Skip to content

xhqing/MarketForecastingAgents

BranchesTags

Repository files navigation

MarketForecastingAgents

An AI-powered market research report auto-generation system based on multi-agent collaboration, covering A-shares, Hong Kong stocks, and US stocks.

中文文档

Features

  • Multi-Agent Collaboration: Leverages four specialized investment agents — Livermore (trend-following), Buffett (value investing), Cathie Wood (disruptive innovation), and Ray Dalio (macro cycle analysis) — each with distinct analytical philosophies, providing multi-perspective market assessments.
  • Full Market Coverage: Supports A-shares (SSE/SZSE), Hong Kong stocks (HKEX), and US stocks (NYSE/NASDAQ), including indices, individual stocks, and ETFs.
  • Real-Time Data Integration: Connects to multiple MCP (Model Context Protocol) servers for real-time market data, financial statements, technical indicators, sector fund flows, and research reports.
  • Automated Validation: Built-in targets_validator.py ensures generated reports fully cover all configured targets with zero omissions, and price_checker.py ensures all target prices are accurate and reasonable with zero price errors.
  • Flexible Target Configuration: All analysis targets are managed through a single targets.json file — add, remove, or modify entries as needed.

System Architecture

Agent Team

Agent Role Analytical Philosophy MCP / Capabilities
@Livermore Market Analyst Trend-following, key-point trading, volume-price confirmation Custom MCP servers + Web Search + Financial market analysis
@Buffett Value Investor Fundamental analysis, margin of safety, long-term holding China Stock MCP + Market Data Fetcher + Research reports
@CathieWood Innovation Investor Disruptive technology, convergence across sectors, 5-year horizon FinanceKit + Market Data Fetcher + Research reports
@RayDalio Macro Cycle Analyst Debt cycle, four-environment framework, liquidity-first AKShare + Market Data Fetcher + Research reports
@SOLO Agent Coordinator Merges outputs from all agents, saves reports, runs validation File system + Command execution

MCP Servers

MCP Server Function Connected Agents
market-data-fetcher Fetches index, stock, and ETF data (Longport / AKShare / Yahoo Finance) Livermore, Buffett, CathieWood, RayDalio
financekit Stock quotes, company info, technical analysis, risk metrics, options chain CathieWood
china-stock-mcp A-share financial data, balance sheets, income statements, cash flows, shareholder info Buffett
mcp-aktools AKShare-based tools: stock indicators, sector fund flows, trading suggestions Livermore
financemcp-dcths Tonghuashun (THS) and Dongfang Caifu (DC) sector indices, members, and daily data Livermore
akshare-one-mcp Historical data, real-time data, financial metrics (HTTP-based) Livermore, Buffett, RayDalio
Time Current time and timezone conversion All agents
FaXianBaoGao (发现报告) Research report search and content retrieval All agents
tushareMcp TuShare financial data API (HTTP-based) Buffett

Workflow

┌──────────────────────────────────────────────────────────────────┐
│                        User Instruction                          │
│  (Sent to SOLO Agent in TRAE CN SOLO mode)                       │
└──────────────────────────┬───────────────────────────────────────┘
                           │
                           ▼
┌──────────────────────────────────────────────────────────────────┐
│  Step 1: Learn from Chat History                                 │
│  All agents study past analysis patterns in chatHistory/          │
└──────────────────────────┬───────────────────────────────────────┘
                           │
                           ▼
┌──────────────────────────────────────────────────────────────────┐
│  Step 2: Macro Framework First (RayDalio)                        │
│  RayDalio diagnoses the economic machine state, determines       │
│  short-term debt cycle position, and evaluates four-environment  │
│  probabilities as macro top-level constraint                     │
└──────────────────────────┬───────────────────────────────────────┘
                           │
                           ▼
┌──────────────────────────────────────────────────────────────────┐
│  Step 3: Gather Real-Time Data                                   │
│  Each agent uses MCP tools and WebSearch to collect latest data   │
│  ⚠️ Each agent MUST return the latest price for EVERY target     │
└──────────────────────────┬───────────────────────────────────────┘
                           │
                           ▼
┌──────────────────────────────────────────────────────────────────┐
│  Step 4: Build Reference Prices                                  │
│  SOLO Agent collects prices from all 4 agents, cross-validates,  │
│  and saves to reference_prices.json                              │
│  Missing prices MUST be supplemented via WebSearch               │
└──────────────────────────┬───────────────────────────────────────┘
                           │
                           ▼
┌──────────────────────────────────────────────────────────────────┐
│  Step 5: Multi-Agent Deliberation                                │
│  All 5 agents discuss and produce a consolidated market forecast  │
│  with probability-weighted scenarios for each target              │
│  (All prices MUST be sourced from reference_prices.json)          │
└──────────────────────────┬───────────────────────────────────────┘
                           │
                           ▼
┌──────────────────────────────────────────────────────────────────┐
│  Step 6: Merge & Save Report                                     │
│  SOLO Agent merges all outputs into a Markdown report             │
│  saved to chatHistory/                                           │
└──────────────────────────┬───────────────────────────────────────┘
                           │
                           ▼
┌──────────────────────────────────────────────────────────────────┐
│  Step 7: Validate Prices                                         │
│  SOLO Agent runs price_checker.py                                │
│  If exit code = 1 → loop back to Step 4 to fix prices           │
│  If exit code = 0 → continue                                     │
└──────────────────────────┬───────────────────────────────────────┘
                           │
                           ▼
┌──────────────────────────────────────────────────────────────────┐
│  Step 8: Validate Coverage                                       │
│  SOLO Agent runs targets_validator.py                            │
│  If exit code = 1 → loop back to Step 1                         │
│  If exit code = 0 → done ✓                                       │
└──────────────────────────────────────────────────────────────────┘

Project Structure

MarketForecastingAgents
├── .trae/
│   ├── mcp.json.example       # MCP server configuration template
│   └── rules/
│       ├── toolcallingrules.md # Tool calling rules (English)
│       └── 工具调用规则.md      # Tool calling rules (Chinese)
├── agents_info/
│   ├── Livermore_info.md      # Livermore agent prompt & instructions
│   ├── Buffet_info.md         # Buffett agent prompt & instructions
│   ├── CathieWood_info.md     # Cathie Wood agent prompt & instructions
│   └── RayDalio_info.md       # Ray Dalio agent prompt & instructions
├── chatHistory/               # Generated research reports & chat logs
│   ├── 全市场走势预判_20260513.md  # Full market forecast report
│   ├── 港股走势预判_20260513.md    # HK market forecast report
│   └── ...
├── targets.json               # Target configuration (indices, stocks, ETFs)
├── targets_validator.py       # Report coverage validation script
├── price_checker.py           # Report price validation script
├── reference_prices.json      # Reference prices for price validation (generated by SOLO Agent)
├── .gitignore
├── README.md                  # This file (English)
└── README_cn.md               # Chinese documentation

Target Configuration (targets.json)

All analysis targets are managed through targets.json in the project root directory, covering A-shares, Hong Kong stocks, and US stocks. The file supports flexible configuration of indices, individual stocks, and ETFs. Users can directly edit this file to add or remove targets.

Structure

{
  "a_shares": {
    "index_major": [{"name": "...", "code": "000001.SH"}],
    "sse_stocks":   [{"name": "...", "code": "688981.SH"}],
    "sse_etf":      [{"name": "...", "code": "513310.SH"}],
    "szse_stocks":  [{"name": "...", "code": "000063.SZ"}],
    "szse_etf":     [{"name": "...", "code": ""}]
  },
  "hk_shares": {
    "index_major":  [{"name": "...", "code": "800000.HK"}],
    "hkex_stocks":  [{"name": "...", "code": "00700.HK"}],
    "hkex_etf":     [{"name": "...", "code": "02800.HK"}]
  },
  "us_shares": {
    "index_major":  [{"name": "...", "code": ".NDX"}],
    "stocks":       [{"name": "...", "code": "NVDA"}],
    "adr":          [{"name": "...", "code": "BABA"}],
    "etf":          [{"name": "...", "code": "QQQ"}]
  }
}

Current Preset Targets

  • A-Share Indices: SSE Composite (000001.SH), SZSE Component (399001.SZ)
  • A-Share Stocks: SMIC (688981.SH), ZTE (000063.SZ)
  • A-Share ETFs: China-Korea Semiconductor ETF (513310.SH)
  • HK Indices: Hang Seng Index (800000.HK), Hang Seng Tech Index (800700.HK)
  • HK Stocks: Tencent, Alibaba, Xiaomi, Kuaishou, JD.com, Meituan, SMIC, Pop Mart, CATL, BYD, HKEX, AIA, China Mobile, NetEase, Baidu, Li Auto, XPeng, Anta, Horizon Robotics, and 20+ more
  • HK ETFs: Tracker Fund (02800.HK), CSOP Hang Seng Tech (03033.HK)
  • US Indices: NASDAQ 100 (.NDX), S&P 500 (.SPX)
  • US Stocks: NVIDIA (NVDA)
  • US ADRs: Alibaba (BABA)
  • US ETFs: QQQ, SPY

Note: Entries with empty name and code fields are placeholders — fill them in or remove them as needed.

Report Validation

Coverage Validation (targets_validator.py)

The targets_validator.py script verifies that a generated research report fully covers all targets defined in targets.json.

Usage

python targets_validator.py <report.md> [--targets <targets.json>]

What It Does

  1. Extracts all valid targets from targets.json (entries where both name and code are non-empty)
  2. Searches the report markdown for each target's name or code
  3. Outputs: missing targets list, coverage statistics

Exit Codes

Code Meaning
0 All targets covered, no omissions
1 Missing or extra targets detected

Price Validation (price_checker.py)

The price_checker.py script verifies that every target's current price in the report is present and reasonable, preventing price fabrication or estimation errors.

Usage

python price_checker.py <report.md> [--targets <targets.json>] [--ref <reference_prices.json>] [--deviation <percentage>]

What It Does

  1. Extracts all valid targets from targets.json
  2. Extracts each target's "当前价格" (current price) field from the report markdown
  3. Checks for four types of issues:
    • Missing prices — No "当前价格" field found for a target in the report
    • Zero prices — Price is 0 or negative
    • Sanity issues — Price is below the minimum reasonable threshold for the market
    • Deviation issues — Price deviates from reference_prices.json by more than the threshold (default 50%)

Exit Codes

Code Meaning
0 All target prices present and valid
1 Missing or abnormal prices detected

Getting Started

Prerequisites

  • TRAE CN IDE with SOLO mode support
  • Python 3.10+ (for running targets_validator.py and price_checker.py)
  • uv (for MCP server installation via uvx)
  • Node.js (for MCP server installation via npx)
  • Optional: Longport account for real-time HK/US market data

Setup

  1. Configure MCP Servers: Copy the example configuration and fill in your credentials: 
    cp .trae/mcp.json.example .trae/mcp.json
    Edit .trae/mcp.json and replace placeholder values:
    • LONGPORT_APP_KEY, LONGPORT_APP_SECRET, LONGPORT_ACCESS_TOKEN — your Longport API credentials
    • TARGETS_JSON_PATH — absolute path to your targets.json
    • tushareMcp URL — replace your_tushare_token_here with your TuShare token
  2. Configure Targets: Edit targets.json to add, remove, or modify the indices, stocks, and ETFs you want to analyze.
  3. Set Up Agents: Import the four agents into TRAE CN using the links in agents_info/:
    • Livermore: See agents_info/Livermore_info.md for the prompt
    • Buffett: See agents_info/Buffet_info.md for the prompt
    • Cathie Wood: See agents_info/CathieWood_info.md for the prompt
    • Ray Dalio: See agents_info/RayDalio_info.md for the prompt

Generating a Research Report

After configuration, switch to SOLO mode in TRAE CN and send the following instruction to the SOLO Agent:

1. You and @Livermore @Buffett @CathieWood @RayDalio study the financial market analysis logic and response framework in this directory (/path/to/MarketForecastingAgents/chatHistory/).
2. First, @RayDalio completes the macro framework analysis: diagnose the economic machine state, determine the short-term debt cycle position, evaluate four-environment probabilities (growth↑+inflation↑, growth↑+inflation↓, growth↓+inflation↑, growth↓+inflation↓), providing the macro top-level constraint for all subsequent analysis.
3. All 5 agents (@Livermore @Buffett @CathieWood @RayDalio @SOLO Agent) use MCP tools or WebSearch to gather as much latest data and information as possible. ⚠️ Each agent MUST return the latest price for EVERY target — missing any target's price data is a failure condition.
4. You (SOLO Agent) collect the latest prices returned by the four investment agents, cross-validate them, and save to reference_prices.json (format: {"000001.SH": 4152, "00700.HK": 460, ...}). If any target's price is missing from all four agents, you MUST search for it via WebSearch before proceeding.
5. All 5 agents (@Livermore @Buffett @CathieWood @RayDalio @SOLO Agent) discuss and produce a forecast for the next 6 months for all targets in this file (/path/to/MarketForecastingAgents/targets.json), assigning probability values and price ranges for each of the following scenarios with reasoning: 1. Range-bound with bullish bias; 2. Range-bound with bearish bias; 3. Volatile uptrend; 4. Volatile downtrend; 5. Direct uptrend; 6. Direct downtrend. ⚠️ All prices in the report MUST be sourced from reference_prices.json — never estimate or fill in from memory.
6. You comprehensively analyze the outputs from all 5 agents (@Livermore @Buffett @CathieWood @RayDalio @SOLO Agent), and save the comprehensive analysis results as a Markdown document in this directory (/path/to/MarketForecastingAgents/chatHistory/).
7. You perform price validation: run this Python file (/path/to/MarketForecastingAgents/price_checker.py); if the exit code is 1, loop back to step 4 to fix price issues until the exit code is 0.
8. Finally, you perform target count validation: run this Python file (/path/to/MarketForecastingAgents/targets_validator.py); if the exit code is 1, repeat steps 1-8 until the exit code is 0.

Important: Replace /path/to/MarketForecastingAgents/ with the actual absolute path on your system.

Forecast Scenarios

Each target in the generated report is evaluated against six probability-weighted scenarios:

# Scenario Description
1 Range-bound with bullish bias Price oscillates within a range but leans upward
2 Range-bound with bearish bias Price oscillates within a range but leans downward
3 Volatile uptrend Price trends upward with significant oscillations
4 Volatile downtrend Price trends downward with significant oscillations
5 Direct uptrend Price moves steadily upward without significant pullbacks
6 Direct downtrend Price moves steadily downward without significant rebounds

Disclaimer

The research reports generated by this project are for reference only and do not constitute any investment advice. Financial markets carry inherent risks — invest with caution and always conduct your own due diligence.

Copyright

© 2025–2026 MarketForecastingAgents Contributors. All rights reserved.

License

This project is licensed under the GNU Affero General Public License v3.0 (AGPL-3.0). See the LICENSE file at the repository root for full terms.

About

基于多智能体协作的AI市场研报自动生成系统，覆盖A股/港股/美股 | AI-powered multi-agent market research report auto-generation system covering A-shares, HK & US stocks

Topics

ai mcp multi-agent investment market-analysis a-shares stock-forecast llm us-stocks hong-kong-stocks

Resources

Readme

License

AGPL-3.0 license
Activity

Stars

0 stars

Watchers

0 watching

Forks

0 forks
Report repository

Releases

No releases published

Packages

 
 
 

Contributors

Languages