trading-agents.skill

Orchestrate a swarm of specialized Claude subagents that simulate a professional trading firm to analyze stocks and produce trading decisions. Based on the TradingAgents framework (arXiv 2412.20138), this skill deploys analyst agents (fundamental, technical, sentiment, news), bull/bear research debaters, a trader, a risk manager, and a portfolio manager — all collaborating to produce a comprehensive trading recommendation. Use this skill whenever the user asks about stock analysis, trading decisions, market research for specific tickers, investment recommendations, portfolio decisions, or wants a multi-perspective analysis of any publicly traded security. Also trigger when the user mentions "trading agents", "multi-agent trading", "stock swarm", or wants an AI-driven trading desk analysis.

TradingAgents: Multi-Agent Trading Analysis Skill

This skill orchestrates a swarm of Claude subagents that mirror the structure of a real trading firm. Each agent has a distinct role, specific tools, and a clear mandate. The agents collaborate through structured reports, adversarial debate, and sequential review — producing a final trading recommendation that reflects diverse analytical perspectives.

Prerequisites

This skill requires Python, pip, and uv to be installed on the system. Before running any analysis, set up the Python environment:

1. Install uv (if not already installed):

   bash

   pip install -U uv
   

2. Sync the project dependencies from the skill directory:

   bash

   cd {SKILL_PATH} && uv sync
   

   This installs all required packages (yfinance, akshare, etc.) into a managed virtual environment based on pyproject.toml. You only need to do this once, or when dependencies change.

All Python scripts in this skill must be executed with uv run to ensure they use the correct environment. For example: uv run scripts/fetch_market_data.py NVDA

Architecture Overview

The system follows a five-stage pipeline inspired by the TradingAgents paper:

text

Stage 1: Analysis (parallel)     → 4 analyst agents gather data simultaneously
Stage 2: Research (debate)       → Bull and bear researchers debate the findings
Stage 3: Trading decision        → Trader synthesizes everything into a signal
Stage 4: Risk review             → Risk manager evaluates the proposed trade
Stage 5: Final approval          → Portfolio manager makes the go/no-go call

How to Use This Skill

When the user asks for a stock analysis or trading decision, follow these steps:

Step 0: Parse the Request

Extract from the user's message:

• Ticker(s): The stock symbol(s) to analyze (e.g., NVDA, AAPL)
• Date context: Whether they want current analysis or historical (default: today)
• Debate rounds: If specified, how many bull/bear rounds (default: 1)
• Focus areas: Any specific concerns (e.g., "worried about earnings", "considering for long-term hold")

If the ticker is ambiguous or missing, ask the user to clarify.

Step 1: Launch Analyst Agents (Parallel)

Spawn four analyst subagents simultaneously using the Agent tool. Each agent gets its own prompt from the agents/ directory. Pass each agent the ticker, date, and any user context.

Read the agent prompts before spawning:

• agents/fundamental_analyst.md — Analyzes financial health, valuation, earnings
• agents/technical_analyst.md — Analyzes price patterns, indicators, chart signals
• agents/sentiment_analyst.md — Gauges market mood from social media and forums
• agents/news_analyst.md — Evaluates recent news and macro events

Each analyst should save their report to a working directory. The prompts instruct them on format and what tools to use (web search, yfinance via the scripts/fetch_market_data.py script, etc.).

Important: Launch all four in a single message to maximize parallelism. Don't wait for one to finish before starting the next.

Step 2: Collect Analyst Reports

Once all four analysts complete, read their reports. Compile them into a single briefing document that will feed into the research phase.

Step 3: Bull/Bear Research Debate

Spawn the debate in rounds. For each round:

1. Spawn bull researcher and bear researcher simultaneously (read agents/bull_researcher.md and agents/bear_researcher.md). Give them all four analyst reports plus any previous debate history.
2. The bull researcher argues for the investment opportunity; the bear researcher argues against.
3. After each round, both researchers can read each other's previous arguments and respond.

Default is 1 round. For deeper analysis, the user can request 2-3 rounds. More rounds means more thorough analysis but also more time and tokens.

After the debate, spawn the research manager (read agents/research_manager.md) to synthesize the debate into a balanced research summary.

Step 3.5: Compile Debate Record

After the research manager produces the summary, compile a standalone debate record document (debate_record.md) that presents the full bull/bear debate process in a clear, readable format. This document should include:

1. 辩论背景: The ticker, date, and number of debate rounds
2. 第 N 轮辩论: For each round, show the bull case followed by the bear case, clearly labeled
3. 研究经理总结: The research manager's balanced synthesis at the end

This gives the reader a single document to see the entire adversarial debate process, rather than having to read multiple separate files. The individual bull_case.md, bear_case.md, and research_summary.md files should still be saved separately as well.

Step 4: Trading Decision

Spawn the trader agent (read agents/trader.md). Give it:

• All four analyst reports
• The research debate summary
• The user's original context/constraints

The trader produces a concrete recommendation: BUY, SELL, or HOLD, with position sizing guidance, entry/exit points, and confidence level.

Step 5: Risk Management Review

Spawn the risk manager (read agents/risk_manager.md). Give it:

• The trader's recommendation
• All analyst reports
• Current portfolio context if available

The risk manager evaluates: position size appropriateness, portfolio concentration risk, volatility assessment, downside scenarios, and liquidity concerns.

Step 6: Portfolio Manager Approval

Spawn the portfolio manager (read agents/portfolio_manager.md). Give it everything:

• Analyst reports, debate summary, trader recommendation, risk assessment

The portfolio manager makes the final call: APPROVE, REJECT, or MODIFY the recommendation, with reasoning.

Step 7: Compile Final Output

Produce two outputs:

1. A comprehensive report file (Markdown) saved to the workspace, containing:

   • Executive summary with the final decision
   • Each analyst's key findings (condensed)
   • Bull/bear debate highlights
   • Trader's recommendation details
   • Risk assessment summary
   • Portfolio manager's final decision and reasoning
   • Disclaimer that this is AI-generated analysis, not financial advice

   Additionally, a debate_record.md file that compiles the entire bull/bear debate process into a single readable document (see Step 3.5).

2. A conversational summary in the chat, covering:

   • The final recommendation (BUY/SELL/HOLD)
   • Top 3 reasons for the decision
   • Key risk factors
   • Confidence level
   • Link to the full report

Configuration Defaults

• Debate rounds: 1 (configurable by user, max 5)
• Data sources: Web search + yfinance (scripts/fetch_market_data.py) + optional APIs
• Output: Both report file + conversational summary

Helper Scripts

• scripts/fetch_market_data.py — Fetches price history, financial statements, and key metrics via yfinance
• scripts/technical_indicators.py — Computes common technical indicators (RSI, MACD, Bollinger Bands, moving averages)

These scripts are used by the analyst agents. Run them from the skill directory using uv run:

bash

uv run scripts/fetch_market_data.py <TICKER> [-o OUTPUT_DIR]
uv run scripts/technical_indicators.py <TICKER> [-o OUTPUT_DIR]

Source Citation & Data Quality Standards

All analyst reports must meet these standards:

• Primary sources first: Financial data should come from first-hand, authoritative sources: company investor relations pages, stock exchange filings (SEC EDGAR, HKEX, SSE/SZSE), official earnings releases, and annual/quarterly reports. Third-party aggregators (Yahoo Finance, Bloomberg, etc.) are acceptable as supplementary sources but should be labeled as such.

• Every key data point must cite its source with a clickable URL, the reporting period (e.g., "FY2025", "Q1 2026"), and the currency/unit (e.g., "人民币/百万元", "USD millions").

• News must come from authoritative media, prioritized in this order: official company announcements > tier-1 financial media (Reuters, Bloomberg, FT, WSJ, 财新, 第一财经) > regional authoritative media > industry publications. Each news item must include the publication date and a clickable link.

• Sentiment claims must be attributed to specific sources with links, not vague statements like "market sentiment is bullish." For Chinese/HK stocks, 雪球 (xueqiu.com) should be the primary sentiment data source.

• Technical indicators must include plain-language explanations so non-expert readers can understand what each indicator means and why it matters.

• Industry-specific analysis is required: Analysts and risk managers must go beyond generic metrics and cover sector-specific KPIs. For example:

  • Banking: capital adequacy ratio, NPL ratio, provision coverage ratio, NIM
  • E-commerce: GMV, take rate, MAU/DAU, customer acquisition cost
  • Consumer electronics: supply chain, market share trends, component cost analysis
  • Automotive/EV: monthly delivery/sales data (12–36 months), competitive comparison
  • Insurance: NBV, embedded value, combined ratio, solvency ratio
  • Multi-segment companies: break down by segment with relevant industry metrics for each

Important Notes

• Not financial advice: Always include a disclaimer in both the report and the summary. This is an AI research tool for educational purposes.
• Data freshness: yfinance data may have delays. Web search helps get the latest news.
• Cost awareness: Each analysis spawns 8-12+ subagents. For users analyzing multiple tickers, suggest doing them one at a time or warn about the computational cost.
• Error handling: If an analyst agent fails (e.g., can't find data), note the gap in the final report rather than blocking the entire pipeline. The system should be resilient to partial failures.
• No internal paths in reports: Never include internal file system paths (e.g., /sessions/..., /tmp/..., working directory paths) in any report that the reader will see. These are implementation details. Reports should reference other reports by filename only (e.g., "详见 fundamental_analysis.md"), not by absolute path.