cointegration-alpha-research

🦅 Beta-Neutral Statistical Arbitrage Engine

---

📊 Executive Summary

This repository contains an institutional-grade Quantitative Research Framework designed to identify, test, and execute Statistical Arbitrage (StatArb) strategies.

Unlike standard correlation-based pairs trading, this engine utilizes Cointegration (Engle-Granger Two-Step Method) to identify stationary time-series relationships. It implements Dynamic Beta Hedging using OLS Regression to neutralize market risk and integrates a Machine Learning (Random Forest) layer to filter signal noise, ensuring trade execution occurs only during high-probability mean-reversion regimes.

Key Edge: Isolates idiosyncratic alpha from broad market volatility (Beta $\approx$ 0), creating a return stream uncorrelated with the S&P 500.

---

🏗️ Technical Architecture

1. Data Pipeline & Scalability

• Ingestion: Fetches high-frequency (Hourly) OHLCV data via yfinance.
• Microstructure Handling: Implements ffill/bfill imputation to handle missing ticks and execution gaps.
• Parallel Computing: Built on Dask to demonstrate horizontal scalability, allowing the engine to process broad-market universes (Russell 3000) without memory bottlenecks.

2. The Alpha Engine (Math & Stats)

• Stationarity Testing: Utilizes the Augmented Dickey-Fuller (ADF) test from statsmodels to validate the cointegration vector ($p < 0.05$).
• Hedge Ratio Optimization: Calculates the rolling Ordinary Least Squares (OLS) beta to dynamically adjust position sizing.
  • Formula: $Spread_t = Price^A_t - (\beta \times Price^B_t)$
• Signal Generation: Computes rolling Z-Scores to normalize spread divergence, entering trades at $\pm 2.0\sigma$ (Bollinger Band logic).

3. Machine Learning Risk Filter

• Model: Random Forest Classifier (scikit-learn).
• Feature Engineering:
  • Spread Volatility (10-period rolling std)
  • Market Velocity (Momentum)
  • Z-Score Extremes
• Function: Acts as a "Regime Filter" to reject mean-reversion signals during momentum breakouts, significantly reducing Max Drawdown.

---

📈 Performance & Tearsheet

The strategy performance is evaluated using institutional risk metrics, accounting for 10bps transaction costs to simulate real-world execution friction.

Metric	Value	Description
Strategy Type	Mean Reversion	High-frequency StatArb
Hedge Ratio	1.60	Dynamic OLS Beta (Long AMD / Short NVDA)
Sharpe Ratio	Dynamic	Risk-adjusted return unit
Max Drawdown	Dynamic	Peak-to-valley loss intensity
Signal Precision	> 60%	ML-filtered entry accuracy

(Note: Detailed performance graphs and equity curves can be found in the outputs/ directory.)

---

📂 Repository Structure

├── 📁 src/                # Source code for the engine
│   ├── data_loader.py     # Dask/Pandas ingestion pipelines
│   ├── statarb_core.py    # ADF tests and Cointegration logic
│   └── ml_filter.py       # Scikit-Learn Random Forest implementation
├── 📁 notebooks/          # Jupyter Notebooks for research & visualization
│   └── Strategy_Walkthrough.ipynb
├── 📁 outputs/            # Generated Tearsheets, Equity Curves, and Z-Score Plots
├── .gitignore             # Financial data exclusion rules
├── requirements.txt       # Production dependencies
└── README.md              # Documentation