tradingagents-astock-multi-agent-framework

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

TradingAgents-Astock Multi-Agent Framework

TradingAgents-Astock 多Agent框架

Skill by ara.so — AI Agent Skills collection.

由 ara.so 提供的Skill — AI Agent技能集合。

Overview

概述

TradingAgents-Astock is a multi-agent investment research framework specifically adapted for Chinese A-share markets. It orchestrates 7 specialized AI analyst agents that generate research reports, engage in bull/bear debates, perform risk assessment, and produce trading decisions. The framework handles A-share specific constraints (T+1 settlement, price limits, minimum lots) and uses free Chinese data sources (mootdx, EastMoney, Sina, THS) instead of Western APIs.

Key Features:

7 specialized analysts (Market, Social, News, Fundamentals, Policy, Hot Money, Lockup)
Bull vs Bear research debate system
3-way risk assessment (Aggressive, Conservative, Neutral)
A-share trading constraints (T+1, 涨跌停, minimum lots, ST rules)
Dual LLM architecture (quick_think + deep_think)
Web UI with real-time progress tracking
Chinese output with English internal reasoning

TradingAgents-Astock是专门适配中国A股市场的多Agent投资研究框架。它协调7个专业AI分析师Agent生成研究报告、开展多空辩论、进行风险评估并生成交易决策。该框架适配A股特有的约束条件（T+1结算、涨跌停限制、最小交易单位），并使用免费的中文数据源（mootdx、东方财富、新浪、同花顺）而非西方API。

核心特性：

7个专业分析师（市场、舆情、新闻、基本面、政策、游资、限售股）
多空研究辩论系统
三元风险评估（激进型、保守型、中性型）
A股交易约束（T+1、涨跌停、最小交易单位、ST规则）
双LLM架构（quick_think + deep_think）
带实时进度追踪的Web UI
中文输出搭配英文内部推理逻辑

Installation

安装

bash

undefined

bash

undefined

Clone the repository

克隆仓库

git clone https://github.com/simonlin1212/tradingagents-astock.git cd tradingagents-astock

Install base package (Python >= 3.10)

安装基础包（Python >= 3.10）

pip install -e .

Optional: Install with Google Gemini support

可选：安装Google Gemini支持

pip install -e ".[google]"

undefined

pip install -e ".[google]"

undefined

Configuration

配置

LLM Provider Setup

LLM提供商设置

Create a

.env

file in the project root with your chosen LLM provider:

bash

undefined

在项目根目录创建

.env

文件，填入所选LLM提供商的密钥：

bash

undefined

MiniMax (Recommended for China, cost-effective)

MiniMax（推荐国内使用，性价比高）

MINIMAX_API_KEY=sk-your-key-here

DeepSeek

DEEPSEEK_API_KEY=sk-your-key-here

Zhipu GLM

智谱GLM

ZHIPU_API_KEY=your-key-here

Alibaba Qwen

阿里通义千问

DASHSCOPE_API_KEY=sk-your-key-here

OpenAI

OPENAI_API_KEY=sk-your-key-here

Anthropic

ANTHROPIC_API_KEY=sk-ant-your-key-here

Kimi (uses Anthropic-compatible API)

Kimi（兼容Anthropic API）

ANTHROPIC_AUTH_TOKEN=your-kimi-token

undefined

ANTHROPIC_AUTH_TOKEN=your-kimi-token

undefined

Graph Configuration Object

图配置对象

python

config = {
    "llm_provider": "minimax",              # Provider: minimax, deepseek, zhipu, qwen, openai, anthropic, google, xai, ollama
    "deep_think_llm": "MiniMax-M2.7",       # Model for Research Manager & Portfolio Manager
    "quick_think_llm": "MiniMax-M2.7-highspeed",  # Model for analysts, researchers, traders
    "output_language": "Chinese",            # Final report language (Chinese/English)
    "backend_url": None,                    # Optional: Custom API endpoint
    "max_debate_rounds": 3,                 # Bull/Bear debate iterations
    "enable_policy_analyst": True,          # A-share specific analysts
    "enable_hotmoney_analyst": True,
    "enable_lockup_analyst": True,
}

python

config = {
    "llm_provider": "minimax",              # 提供商：minimax, deepseek, zhipu, qwen, openai, anthropic, google, xai, ollama
    "deep_think_llm": "MiniMax-M2.7",       # 用于研究经理和投资组合经理的模型
    "quick_think_llm": "MiniMax-M2.7-highspeed",  # 用于分析师、研究员、交易员的模型
    "output_language": "Chinese",            # 最终报告语言（Chinese/English）
    "backend_url": None,                    # 可选：自定义API端点
    "max_debate_rounds": 3,                 # 多空辩论轮次
    "enable_policy_analyst": True,          # 启用A股专属分析师
    "enable_hotmoney_analyst": True,
    "enable_lockup_analyst": True,
}

Core API Usage

核心API使用

Basic Analysis

基础分析

python

from tradingagents.graph.trading_graph import TradingAgentsGraph

python

from tradingagents.graph.trading_graph import TradingAgentsGraph

Initialize with MiniMax

使用MiniMax初始化

config = { "llm_provider": "minimax", "deep_think_llm": "MiniMax-M2.7", "quick_think_llm": "MiniMax-M2.7-highspeed", "output_language": "Chinese", }

ta = TradingAgentsGraph(debug=True, config=config)

config = { "llm_provider": "minimax", "deep_think_llm": "MiniMax-M2.7", "quick_think_llm": "MiniMax-M2.7-highspeed", "output_language": "Chinese", }

ta = TradingAgentsGraph(debug=True, config=config)

Run analysis for stock 688017 on 2026-05-12

对股票688017在2026-05-12进行分析

final_state, decision = ta.propagate("688017", "2026-05-12")

Access decision

获取决策结果

print(f"Signal: {decision['signal']}") # BUY, HOLD, or SELL print(f"Confidence: {decision['confidence']}") # 0-100 print(f"Position: {decision['position_size']}") # Suggested position size print(f"Reasoning: {decision['reasoning']}")

undefined

print(f"信号: {decision['signal']}") # BUY, HOLD, 或 SELL print(f"置信度: {decision['confidence']}") # 0-100 print(f"仓位: {decision['position_size']}") # 建议仓位大小 print(f"推理逻辑: {decision['reasoning']}")

undefined

Using DeepSeek

使用DeepSeek

python

config = {
    "llm_provider": "deepseek",
    "deep_think_llm": "deepseek-chat",
    "quick_think_llm": "deepseek-chat",
    "output_language": "Chinese",
}

ta = TradingAgentsGraph(config=config)
final_state, decision = ta.propagate("600519", "2026-05-15")

python

config = {
    "llm_provider": "deepseek",
    "deep_think_llm": "deepseek-chat",
    "quick_think_llm": "deepseek-chat",
    "output_language": "Chinese",
}

ta = TradingAgentsGraph(config=config)
final_state, decision = ta.propagate("600519", "2026-05-15")

Using Anthropic with Kimi Backend

使用Anthropic搭配Kimi后端

python

config = {
    "llm_provider": "anthropic",
    "deep_think_llm": "claude-sonnet-4-6",
    "quick_think_llm": "claude-sonnet-4-6",
    "backend_url": "https://api.kimi.com/coding/",
    "output_language": "Chinese",
}

ta = TradingAgentsGraph(config=config)
final_state, decision = ta.propagate("000001", "2026-05-20")

python

config = {
    "llm_provider": "anthropic",
    "deep_think_llm": "claude-sonnet-4-6",
    "quick_think_llm": "claude-sonnet-4-6",
    "backend_url": "https://api.kimi.com/coding/",
    "output_language": "Chinese",
}

ta = TradingAgentsGraph(config=config)
final_state, decision = ta.propagate("000001", "2026-05-20")

Accessing Individual Analyst Reports

获取单个分析师报告

python

final_state, decision = ta.propagate("688017", "2026-05-12")

python

final_state, decision = ta.propagate("688017", "2026-05-12")

Access 7 analyst reports

获取7份分析师报告

market_report = final_state["market_analyst_report"] social_report = final_state["social_analyst_report"] news_report = final_state["news_analyst_report"] fundamentals_report = final_state["fundamentals_analyst_report"] policy_report = final_state["policy_analyst_report"] hotmoney_report = final_state["hotmoney_analyst_report"] lockup_report = final_state["lockup_analyst_report"]

Access debate and risk assessment

获取辩论和风险评估结果

bull_research = final_state["bull_researcher_report"] bear_research = final_state["bear_researcher_report"] risk_assessment = final_state["risk_assessment"] trading_plan = final_state["trading_plan"]

undefined

bull_research = final_state["bull_researcher_report"] bear_research = final_state["bear_researcher_report"] risk_assessment = final_state["risk_assessment"] trading_plan = final_state["trading_plan"]

undefined

CLI Commands

CLI命令

Interactive Mode

交互模式

bash

undefined

bash

undefined

Launch interactive CLI

启动交互式CLI

tradingagents

CLI will prompt for:

CLI会提示输入：

- Stock code (e.g., 688017, 600519)

- 股票代码（例如：688017, 600519）

- Analysis date (YYYY-MM-DD)

- 分析日期（YYYY-MM-DD）

- LLM provider selection

- LLM提供商选择

undefined

undefined

Direct Execution

直接执行

bash

undefined

bash

undefined

Analyze with specific parameters

使用指定参数进行分析

tradingagents --stock 688017 --date 2026-05-12 --provider minimax

Use different models

使用不同模型

tradingagents --stock 600519 --date 2026-05-15
--provider deepseek
--deep-model deepseek-chat
--quick-model deepseek-chat

Enable debug output

启用调试输出

tradingagents --stock 000001 --date 2026-05-20 --debug

undefined

tradingagents --stock 000001 --date 2026-05-20 --debug

undefined

Web UI

bash

undefined

bash

undefined

Launch Streamlit web interface

启动Streamlit Web界面

tradingagents-web

Alternative

替代方式

streamlit run web/app.py

Access at http://localhost:8501

访问地址：http://localhost:8501

undefined

undefined

Data Source Integration

数据源集成

Available Data Tools

可用数据工具

The framework provides these tools to analysts (all free, no API keys needed):

python

undefined

框架为分析师提供以下工具（均免费，无需API密钥）：

python

undefined

Market data (OHLCV, indicators)

市场数据（OHLCV、指标）

get_stock_data(ticker, start_date, end_date) get_indicators(ticker, start_date, end_date)

Fundamental data

基本面数据

get_fundamentals(ticker) get_balance_sheet(ticker) get_cashflow(ticker) get_income_statement(ticker)

News and sentiment

新闻与舆情

get_news(ticker, start_date, end_date) get_global_news(start_date, end_date)

A-share specific

A股专属数据

get_insider_transactions(ticker) # Lockup releases, insider trading

get_insider_transactions(ticker) # 限售股解禁、内幕交易

Dragon-Tiger list data available via get_news

龙虎榜数据可通过get_news获取

undefined

undefined

Data Provider Mapping

数据源映射

Data Type	Provider	Protocol
OHLCV K-lines	mootdx	TCP 7709
PE/PB/Market Cap	Tencent Finance	HTTP
Dragon-Tiger List	EastMoney	HTTP
Lockup Schedule	EastMoney	HTTP
Financial Statements	Sina Finance	HTTP
EPS Consensus	THS (10jqka)	HTTP
News Feed	CLS.cn	HTTP
Sector Classification	Baidu Finance	HTTP

数据类型	提供商	协议
OHLCK线	mootdx	TCP 7709
PE/PB/市值	腾讯财经	HTTP
龙虎榜	东方财富	HTTP
限售股时间表	东方财富	HTTP
财务报表	新浪财经	HTTP
EPS一致预期	同花顺（10jqka）	HTTP
新闻资讯	CLS.cn	HTTP
行业分类	百度财经	HTTP

Agent Pipeline Architecture

Agent流水线架构

12-Stage Execution Flow

12阶段执行流程

python

undefined

python

undefined

Stage 1-7: Analyst Reports

阶段1-7：分析师报告

stages = [ "market_analyst", # Technical analysis "social_analyst", # Social sentiment "news_analyst", # News and events "fundamentals_analyst", # Financial statements "policy_analyst", # Regulatory policy (A-share specific) "hotmoney_analyst", # Dragon-Tiger list tracking (A-share specific) "lockup_analyst", # Share lockup monitoring (A-share specific) ]

stages = [ "market_analyst", # 技术分析 "social_analyst", # 舆情分析 "news_analyst", # 新闻与事件分析 "fundamentals_analyst", # 财务报表分析 "policy_analyst", # 监管政策分析（A股专属） "hotmoney_analyst", # 龙虎榜追踪（A股专属） "lockup_analyst", # 限售股监控（A股专属） ]

Stage 8: Quality Gate

阶段8：质量校验

Checks if analysts provided sufficient data

检查分析师是否提供了足够的数据

Stage 9-10: Bull/Bear Debate

阶段9-10：多空辩论

Bull and Bear researchers debate up to N rounds

多头和空头研究员进行最多N轮辩论

Stage 11: Research Manager

阶段11：研究经理

Deep-think LLM synthesizes debate into investment plan

深度思考LLM将辩论内容整合为投资方案

Stage 12: Trader + Risk Assessment

阶段12：交易员 + 风险评估

Trader proposes execution plan

交易员提出执行方案

3 risk debaters (Aggressive, Conservative, Neutral) assess

3位风险辩论者（激进型、保守型、中性型）进行评估

Stage 13: Portfolio Manager

阶段13：投资组合经理

Deep-think LLM makes final BUY/HOLD/SELL decision

深度思考LLM做出最终的买入/持有/卖出决策

undefined

undefined

Custom Analyst Configuration

自定义分析师配置

python

undefined

python

undefined

Disable A-share specific analysts

禁用A股专属分析师

config = { "llm_provider": "minimax", "deep_think_llm": "MiniMax-M2.7", "quick_think_llm": "MiniMax-M2.7-highspeed", "output_language": "Chinese", "enable_policy_analyst": False, "enable_hotmoney_analyst": False, "enable_lockup_analyst": False, }

ta = TradingAgentsGraph(config=config)

undefined

ta = TradingAgentsGraph(config=config)

undefined

A-Share Trading Constraints

A股交易约束

The framework automatically applies A-share market rules:

python

undefined

框架会自动应用A股市场规则：

python

undefined

T+1 Settlement

T+1结算

Cannot sell shares bought today

当日买入的股票无法卖出

Price Limits

涨跌停限制

ST stocks: ±5%

ST股：±5%

Regular stocks: ±10% (or ±20% for ChiNext/STAR)

普通股票：±10%（创业板/科创板为±20%）

Minimum Lot Size

最小交易单位

100 shares (1 lot)

100股（1手）

Must trade in multiples of 100

必须以100股的倍数进行交易

Trading Hours

交易时间

Morning: 09:30-11:30

上午：09:30-11:30

Afternoon: 13:00-15:00

下午：13:00-15:00

Call auction: 09:15-09:25, 14:57-15:00

集合竞价：09:15-09:25、14:57-15:00


These constraints are built into the Trader agent's decision logic.


这些约束已内置到交易员Agent的决策逻辑中。

Common Patterns

常见用法

Batch Analysis

批量分析

python

from tradingagents.graph.trading_graph import TradingAgentsGraph
from datetime import datetime, timedelta

config = {
    "llm_provider": "minimax",
    "deep_think_llm": "MiniMax-M2.7",
    "quick_think_llm": "MiniMax-M2.7-highspeed",
    "output_language": "Chinese",
}

ta = TradingAgentsGraph(config=config)

python

from tradingagents.graph.trading_graph import TradingAgentsGraph
from datetime import datetime, timedelta

config = {
    "llm_provider": "minimax",
    "deep_think_llm": "MiniMax-M2.7",
    "quick_think_llm": "MiniMax-M2.7-highspeed",
    "output_language": "Chinese",
}

ta = TradingAgentsGraph(config=config)

Analyze portfolio of stocks

分析股票组合

stocks = ["600519", "000858", "600036", "601318"] date = "2026-05-15"

results = {} for ticker in stocks: try: final_state, decision = ta.propagate(ticker, date) results[ticker] = decision print(f"{ticker}: {decision['signal']} (confidence: {decision['confidence']})") except Exception as e: print(f"Error analyzing {ticker}: {e}") continue

undefined

stocks = ["600519", "000858", "600036", "601318"] date = "2026-05-15"

results = {} for ticker in stocks: try: final_state, decision = ta.propagate(ticker, date) results[ticker] = decision print(f"{ticker}: {decision['signal']} (置信度: {decision['confidence']})") except Exception as e: print(f"分析{ticker}出错: {e}") continue

undefined

Time Series Analysis

时间序列分析

python

from datetime import datetime, timedelta

ta = TradingAgentsGraph(config=config)
ticker = "688017"

python

from datetime import datetime, timedelta

ta = TradingAgentsGraph(config=config)
ticker = "688017"

Analyze over 5 trading days

分析连续5个交易日

base_date = datetime(2026, 5, 10) signals = []

for i in range(5): analysis_date = (base_date + timedelta(days=i)).strftime("%Y-%m-%d") try: final_state, decision = ta.propagate(ticker, analysis_date) signals.append({ "date": analysis_date, "signal": decision["signal"], "confidence": decision["confidence"], }) except Exception as e: print(f"Error on {analysis_date}: {e}") continue

base_date = datetime(2026, 5, 10) signals = []

Track signal consistency

追踪信号一致性

print(f"Signal history for {ticker}:") for s in signals: print(f"{s['date']}: {s['signal']} ({s['confidence']}%)")

undefined

print(f"{ticker}的信号历史:") for s in signals: print(f"{s['date']}: {s['signal']} ({s['confidence']}%)")

undefined

Custom LLM Backend

自定义LLM后端

python

undefined

python

undefined

Use local Ollama instance

使用本地Ollama实例

config = { "llm_provider": "ollama", "deep_think_llm": "qwen2.5:32b", "quick_think_llm": "qwen2.5:14b", "backend_url": "http://localhost:11434", "output_language": "Chinese", }

ta = TradingAgentsGraph(config=config) final_state, decision = ta.propagate("600519", "2026-05-15")

undefined

config = { "llm_provider": "ollama", "deep_think_llm": "qwen2.5:32b", "quick_think_llm": "qwen2.5:14b", "backend_url": "http://localhost:11434", "output_language": "Chinese", }

ta = TradingAgentsGraph(config=config) final_state, decision = ta.propagate("600519", "2026-05-15")

undefined

Extracting Structured Data

提取结构化数据

python

final_state, decision = ta.propagate("688017", "2026-05-12")

python

final_state, decision = ta.propagate("688017", "2026-05-12")

Extract key metrics from analyst reports

从分析师报告中提取关键指标

def extract_pe_ratio(fundamentals_report): # Parse PE from report text import re match = re.search(r'PE.*?(\d+.\d+)', fundamentals_report) return float(match.group(1)) if match else None

def extract_lockup_events(lockup_report): # Parse lockup schedule events = [] lines = lockup_report.split('\n') for line in lines: if '解禁' in line: events.append(line.strip()) return events

pe_ratio = extract_pe_ratio(final_state["fundamentals_analyst_report"]) lockup_events = extract_lockup_events(final_state["lockup_analyst_report"])

print(f"PE Ratio: {pe_ratio}") print(f"Upcoming lockups: {lockup_events}")

undefined

def extract_pe_ratio(fundamentals_report): # 从报告文本中解析PE值 import re match = re.search(r'PE.*?(\d+.\d+)', fundamentals_report) return float(match.group(1)) if match else None

def extract_lockup_events(lockup_report): # 解析限售股解禁时间表 events = [] lines = lockup_report.split('\n') for line in lines: if '解禁' in line: events.append(line.strip()) return events

pe_ratio = extract_pe_ratio(final_state["fundamentals_analyst_report"]) lockup_events = extract_lockup_events(final_state["lockup_analyst_report"])

print(f"PE比率: {pe_ratio}") print(f"即将到来的解禁事件: {lockup_events}")

undefined

Troubleshooting

故障排除

API Key Issues

API密钥问题

python

undefined

python

undefined

Verify env vars are loaded

验证环境变量是否加载

import os from dotenv import load_dotenv

load_dotenv() print(f"MINIMAX_API_KEY exists: {bool(os.getenv('MINIMAX_API_KEY'))}")

import os from dotenv import load_dotenv

load_dotenv() print(f"MINIMAX_API_KEY已存在: {bool(os.getenv('MINIMAX_API_KEY'))}")

If still failing, pass key directly (not recommended for production)

如果仍失败，直接传入密钥（不推荐生产环境使用）

config = { "llm_provider": "minimax", "api_key": "sk-your-key-here", # Override env var "deep_think_llm": "MiniMax-M2.7", "quick_think_llm": "MiniMax-M2.7-highspeed", }

undefined

config = { "llm_provider": "minimax", "api_key": "sk-your-key-here", # 覆盖环境变量 "deep_think_llm": "MiniMax-M2.7", "quick_think_llm": "MiniMax-M2.7-highspeed", }

undefined

Data Source Failures

数据源获取失败

python

undefined

python

undefined

Test data connectivity

测试数据连接

from tradingagents.tools.astock_tools import get_stock_data

try: data = get_stock_data("688017", "2026-04-01", "2026-05-01") print(f"Successfully fetched {len(data)} rows") except Exception as e: print(f"Data fetch failed: {e}") # Fallback: Check network, try different date range

undefined

from tradingagents.tools.astock_tools import get_stock_data

try: data = get_stock_data("688017", "2026-04-01", "2026-05-01") print(f"成功获取{len(data)}条数据") except Exception as e: print(f"数据获取失败: {e}") # 备选方案：检查网络，尝试不同日期范围

undefined

Invalid Stock Code

无效股票代码

python

undefined

python

undefined

Validate A-share ticker format

验证A股代码格式

import re

def validate_astock_ticker(ticker): # A-share codes: 6 digits # Shanghai: 600xxx, 601xxx, 603xxx, 688xxx (STAR) # Shenzhen: 000xxx, 001xxx, 002xxx, 003xxx, 300xxx (ChiNext) pattern = r'^(600|601|603|688|000|001|002|003|300)\d{3}$' return bool(re.match(pattern, ticker))

ticker = "688017" if not validate_astock_ticker(ticker): print(f"Invalid ticker: {ticker}")

undefined

import re

def validate_astock_ticker(ticker): # A股代码：6位数字 # 沪市：600xxx, 601xxx, 603xxx, 688xxx（科创板） # 深市：000xxx, 001xxx, 002xxx, 003xxx, 300xxx（创业板） pattern = r'^(600|601|603|688|000|001|002|003|300)\d{3}$' return bool(re.match(pattern, ticker))

ticker = "688017" if not validate_astock_ticker(ticker): print(f"无效代码: {ticker}")

undefined

Rate Limiting

请求速率限制

python

undefined

python

undefined

Add delays for batch analysis

批量分析时添加延迟

import time

ta = TradingAgentsGraph(config=config) tickers = ["600519", "000858", "600036"]

for ticker in tickers: final_state, decision = ta.propagate(ticker, "2026-05-15") print(f"{ticker}: {decision['signal']}") time.sleep(5) # 5 second delay between requests

undefined

import time

ta = TradingAgentsGraph(config=config) tickers = ["600519", "000858", "600036"]

for ticker in tickers: final_state, decision = ta.propagate(ticker, "2026-05-15") print(f"{ticker}: {decision['signal']}") time.sleep(5) # 请求间隔5秒

undefined

Debug Mode

调试模式

python

undefined

python

undefined

Enable verbose logging

启用详细日志

ta = TradingAgentsGraph(debug=True, config=config) final_state, decision = ta.propagate("688017", "2026-05-12")

Inspect intermediate states

查看中间状态

print("Market Analyst Report:") print(final_state["market_analyst_report"]) print("\nBull Research:") print(final_state["bull_researcher_report"]) print("\nBear Research:") print(final_state["bear_researcher_report"])

undefined

print("市场分析师报告:") print(final_state["market_analyst_report"]) print("\n多头研究报告:") print(final_state["bull_researcher_report"]) print("\n空头研究报告:") print(final_state["bear_researcher_report"])

undefined

LLM Response Parsing Errors

LLM响应解析错误

python

undefined

python

undefined

If LLM returns malformed JSON, enable retry logic

如果LLM返回格式错误的JSON，启用重试逻辑

config = { "llm_provider": "minimax", "deep_think_llm": "MiniMax-M2.7", "quick_think_llm": "MiniMax-M2.7-highspeed", "max_retries": 3, # Retry up to 3 times on parse errors "temperature": 0.7, # Lower temperature for more consistent formatting }

ta = TradingAgentsGraph(config=config)

undefined

config = { "llm_provider": "minimax", "deep_think_llm": "MiniMax-M2.7", "quick_think_llm": "MiniMax-M2.7-highspeed", "max_retries": 3, # 解析错误时最多重试3次 "temperature": 0.7, # 降低温度以获得更一致的格式 }

ta = TradingAgentsGraph(config=config)

undefined

Project Structure

项目结构

tradingagents-astock/
├── tradingagents/
│   ├── graph/
│   │   ├── trading_graph.py      # Main TradingAgentsGraph class
│   │   └── nodes.py               # Individual agent node implementations
│   ├── tools/
│   │   ├── astock_tools.py        # A-share data fetching tools
│   │   └── tool_registry.py       # Tool registration system
│   ├── llm/
│   │   ├── llm_factory.py         # LLM provider abstraction
│   │   └── providers/             # Provider-specific implementations
│   └── agents/
│       ├── analysts/              # 7 analyst agent prompts
│       ├── researchers/           # Bull/Bear researchers
│       ├── risk/                  # 3 risk debaters
│       └── portfolio_manager/     # Final decision maker
├── web/
│   └── app.py                     # Streamlit UI
├── .env.example                   # Environment variable template
└── README.md

tradingagents-astock/
├── tradingagents/
│   ├── graph/
│   │   ├── trading_graph.py      # 核心TradingAgentsGraph类
│   │   └── nodes.py               # 单个Agent节点实现
│   ├── tools/
│   │   ├── astock_tools.py        # A股数据获取工具
│   │   └── tool_registry.py       # 工具注册系统
│   ├── llm/
│   │   ├── llm_factory.py         # LLM提供商抽象层
│   │   └── providers/             # 各提供商的具体实现
│   └── agents/
│       ├── analysts/              # 7个分析师Agent的提示词
│       ├── researchers/           # 多空研究员
│       ├── risk/                  # 3位风险辩论者
│       └── portfolio_manager/     # 最终决策制定者
├── web/
│   └── app.py                     # Streamlit UI应用
├── .env.example                   # 环境变量模板
└── README.md

Best Practices

最佳实践

Always use environment variables for API keys
Start with debug=True to understand the pipeline
Use MiniMax or DeepSeek for cost-effective China-based inference
Check data availability before running batch analyses (market holidays, weekends)
Monitor LLM costs — each analysis requires 30-50 API calls
Validate stock codes before passing to
```
propagate()
```
Set appropriate
max_debate_rounds
— more rounds = higher cost but potentially better decisions
Use quick_think_llm for analysts and researchers, reserve deep_think_llm for final decisions

始终使用环境变量存储API密钥
启动时开启debug=True以理解流水线运行逻辑
使用MiniMax或DeepSeek进行高性价比的国内推理
批量分析前检查数据可用性（市场节假日、周末）
监控LLM成本 — 每次分析需要30-50次API调用
传入propagate()前验证股票代码
设置合适的max_debate_rounds — 轮次越多成本越高，但决策可能更优
分析师和研究员使用quick_think_llm，将deep_think_llm留作最终决策使用