add your API keys to the .env file:

# Binance API keys (required for data access)
BINANCE_API_KEY=your-binance-api-key
BINANCE_API_SECRET=your-binance-api-secret
# For running LLMs hosted by openai (gpt-4o, gpt-4o-mini, etc.)
# Get your OpenAI API key from https://platform.openai.com/
OPENAI_API_KEY=your-open-ai-key
GROQ_API_KEY=your-open-ai-key
OPENROUTER_API_KEY=your-open-ai-key
ANTHROPIC_API_KEY=your-anthropic-api-key
GOOGLE_API_KEY=your-google-api-key

Configuration

The system is configured through the config.yaml file. Here's what each setting means:

mode: backtest  # Options: backtest, live
start_date: 2025-04-20  # Start date for backtesting
end_date: 2025-05-01  # End date for backtesting
primary_interval: 1h  # Main timeframe for analysis
initial_cash: 100000  # Starting capital
margin_requirement: 0.0  # Margin requirements for short positions
show_reasoning: false  # Whether to show strategy reasoning
show_agent_graph: true  # Whether to show the agent workflow graph
signals:
  intervals: ["30m", "1h", "4h"]  # Timeframes to analyze
  tickers: ["BTCUSDT", "ETHUSDT"]  # Trading pairs
  strategies: ['MacdStrategy']  # Strategies to use
model:
  name: "gpt-4o-mini" # config your model
  provider: "openai"  # config your llm provider, support openai， groq， openrouter，gemini，anthropic，ollama
#  base_url: "https://api.openai.com/v1"  # not required, but if you want to set the base_url

Supported providers include OpenAI, Groq, OpenRouter, Gemini, Anthropic, and Ollama. Most LLMs are compatible with the OpenAI SDK, so you can often switch models by simply adjusting the model name and base_url to match the desired provider.

Usage