Advanced Portfolio Hedging & Risk Analytics

📌 Project Overview

This project implements a sophisticated risk management system designed for Ultra High Net Worth (UHNW) clients holding concentrated single-stock positions. The core constraint is tax efficiency: reducing portfolio risk without triggering capital gains taxes by selling the underlying asset.

We compare two distinct hedging approaches:

1. Quantitative Factor Model: A traditional Barra-style risk model using Bloomberg factors (Size, Value, Momentum) and robust regression.
2. NLP Semantic Model: A novel approach using Large Language Model (LLM) embeddings (Nomic v1.5) to identify "fundamental peers" based on semantic business similarity.

🚀 Key Results

The backtest results (Notebook 03) reveal that NLP-based hedging outperforms traditional factor models for idiosyncratic companies where sector labels are insufficient.

Strategy	Win Case Example	Rationale
NLP Hedge	Flextronics (FLEX)	+402 bps risk reduction. NLP correctly identified niche electronics manufacturing peers that generic "Tech" factors missed.
NLP Hedge	Mosaic (MOS)	+392 bps risk reduction. Semantic search captured the specific fertilizer/commodity risk better than broad "Materials" sector factors.
Factor Hedge	Apple (AAPL)	-200 bps. For mega-cap stocks driven by broad market flows, the systematic factor model proved superior to semantic matching.

📂 Repository Structure

.
├── data/                   # Raw and Processed data (Bloomberg, Wikipedia, Embeddings)
├── notebooks/              # Jupyter Notebooks (Sequential Logic)
│   ├── 00_exploratory_data_analysis.ipynb   # Data Cleaning & Veralto/UMB Fixes
│   ├── 01_factor_model_construction.ipynb   # Huber Robust Regression & Factor Returns
│   ├── 02_nlp_embedding_generation.ipynb    # Nomic v1.5 Embeddings & Context-Aware Chunking
│   ├── 03_hedging_strategy_comparison.ipynb # The Backtest: Factor Optimization vs. NLP
│   └── 04_ai_revolution_clustering.ipynb    # Extra Credit: Unsupervised AI Clustering
├── scripts/                # Production scripts for batch jobs
├── src/                    # Source code package (adv_hedging)
│   ├── hedging/            # Optimization & Metrics
│   ├── nlp/                # Text Processing & Embeddings
│   └── risk_model/         # Factor Engine
├── environment.yml         # Conda environment definition
└── pyproject.toml          # Python dependencies

🛠 Installation & Setup

This project uses a custom Conda environment (hedging_clean) with Python 3.10.

1. Clone the Repository:

   git clone [https://github.com/your-username/advanced-portfolio-hedging.git](https://github.com/your-username/advanced-portfolio-hedging.git)

cd advanced-portfolio-hedging ```

2. Create Environment:

   conda env create -f environment.yml
   conda activate hedging_clean

3. Install Local Package:

   pip install -e .

🧠 Methodology Details

1. Factor Risk Model

   • Data: 7 Bloomberg Risk Factors (Size, Value, Momentum, Volatility, Profitability, Leverage, Trading Activity).

   • Estimation: Uses Huber Robust Regression (epsilon=1.35) to estimate daily factor returns, minimizing the impact of outliers (meme stocks).

   • Covariance: Factor covariance matrix estimated on a 2-year rolling window.

2. NLP Engine

   • Model: nomic-ai/nomic-embed-text-v1.5 (Matryoshka embeddings).

   • Innovation: Implements Context-Aware Chunking. Every text chunk includes the company metadata header ("Title: Apple Inc...") to prevent context loss in long documents.

   • Evaluation: Validated using Silhouette Scores on GICS sectors, outperforming standard MPNet and BGE models.

3. Hedging Optimization (Part 3)

   • Objective: Minimize Active Risk (Tracking Error) against the target stock.

   • Constraints:

     • Max 10 positions (Cardinality constraint for operational simplicity).
     • Max weight 25% per position.
     • Hedge Ratio: 100% (Dollar Neutral).

4. AI Revolution Clustering (Extra)

   • Goal: Challenge expert "Maker vs. User" labels using unsupervised learning.
   • Technique: UMAP dimensionality reduction + HDBScan density clustering.
   • Insight: The model identified "Hybrid" clusters (e.g., Cloud Hyperscalers like AMZN/GOOGL) that act as both Makers and Users, defying binary classification.

📊 Usage

To reproduce the full analysis, run the notebooks in order:

1. 00_exploratory_data_analysis.ipynb: Verifies data integrity.

2. 01_factor_model_construction.ipynb: Builds the risk model.

3. 02_nlp_embedding_generation.ipynb: Generates Nomic embeddings (requires GPU/MPS).

4. 03_hedging_strategy_comparison.ipynb: Runs the 50-stock backtest loop.

Alternatively, use the command-line interface:

python scripts/run_hedge_backtest.py

-- Last modified Dec 20, 2025.