AI-Driven Regime-Aware Pairs Trading in Healthcare

Statistical Arbitrage · Graph Attention Network · Hidden Markov Model · PPO Reinforcement Learning · Python

FE-670 Final Project — MS Financial Engineering, Stevens Institute of Technology

Authors: Swara Dave · Parth Vora · Robin Singh

---

📌 Overview

Classic pairs trading assumes stable relationships between assets — but markets switch between calm and crisis regimes, and naïve Z-score rules often over-trade in noisy environments. This project builds an end-to-end AI-driven, regime-aware pairs trading framework in the U.S. healthcare and biopharma sector combining:

1. Statistical screening — Johansen cointegration tests across 741 pairs in a 39-asset universe
2. Graph Attention Network (GAT) — learns a data-driven similarity score over all assets using network structure
3. Hidden Markov Model (HMM) — detects 3 market regimes (Calm, Normal, Volatile) to condition trade execution
4. PPO Reinforcement Learning agent — executes mean-reversion trades in a custom gym environment with transaction costs

---

📊 Key Results

• Screened 741 pairs across 39 healthcare equities → 36 cointegrated candidates (Johansen p < 0.05)
• GAT Spearman rank correlation between model scores and realized P&L: ρ = 0.5629 (p = 0.0004)
• PPO agent achieved 94–96% win rates across COVID and post-COVID regimes with controlled drawdowns
• Top performing pairs: BNTX–RGEN and AMGN–LLY — high win rates and stable mean-reverting spreads
• GNN Precision-at-K: top 5 and top 10 ranked pairs had substantially higher profitable strategy rates than the full set

---

🗂️ Data

Daily adjusted close prices from Feb 6, 2020 to Dec 5, 2025 for 39 U.S. healthcare and biopharma tickers.

Features engineered per asset:

• Log returns, annualized volatility, rolling Sharpe-like ratio
• Pairwise Johansen cointegration p-values
• Cross-sectional average return, volatility, and correlation (HMM inputs)

---

⚙️ Methodology

Pipeline Overview

Data & EDA → Cointegration Screening → HMM Regime Detection → GAT Pair Scoring → PPO Execution → Evaluation

1. Statistical Screening

• Computed correlation matrix and pairwise Johansen cointegration p-values for all 741 pairs
• Filtered to 36 "strong pairs" at p < 0.05 threshold

2. Hidden Markov Model (HMM)

• 3-state HMM on universe-level factors: cross-sectional average return, volatility, average pairwise correlation
• Regime labels: Calm (low vol, moderate returns), Normal (average vol), Volatile (elevated variance, extreme moves)
• Regime label fed into PPO state vector — same Z-score has different meaning across regimes

3. Graph Attention Network (GAT)

• Nodes: 39 assets with features (return, volatility, Sharpe ratio, sector indicators)
• Edges: cointegrated pairs + highly correlated pairs (weighted by statistical strength)
• Architecture: 2 attention layers → 32-dim hidden space (4 heads) → 128-dim embedding → 16-dim output
• Training: binary cross-entropy loss, Adam optimizer, dropout 0.2
• Output: dot-product similarity score per pair → ranked list of trading candidates

4. PPO Reinforcement Learning Execution

• Custom OpenAI Gym environment per pair
• State vector: spread Z-score, spread velocity, HMM regime, current position, unrealized PnL (normalized)
• Action space: hold / enter long spread / enter short spread / exit
• Reward: PnL change net of transaction costs — penalizes over-trading
• Trained on rolling windows with clipped policy gradients for stable learning

---

📁 Repository Structure

PairsTrading/
├── 1_Data_Ingestion_&_EDA.ipynb                    # Data loading, EDA, correlation & cointegration analysis
├── 2_Hidden_Markov_Model_(Regime_Detection).ipynb  # HMM regime detection
├── 3_DRL_(PPO).ipynb                               # PPO agent training & backtesting
├── 4_GNN.ipynb                                     # Graph Attention Network pair scoring
└── Results/
    ├── cointegration_pvalues.csv                   # Johansen test p-values for all 741 pairs
    ├── correlation_matrix.csv                      # Pairwise correlation matrix for 39 assets
    ├── strong_pairs.csv                            # 36 cointegrated pairs (p < 0.05)
    ├── hmm_regimes.csv                             # Daily HMM regime labels
    ├── gnn_all_pairs_scores.csv                    # GAT similarity scores for all pairs
    ├── all_pairs_results.csv                       # PPO trading results across all 36 pairs
    └── gnn_precision_metrics.csv                   # GNN validation and precision-at-K metrics

🚀 How to Run

1. Clone the repo
2. Install dependencies:

pip install numpy pandas matplotlib seaborn scikit-learn torch torch-geometric stable-baselines3 gymnasium hmmlearn yfinance

3. Run notebooks in order:
   • 1_Data_Ingestion_&_EDA.ipynb — fetches stock data via yfinance, runs EDA and cointegration screening
   • 2_Hidden_Markov_Model_(Regime_Detection).ipynb — fit HMM and label regimes
   • 4_GNN.ipynb — train GAT and score pairs
   • 3_DRL_(PPO).ipynb — train PPO agents and evaluate

Note: Stock data is pulled automatically via yfinance inside the first notebook. No manual data download required. Pre-computed results are available in the Results/ folder.

👤 Authors

Swara Dave — MS Financial Engineering, Stevens Institute of Technology