🚀 Hull Tactical Market Prediction Engine

An end-to-end ensemble LightGBM system that directly predicts optimal S&P 500 positions under volatility and return penalties for the Hull Tactical Kaggle Competition.

🎯 Business Problem

"Wisdom from most personal finance experts would suggest that it's irresponsible to try and time the market... But in the age of machine learning, is it irresponsible to not try?"

The Hull Tactical competition challenges the Efficient Market Hypothesis by evaluating models on a modified Sharpe ratio. The aim is to build a model that predicts excess returns and includes a betting strategy designed to outperform the S&P 500 while staying within a 120% volatility constraint.

🏗 Solution Architecture

1. Custom Target Engineering

Instead of predicting raw returns (regression), I formulated a Risk-Adjusted Optimal Position Target (classification) for supervised learning. This target incorporates:

• Forward-looking market returns.
• Under-performance and volatility penalties.
• Aligns the model's loss function with the competition's specific Sharpe Ratio objective.

2. Ensemble LightGBM Strategy

• Dual-Window Training: Trained two independent models on different time horizons to capture both long-term regime stability and short-term market shifts.
• Hyperparameter Tuning: Optimized using Optuna to maximize generalization on out-of-sample data.
• Ensemble Logic: Final trading signal is a weighted ensemble of the two models, reducing variance and smoothing signal noise.
• Adaptive Retraining: Implemented a periodic retraining loop to allow the model to adapt to new market regimes during the inference phase.

3. Feature Engineering

Enhanced the provided 98 "black-box" features with domain-specific technical indicators:

• Momentum Indicators: Capturing price velocity and trend strength.
• RSI (Relative Strength Index): Identifying overbought/oversold regimes.
• Lagged Return Patterns: Encoding historical market memory.

🛠 Technical Skills Demonstrated

• Quantitative Risk Modeling: Formulating custom loss functions that penalize volatility (Sharpe-aware learning).
• Ensemble ML Engineering: Designing a multi-model system with time-window diversity to improve robustness.
• Automated Retraining Pipelines: Building a system that updates itself with new data, simulating a live trading environment.
• High-Performance ML: Utilizing LightGBM for efficient handling of high-dimensional feature spaces (9000+ samples trained in <5 mins).

📊 Results

• Status: Inference in progress (Live Testing Period).

📂 Dataset & Structure

Data Source: Hull Tactical Competition

• 98 Features: Market dynamics, Macro-economic indicators, Proprietary signals.
• Target: S&P 500 daily returns (transformed into optimal position targets).

├── notebooks/                  # Feature Engineering & Training Logic
├── src/                        # Inference pipeline scripts
├── requirements.txt            # Python dependencies
└── README.md                   # Documentation