StrategyIdeationEngine (StrategyAgent)

Institutional-Grade AI Agent for Quantitative Strategy Research

The StrategyIdeationEngine is a sophisticated, self-correcting AI agent designed to generate, critique, and empirically validate mid-term (days to weeks) trading hypotheses. Unlike standard LLM "narrative" generators, this engine uses a Three-Tier Intelligence Architecture to force every idea through a rigorous internal "Red Team" and a high-fidelity historical backtest.

---

🚀 Core Architecture: The Tiered Pipeline

The engine operates on a multi-agent orchestration designed to reduce hallucinations and maximize alpha quality:

1. Phase 1: Event Triggering: Manual ticker research or an automated scanner bot.
2. Phase 2: Ground Truth Ingestion: Unified data gathering from yfinance, Alpha Vantage, and FRED.
3. Phase 3: The Scout (Tier 1): Powered by fast models (e.g., Llama-4 17B). Condenses thousands of tokens from ArXiv papers, Tavily news, and SEC filings into a distilled Numerical Fact Sheet.
4. Phase 4: The Strategist (Tier 2): Powered by high-reasoning models (e.g., Llama-3.3 70B).
   • Thesis Architect: Develops a multi-leg economic rationale with citations.
   • Trade Architect: Translates the thesis into concrete math rules (using 130+ indicators via pandas_ta).
5. Phase 5: Double Validation & Self-Correction:
   • Adversarial Critique (Tier 3): An internal Red Team LLM analyzes the strategy for logical flaws and horizon-specific risks.
   • High-Fidelity Backtest: A quantitative module that simulates the entire strategy portfolio as a co-integrated unit, including transaction costs and slippage.
   • Sector-Aware Benchmarking: Automatically identifies the relevant sector ETF (e.g., XLK, XLE) to isolate true Alpha from Sector Beta.
   • Refinement Loop: If scores are low (< 0.7), the engine passes a Self-Correction Plan back to the Strategist for up to 3 polish cycles.
6. Phase 6: Professional Export: Generates detailed Markdown reports.

---

🚦 Getting Started

1. Prerequisites

• Python 3.10+
• API Keys for your chosen LLM providers: Groq, OpenAI, Anthropic, or Google Gemini.
• Optional Data APIs: Tavily (Search), FRED (Macro), Alpha Vantage (Technicals).

2. Configuration

Create a .env file in the root directory (see .env.example). You can now mix and match models from different providers for each tier:

# API Keys (Set the ones you intend to use)
GROQ_API_KEY=your_groq_key
OPENAI_API_KEY=your_openai_key
ANTHROPIC_API_KEY=your_anthropic_key
GEMINI_API_KEY=your_gemini_key

# Model Tiering (Supports prefixes: groq/, openai/, anthropic/, gemini/)
SCOUT_MODEL=groq/llama-3.1-8b-instant
STRATEGIST_MODEL=openai/gpt-4o
CRITIC_MODEL=anthropic/claude-3-5-sonnet

# Data APIs
TAVILY_API_KEY=your_key
FRED_API_KEY=your_key
ALPHA_VANTAGE_API_KEY=your_key

3. Usage

Manual Run:

python main.py run --ticker NVDA --reason "Semiconductor supply chain shift"

Watcher Mode:

python main.py watch --interval 3600

View History:

python main.py history

---

🛠 Operational Robustness

• Dynamic Context Scaling: Uses a custom TokenBudgeter with tiktoken to maximize research signal while staying within model TPM limits.
• Caching Layer: S&P 500 tickers, FRED Macro data, and technical indicators are cached in the .cache/ directory.
• DSPy Optimization: Refactored into a unified dspy.Module for programmatic prompt optimization and self-correction.
• Sector-Aware Alpha: Automatically maps tickers to their respective Sector ETFs to calculate true Alpha.

🧪 Testing

# Run all tests
python -m pytest tests

🛡 Security & Risk Disclaimer

This software is a research tool and not financial advice. Trading involves significant risk of loss. Backtest results are historical and do not guarantee future performance.

---

Developed for Quantitative Strategy Research v2.0