Trading agents llm

trading-agents-llm

USE FOR:

• "build multi-agent trading system"
• "LLM-powered stock analysis pipeline"
• "analyst + researcher + trader + risk manager agent workflow"
• "AI agent debate for trading decisions"
• "integrate Claude / GPT / Gemini into trading research"
• "A-share / HK / US equity LLM analysis"
• "automate fundamental + sentiment + news + technical analysis" tags: [multi-agent, LLM, trading, AI, equities, fundamentals, sentiment, technical, risk, LangGraph, Claude, GPT, research] kind: framework category: quant-ml-trading

---

What Is TradingAgents?

Open-source multi-agent LLM framework that simulates a trading firm:

• Specialized agents collaborate across the full research → decision pipeline
• Uses LangGraph for agent orchestration
• Supports 6+ LLM providers including Anthropic Claude
• Research only — not financial advice

Repos:

• Original: https://github.com/TauricResearch/TradingAgents
• CN Enhanced Fork: https://github.com/hsliuping/TradingAgents-CN

---

Agent Architecture

┌─────────────────── ANALYST TEAM ───────────────────┐
│  Fundamentals Analyst  →  Financial metrics & value │
│  Sentiment Analyst     →  Social media & mood       │
│  News Analyst          →  Macro news & events       │
│  Technical Analyst     →  MACD, RSI, patterns       │
└─────────────────────────────────────────────────────┘
            ↓ Reports fed into ↓
┌─────────────── RESEARCHER TEAM ────────────────────┐
│  Bullish Researcher  ↔  Bearish Researcher (debate) │
│  Critical assessment of analyst findings            │
└─────────────────────────────────────────────────────┘
            ↓ Debate synthesis ↓
┌─────────────── TRADER AGENT ───────────────────────┐
│  Synthesizes all reports → trading decision         │
│  Determines timing and position magnitude           │
└─────────────────────────────────────────────────────┘
            ↓ Proposal submitted ↓
┌─────────── RISK MANAGEMENT TEAM ───────────────────┐
│  Portfolio Manager   → approves / rejects trades    │
│  Risk evaluator      → volatility + liquidity check │
└─────────────────────────────────────────────────────┘

---

Installation (Original)

git clone https://github.com/TauricResearch/TradingAgents.git
cd TradingAgents
conda create -n tradingagents python=3.13
conda activate tradingagents
pip install -r requirements.txt

Required API Keys:

export OPENAI_API_KEY="sk-..."         # or any supported provider
export ANTHROPIC_API_KEY="sk-ant-..." # for Claude
export ALPHA_VANTAGE_API_KEY="..."     # market data

---

Usage

CLI (Interactive)

python -m cli.main
# Select: ticker, date, LLM provider, research depth

Python API

from tradingagents.graph.trading_graph import TradingAgentsGraph
from tradingagents.default_config import DEFAULT_CONFIG

config = DEFAULT_CONFIG.copy()
config["llm_provider"] = "anthropic"          # Use Claude
config["deep_think_llm"] = "claude-opus-4-6"  # Complex reasoning
config["quick_think_llm"] = "claude-haiku-4-5-20251001"  # Fast tasks
config["max_debate_rounds"] = 3               # Researcher debate depth
config["online_tools"] = True                 # Live market data

ta = TradingAgentsGraph(debug=True, config=config)
state, decision = ta.propagate("NVDA", "2026-01-15")
print(decision)  # BUY / SELL / HOLD + rationale

---

LLM Provider Configuration

Provider	llm_provider	Models
Anthropic	"anthropic"	claude-opus-4-6, claude-sonnet-4-6, claude-haiku-4-5
OpenAI	"openai"	gpt-4o, gpt-4o-mini, o1
Google	"google"	gemini-2.0-flash, gemini-1.5-pro
xAI	"xai"	grok-2
OpenRouter	"openrouter"	Any model via router
Ollama	"ollama"	Local models (llama3, mistral, etc.)
DeepSeek	"deepseek"	deepseek-chat (CN fork)
Alibaba	"alibaba"	qwen models (CN fork)

---

CN Fork (TradingAgents-CN) — Key Enhancements

Architecture Upgrade

• Original: Streamlit UI
• CN Fork: FastAPI + Vue 3 (enterprise-grade)

Regional Market Support

Market	Data Source
A-shares (China)	Tushare, AkShare, BaoStock
HK Stocks	AkShare
US Equities	Alpha Vantage

Additional Features

• MongoDB + Redis dual database (persistent sessions, caching)
• Docker support (amd64 + ARM64)
• Report export: Markdown, Word, PDF
• Batch portfolio analysis
• SSE + WebSocket real-time progress
• News quality filtering + multi-layer assessment
• User auth + operation logging

CN Fork Installation

git clone https://github.com/hsliuping/TradingAgents-CN.git
cd TradingAgents-CN
docker-compose up -d  # Easiest path (MongoDB + Redis included)
# or
pip install -r requirements.txt

---

Trading Workflow (Step-by-Step)

1. Input: ticker + date
2. Analysts run in parallel → 4 reports
3. Researcher debate (N rounds) → bull/bear synthesis
4. Trader synthesizes → trade proposal (BUY/SELL/HOLD + size)
5. Risk manager evaluates volatility + liquidity
6. Portfolio manager: APPROVE or REJECT
7. Output: final decision + reasoning chain

---

Integration With Claude

Use Claude as the reasoning backbone:

config = {
    "llm_provider": "anthropic",
    "deep_think_llm": "claude-opus-4-6",    # Analyst/Researcher deep work
    "quick_think_llm": "claude-sonnet-4-6", # Fast classification tasks
    "max_debate_rounds": 2,
    "online_tools": True,
}

Claude's strength in structured reasoning makes it ideal for:

• Fundamental analysis reports (long-form reasoning)
• Researcher debate synthesis
• Risk rationale explanation

---

Key Design Patterns (for building similar agents)

# Pattern: Analyst role definition
analyst_prompt = """
You are a Fundamental Analyst. Evaluate the company's:
- Revenue growth, margins, P/E, debt ratios
- Competitive moat and sector dynamics
Return: structured report with BUY/NEUTRAL/SELL signal + confidence
"""

# Pattern: Debate orchestration (LangGraph)
from langgraph.graph import StateGraph

graph = StateGraph(TradingState)
graph.add_node("fundamentals_analyst", run_fundamentals)
graph.add_node("sentiment_analyst", run_sentiment)
graph.add_node("researcher_debate", run_debate)
graph.add_node("trader_decision", run_trader)
graph.add_node("risk_check", run_risk_manager)
graph.add_edge("fundamentals_analyst", "researcher_debate")
# ...

---