charieswei
/
TradingAgents
mirror of https://github.com/TauricResearch/TradingAgents.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity
TradingAgents/CLAUDE.md

5.2 KiB
Raw Blame History

TradingAgents Framework - Project Knowledge

Project Overview

Multi-agent LLM trading framework using LangGraph for financial analysis and decision making.

Development Environment

Conda Environment: tradingagents

Before starting any development work, activate the conda environment:

conda activate tradingagents

Architecture

  • Agent Factory Pattern: create_X(llm) → closure pattern
  • 3-Tier LLM System:
    • Quick thinking (fast responses)
    • Mid thinking (balanced analysis)
    • Deep thinking (complex reasoning)
  • Data Vendor Routing: yfinance (primary), Alpha Vantage (fallback)
  • Graph-Based Workflows: LangGraph for agent coordination

Key Directories

  • tradingagents/agents/ - Agent implementations
  • tradingagents/graph/ - Workflow graphs and setup
  • tradingagents/dataflows/ - Data access layer
  • cli/ - Command-line interface

Agent Flow (Existing Trading Analysis)

  1. Analysts (parallel): Fundamentals, Market, News, Social Media
  2. Bull/Bear Debate
  3. Research Manager
  4. Trader
  5. Risk Debate
  6. Risk Judge

Scanner Flow (New Market-Wide Analysis)

START ──┬── Geopolitical Scanner (quick_think) ──┐
        ├── Market Movers Scanner (quick_think) ──┼── Industry Deep Dive (mid_think) ── Macro Synthesis (deep_think) ── END
        └── Sector Scanner (quick_think) ─────────┘
  • Phase 1: Parallel execution of 3 scanners
  • Phase 2: Industry Deep Dive cross-references all outputs
  • Phase 3: Macro Synthesis produces top-10 watchlist

Data Vendors

  • yfinance (primary, free): Screener(), Sector(), Industry(), index tickers
  • Alpha Vantage (alternative, API key required): TOP_GAINERS_LOSERS endpoint only (fallback for market movers)

LLM Providers

OpenAI, Anthropic, Google, xAI, OpenRouter, Ollama

CLI Entry Point

cli/main.py with Typer:

  • analyze (per-ticker analysis)
  • scan (new, market-wide scan)

Configuration

tradingagents/default_config.py:

  • LLM tiers configuration
  • Vendor routing
  • Debate rounds settings
  • All values overridable via TRADINGAGENTS_<KEY> env vars (see .env.example)

Patterns to Follow

  • Agent creation (trading): tradingagents/agents/analysts/news_analyst.py
  • Agent creation (scanner): tradingagents/agents/scanners/geopolitical_scanner.py
  • Tools: tradingagents/agents/utils/news_data_tools.py
  • Scanner tools: tradingagents/agents/utils/scanner_tools.py
  • Graph setup (trading): tradingagents/graph/setup.py
  • Graph setup (scanner): tradingagents/graph/scanner_setup.py
  • Inline tool loop: tradingagents/agents/utils/tool_runner.py

Critical Patterns (from past mistakes — see MISTAKES.md)

  • Tool execution: Trading graph uses ToolNode in graph. Scanner agents use run_tool_loop() inline. If bind_tools() is used, there MUST be a tool execution path.
  • yfinance DataFrames: top_companies has ticker as INDEX, not column. Always check .index and .columns.
  • yfinance Sector/Industry: Sector.overview has NO performance data. Use ETF proxies for performance.
  • Vendor fallback: Functions inside route_to_vendor must RAISE on failure, not embed errors in return values. Catch (AlphaVantageError, ConnectionError, TimeoutError), not just RateLimitError.
  • LangGraph parallel writes: Any state field written by parallel nodes MUST have a reducer (Annotated[str, reducer_fn]).
  • Ollama remote host: Never hardcode localhost:11434. Use configured base_url.
  • .env loading: load_dotenv() runs at module level in default_config.py — import-order-independent. Check actual env var values when debugging auth.
  • Rate limiter locks: Never hold a lock during sleep() or IO. Release, sleep, re-acquire.
  • Config fallback keys: llm_provider and backend_url must always exist at top level — scanner_graph.py and trading_graph.py use them as fallbacks.

Project Tracking Files

  • DECISIONS.md — Architecture decision records (vendor strategy, LLM setup, tool execution, env overrides)
  • PROGRESS.md — Feature progress, what works, TODOs
  • MISTAKES.md — Past bugs and lessons learned (10 documented mistakes)

LLM Configuration

Per-tier provider overrides in tradingagents/default_config.py:

  • Each tier (quick_think, mid_think, deep_think) can have its own _llm_provider and _backend_url
  • Falls back to top-level llm_provider and backend_url when per-tier values are None
  • All config values overridable via TRADINGAGENTS_<KEY> env vars
  • Keys for LLM providers: .env file (e.g., OPENROUTER_API_KEY, ALPHA_VANTAGE_API_KEY)

Env Var Override Convention

# Pattern: TRADINGAGENTS_<UPPERCASE_KEY>=value
TRADINGAGENTS_LLM_PROVIDER=openrouter
TRADINGAGENTS_DEEP_THINK_LLM=deepseek/deepseek-r1-0528
TRADINGAGENTS_MAX_DEBATE_ROUNDS=3
TRADINGAGENTS_VENDOR_SCANNER_DATA=alpha_vantage

Empty or unset vars preserve the hardcoded default. None-default fields (like mid_think_llm) stay None when unset, preserving fallback semantics.

Running the Scanner

conda activate tradingagents
python -m cli.main scan --date 2026-03-17

Running Tests

conda activate tradingagents
pytest tests/ -v