QQQ Options Alpha Research

Forecasting QQQ with its Own Options Data: A Ensemble Machine Learning Approach

This repository contains a robust, production-ready trading strategy that uses end-of-day QQQ options data to forecast next-day directional movement and generate daily trading signals with leverage between -1.0x and +1.5x.

🎯 Objective

Design a model that systematically deciphers sentiment, risk appetite, and positioning embedded within the QQQ options market to gain an edge on future price action.

Key Performance Target:

• Calmar Ratio > 2.0 (Risk-adjusted returns)
• Robustness: Strategy stable to ±10% parameter variations
• Leverage Range: -1.0x (full short) to +1.5x (leveraged long)

📊 Results Summary

Test Set Performance (2024-07-26 to 2025-09-17)

Metric	Value
Calmar Ratio	2.14 ✅
Sharpe Ratio	1.92
Total Return	23.5%
Max Drawdown	-13.7%
Win Rate	~55%

Robustness Check: Strategy maintains Calmar > 1.5 across all parameter variations (±10%).

🏗️ Repository Structure

qqq-options-alpha-research/
├── data/
│   ├── options_eod_QQQ.csv          # Raw options data (>5M rows)
│   └── daily_features.parquet        # Preprocessed features (generated)
├── notebooks/
│   ├── eda.ipynb                     # Feature engineering & EDA (NEW - Modular)
│   └── model.ipynb                   # Model training & evaluation (NEW - Modular)
├── src/
│   ├── feature_engineering.py        # Robust feature generation (100+ features)
│   ├── ensemble_model.py             # Ensemble ML model (LightGBM + XGBoost + RF + Ridge)
│   ├── visualization.py              # Plotting utilities (NEW)
│   └── backtesting.py                # Performance metrics & robustness testing (NEW)
├── docs/
│   ├── strategy_logic.md             # Detailed strategy rationale
│   ├── feature_engineering.md        # Feature descriptions
│   └── methodology.md                # Model methodology & assumptions
├── outputs/                          # Generated results & plots
├── environment.yml                   # Conda environment configuration
├── README.md                         # This file
└── .gitignore                        # Git ignore rules

🚀 Quick Start

1. Setup Environment

# Clone repository
git clone https://github.com/SMalaekeh/qqq-options-alpha-research.git
cd qqq-options-alpha-research

# Create conda environment (recommended)
conda env create -f environment.yml
conda activate qqq-options-alpha

# OR install dependencies with pip
pip install pandas numpy scikit-learn lightgbm xgboost matplotlib seaborn jupyter notebook ipykernel pyarrow

2. Run Feature Engineering

# Open and run eda.ipynb
jupyter notebook notebooks/eda.ipynb

This will:

• Load raw options data
• Generate 100+ features including:
  • Volatility surface (IV by moneyness × tenor)
  • Greeks & GEX (Gamma Exposure)
  • Variance Risk Premium (VRP)
  • Put/Call ratios & flow metrics
  • Regime detection (High/Low volatility)
• Save features to data/daily_features.parquet

3. Train & Evaluate Model

# Open and run model.ipynb
jupyter notebook notebooks/model.ipynb

This will:

• Train ensemble model (LightGBM + XGBoost + RF + Ridge)
• Generate trading signals with volatility targeting
• Evaluate performance (Sharpe, Calmar, drawdown)
• Run robustness analysis

📈 Strategy Overview

Core Philosophy

The options market is a sentiment barometer and positioning indicator. Large institutional traders must hedge their positions, creating predictable flows. By analyzing these flows and implied volatility dynamics, we can forecast short-term QQQ movements.

Key Signals

1. Variance Risk Premium (VRP)

   • Spread between implied volatility and realized volatility
   • High VRP → Market overpricing risk → Mean reversion opportunity

2. Gamma Exposure (GEX)

   • Measures dealer hedging needs
   • High GEX → Price suppression (dealers hedging)
   • Low/Negative GEX → Increased volatility

3. Put/Call Ratios

   • Sentiment indicator (fear vs. greed)
   • Elevated put buying → Potential reversal signal

4. Volatility Skew

   • Difference between OTM put and call IVs
   • Steepening skew → Rising hedging demand → Bearish signal

Model Architecture

Why Ensemble Trees over Deep Learning and aggressive hyperparameter optimization using Optuna?

During the R&D phase, I conducted extensive experiments comparing Deep Learning architectures against Tree-based Ensembles.

• Deep Learning Experiments: I implemented Transformer encoders (for attention mechanisms on volatility surfaces) and LSTM networks to capture sequential dependencies.
• The Findings: While DL models achieved high training accuracy, they suffered from significant variance in the test set. Given the noise-to-signal ratio in financial data, the DL models tended to overfit to specific market regimes.
• The Decision: I selected an Ensemble Approach (LightGBM + XGBoost + Random Forest + Ridge).
  • Interpretability: Tree-based models allow for explicit feature importance analysis (Gini/Gain), aligning with the challenge's requirement for a "logical framework."
  • Robustness: Bagging and Boosting proved more stable across changing volatility regimes than neural networks.

Ensemble Approach:

• LightGBM (30%): Fast gradient boosting for feature interactions
• XGBoost (30%): Robust gradient boosting with regularization
• Random Forest (30%): Bagging for stability
• Ridge Regression (10%): Linear anchor to prevent overfitting

Signal Generation:

1. Raw predictions → EMA smoothing (reduce noise)
2. Volatility targeting (scale positions by realized vol)
3. Regime filter (reduce/eliminate positions during crises)

Risk Management:

• Leverage capped at -1.0x to +1.5x
• Kill switch: Go to cash if vol > 50% (crisis mode)
• Reduce size by 50% if vol > 35% (elevated risk)

📚 Documentation

Detailed documentation in /docs/:

• Strategy Logic: Why the model works (financial rationale)
• Feature Engineering: Description of all 100+ features
• Methodology: Model training, validation, and robustness testing

🔬 Robustness & Anti-Overfitting Measures

1. Chronological Splitting: 60/20/20 train/val/test (no look-ahead)
2. Rolling Z-scores: Features converted to relative values (stationarity)
3. Feature Selection: Top 15 features selected via F-statistics
4. Regime Detection: Model aware of market volatility state
5. Outlier Clipping: All features winsorized to prevent extreme values
6. Parameter Stability: Calmar > 1.5 across all tested variations

📊 Key Features Generated

Volatility Surface (20 features)

• IV by moneyness: Deep OTM Put, OTM Put, ATM, OTM Call, Deep OTM Call
• IV by tenor: Weekly (0-10d), Monthly (10-45d), Quarterly (45-90d), Long (90d+)

Greeks & Positioning (15 features)

• GEX (Gamma Exposure): Total, ATM, Call, Put
• Vega exposure by call/put
• Delta exposure & net flow

Sentiment & Flow (10 features)

• Put/Call ratios: Volume, OI, OTM
• Volume by moneyness and tenor
• Dollar volume flows

Momentum & Volatility (8 features)

• Realized volatility: 5d, 10d, 20d, 60d
• Price momentum: 5d, 10d, 20d, 60d

Advanced (20+ features)

• VRP (Variance Risk Premium)
• Volatility skew (put premium over calls)
• Term structure (short-term vs. long-term IV)
• Interaction terms (GEX × momentum, VRP × skew, etc.)

🎓 References & Inspiration

• Variance Risk Premium: Carr & Wu (2009), Bollerslev et al. (2009)
• Gamma Exposure: SqueezeMetrics, SpotGamma research
• Volatility Skew: Bates (1991), Rubinstein (1994)
• Options Market Microstructure: Gârleanu, Pedersen, Poteshman (2009)

🛠️ Technical Details

Dependencies:

• Python 3.8+
• pandas, numpy, scikit-learn
• lightgbm, xgboost
• matplotlib, seaborn

Data Requirements:

• QQQ end-of-day options data (strike, IV, Greeks, volume, OI)
• Minimum 2+ years of history for proper training

📝 License

This project is for educational and research purposes.

🙏 Acknowledgments

Developed as part of the Quanta Options Big Data Challenge. Special thanks to the quantitative finance community for open research on options market microstructure.

---

Disclaimer: This is a research project. Past performance does not guarantee future results. Trade at your own risk.