182 lines
5.1 KiB
Markdown
182 lines
5.1 KiB
Markdown
---
|
||
created: "2026-03-21"
|
||
type: resource
|
||
tags: [trading, multi-agent, langgraph, architecture, reference]
|
||
source: "https://github.com/TauricResearch/TradingAgents"
|
||
---
|
||
|
||
# TradingAgents 原始架构分析
|
||
|
||
## 概述
|
||
|
||
[TradingAgents](https://github.com/TauricResearch/TradingAgents) 是一个基于 LangGraph 的多角色 LLM 交易决策框架，模拟真实交易公司的协作流程：分析师团队 → 多空辩论 → 交易决策 → 风控辩论 → 最终裁决。
|
||
|
||
---
|
||
|
||
## 核心文件结构
|
||
|
||
| 文件 | 职责 |
|
||
|------|------|
|
||
| `tradingagents/graph/trading_graph.py` | 主入口，编排整个系统 |
|
||
| `tradingagents/graph/setup.py` | 构建 LangGraph 状态机 |
|
||
| `tradingagents/graph/conditional_logic.py` | 条件路由（工具调用 / 辩论轮次） |
|
||
| `tradingagents/agents/analysts/*.py` | 4 个分析师（市场/社媒/新闻/基本面） |
|
||
| `tradingagents/agents/researchers/*.py` | Bull/Bear 研究员 + Research Manager |
|
||
| `tradingagents/agents/risk_mgmt/*.py` | 激进/保守/中性风控辩论者 |
|
||
| `tradingagents/agents/trader/trader.py` | 交易员 |
|
||
| `tradingagents/agents/utils/agent_states.py` | TypedDict 状态定义 |
|
||
| `tradingagents/agents/utils/memory.py` | BM25 记忆系统 |
|
||
| `tradingagents/dataflows/interface.py` | 数据供应商路由 + 降级策略 |
|
||
| `tradingagents/dataflows/y_finance.py` | yfinance 数据实现 |
|
||
| `tradingagents/graph/reflection.py` | 交易后复盘，更新记忆 |
|
||
| `tradingagents/graph/signal_processing.py` | 从冗长输出提取 BUY/SELL/HOLD |
|
||
|
||
---
|
||
|
||
## 分析师团队（并行执行）
|
||
|
||
### Market Analyst
|
||
- 技术分析：SMA, EMA, MACD, RSI, Bollinger Bands, ATR, VWMA
|
||
- 工具：`get_stock_data`, `get_indicators`
|
||
- 数据源：yfinance + stockstats
|
||
|
||
### Social Media Analyst
|
||
- 公司新闻 + 情感分析
|
||
- 工具：`get_news`
|
||
|
||
### News Analyst
|
||
- 宏观经济 + 世界大事
|
||
- 工具：`get_news`, `get_global_news`
|
||
|
||
### Fundamentals Analyst
|
||
- 财务报表 + 公司指标
|
||
- 工具：`get_fundamentals`, `get_balance_sheet`, `get_cashflow`, `get_income_statement`
|
||
|
||
---
|
||
|
||
## 辩论机制
|
||
|
||
### 投资辩论（Bull vs Bear）
|
||
|
||
1. Bull Researcher 阅读所有报告 + 检索历史相似情况 → 提出看多论点
|
||
2. Bear Researcher 阅读报告 + 历史 → 反驳
|
||
3. 循环 `max_debate_rounds` 轮（默认 1 轮）
|
||
4. Research Manager（裁判）读完整辩论历史 + 过去的裁决记忆 → 输出投资方案
|
||
|
||
### 风控辩论（三方）
|
||
|
||
1. Aggressive Analyst：高风险高回报
|
||
2. Conservative Analyst：稳健保守
|
||
3. Neutral Analyst：平衡两方
|
||
4. 循环 `max_risk_discuss_rounds` 轮（默认 1 轮 × 3 人 = 3 turns）
|
||
5. Risk Manager（裁判）修正交易方案 → 最终决策
|
||
|
||
### 辩论路由逻辑
|
||
|
||
```python
|
||
# conditional_logic.py
|
||
def should_continue_debate(state):
|
||
# 辩论历史长度 >= max_debate_rounds * 2 → 结束
|
||
# 否则交替 Bull → Bear → Bull → Bear
|
||
|
||
def should_continue_risk_analysis(state):
|
||
# 长度 >= max_risk_discuss_rounds * 3 → 结束
|
||
# 循环 Aggressive → Conservative → Neutral
|
||
```
|
||
|
||
---
|
||
|
||
## BM25 记忆系统
|
||
|
||
### 结构
|
||
|
||
```python
|
||
class FinancialSituationMemory:
|
||
# 存储: (market_situation_description, recommendation) 元组列表
|
||
# 检索: BM25 相似度，取 top-n 最相似历史情况
|
||
# 无需 API，完全离线
|
||
```
|
||
|
||
### 5 个独立记忆库
|
||
|
||
1. Bull Researcher Memory
|
||
2. Bear Researcher Memory
|
||
3. Research Manager Memory
|
||
4. Trader Memory
|
||
5. Risk Manager Memory
|
||
|
||
### 复盘机制
|
||
|
||
```python
|
||
# reflection.py
|
||
class Reflector:
|
||
def reflect_and_remember(self, returns_losses):
|
||
# 分析交易结果（盈亏）
|
||
# 生成教训总结
|
||
# 更新各角色的记忆库
|
||
```
|
||
|
||
---
|
||
|
||
## LangGraph 状态定义
|
||
|
||
```python
|
||
class AgentState(TypedDict):
|
||
company_of_interest: str # 股票代码
|
||
trade_date: str # 交易日期
|
||
market_report: str # 市场分析报告
|
||
sentiment_report: str # 情感分析报告
|
||
news_report: str # 新闻分析报告
|
||
fundamentals_report: str # 基本面报告
|
||
invest_debate_state: dict # 多空辩论状态
|
||
risk_debate_state: dict # 风控辩论状态
|
||
final_trade_decision: str # 最终决策
|
||
```
|
||
|
||
---
|
||
|
||
## 数据供应商抽象层
|
||
|
||
```python
|
||
# interface.py - 供应商路由
|
||
def route_to_vendor(tool_name, category, params):
|
||
vendor = config["data_vendors"].get(category, "yfinance")
|
||
try:
|
||
return vendors[vendor][tool_name](params)
|
||
except RateLimit:
|
||
return vendors["yfinance"][tool_name](params) # 降级
|
||
```
|
||
|
||
支持的供应商：
|
||
- yfinance（默认，免费）
|
||
- Alpha Vantage（需 API key）
|
||
- 按 category 级别配置，支持 tool 级别覆盖
|
||
|
||
---
|
||
|
||
## 调用方式
|
||
|
||
```python
|
||
ta = TradingAgentsGraph(config={
|
||
"llm_provider": "openai",
|
||
"deep_think_llm": "gpt-5.2",
|
||
"quick_think_llm": "gpt-5-mini",
|
||
"max_debate_rounds": 1,
|
||
"max_risk_discuss_rounds": 1,
|
||
})
|
||
final_state, decision = ta.propagate("NVDA", "2024-05-10")
|
||
# decision = "BUY" | "SELL" | "HOLD"
|
||
|
||
# 复盘
|
||
ta.reflect_and_remember(returns_losses=1500)
|
||
```
|
||
|
||
---
|
||
|
||
## Related
|
||
|
||
- [[Trading Agents 混合架构方案]]
|
||
- [[Trading Agents 使用指南]]
|
||
- [[openbb-invest-api]]
|
||
- [[OpenClaw 部署配置分析]]
|