AgentQuant: Autonomous Quantitative Research Agent

A fully autonomous AI agent that researches, generates, and validates trading strategies.

🚀 Update (Nov 2025): Now powered by Google Gemini 2.5 Flash. The agent is fully functional and no longer uses random simulation. It actively analyzes market regimes and proposes context-aware strategies.

🎯 What This Project Is

AgentQuant is an AI-powered research platform that automates the quantitative workflow. It replaces the manual work of a junior quant researcher:

1. 1. Market Analysis: Detects regimes (Bull, Bear, Crisis) using VIX and Momentum.
2. 2. Strategy Generation: Uses Gemini 2.5 Flash to propose mathematical strategy parameters optimized for the current regime.
3. 3. Validation: Runs rigorous Walk-Forward Analysis and Ablation Studies to prove strategy robustness.
4. 4. Backtesting: Executes vectorized backtests to verify performance.

🏗️ System Architecture

graph TD
    subgraph "User Interface"
        UI[Streamlit Dashboard]
        Config[config.yaml]
    end

    subgraph "Data Layer"
        Ingest[Data Ingestionyfinance]
        Features[Feature EngineIndicators]
        Regime[Regime DetectionVIX/Momentum]
    end

    subgraph "Agent Core (Gemini 2.5 Flash)"
        Planner[Strategy Planner]
        Context[Market ContextAnalysis]
    end

    subgraph "Execution Layer"
        Strategies[Strategy RegistryMomentum, MeanRev, etc.]
        Backtest[Backtest EngineVectorBT/Pandas]
    end

    subgraph "Validation"
        WalkForward[Walk-ForwardValidation]
        Ablation[AblationStudy]
    end

    UI --> Config
    Config --> Ingest
    Ingest --> Features
    Features --> Regime
    
    Regime --> Context
    Features --> Context
    Context --> Planner
    
    Planner -->|Proposes Params| Strategies
    Strategies --> Backtest
    
    Backtest --> UI
    Backtest --> WalkForward
    Backtest --> Ablation

🧠 The "Brain" (Gemini 2.5 Flash)

The agent uses a sophisticated prompt engineering framework to:

• Analyze technical indicators (RSI, MACD, Volatility).
• Understand market context (e.g., "High Volatility Bear Market").
• Propose specific parameters (e.g., "Use a shorter 20-day lookback for momentum in this volatile regime").

🔬 Scientific Validation

We have implemented rigorous experiments to validate the agent's intelligence:

1. Ablation Study (experiments/ablation_study.py)

• Hypothesis: Does giving the AI "Market Context" improve performance?
• Method: Compare an agent with access to market data vs. a "blind" agent.
• Result: Context-aware agents significantly outperform blind agents in Sharpe Ratio.