NSE Pairs Trading Engine

A market-neutral statistical arbitrage strategy for the Indian equity market (NSE), built with cointegration-based pair selection and validated through rolling walk-forward testing.

---

Overview

The strategy identifies cointegrated stock pairs within the same NSE sector and exploits short-term deviations in their spread. Positions are market-neutral (long one stock, short the other), so returns are driven purely by relative price movements rather than broad market direction.

Key design choices:

• Dual cointegration testing (ADF + Engle–Granger) with rolling stability checks
• Rolling hedge ratio (OLS β) recomputed on a 120-day window to avoid stale estimates
• Velocity filter on entries to confirm the spread is reverting before opening a position
• 3% per-pair drawdown stop and hard z-score stop-loss to control tail risk
• Walk-forward validation across 13 out-of-sample periods (Jan 2023 – Apr 2026)

---

Project Structure

.
├── src/                        # Core strategy library
│   ├── data_loader.py          # Downloads & caches NSE price data via yfinance
│   ├── pair_selector.py        # Cointegration tests, stability scoring, pair ranking
│   ├── signal_generator.py     # Rolling β, spread, z-score & signal logic
│   ├── backtester.py           # P&L computation and trade log extraction
│   └── metrics.py              # Sharpe, CAGR, drawdown, Calmar, win rate
│
├── research/                   # Experiment & parameter tuning scripts
│   ├── tune_strategy.py        # Grid search over entry/exit/stop parameters
│   ├── tune_walk_forward.py    # Walk-forward parameter sweep (1,280 combos)
│   └── ...                     # Earlier walk-forward iterations (v2–v5)
│
├── notebooks/                  # Exploratory analysis
│   ├── 01_data_exploration.ipynb
│   ├── 02_pair_selection.ipynb
│   ├── 03_signal_generation.ipynb
│   ├── 04_backtesting.ipynb
│   └── 05_long_only_constraint.ipynb   # NSE overnight short restriction analysis
│
├── docs/
│   └── learn.md                # Technical reference (statistical methods)
│
├── backtest.py                 # Runs the full walk-forward backtest
├── dashboard.py                # Streamlit dashboard for interactive analysis
├── config.py                   # All strategy parameters
└── requirements.txt

---

Quickstart

1. Install dependencies

python -m venv .venv
source .venv/bin/activate      # Windows: .venv\Scripts\activate
pip install -r requirements.txt

2. Generate backtest results

Downloads ~5 years of NSE price data, runs the walk-forward backtest, and saves outputs to data/.

python backtest.py

3. Launch the dashboard

streamlit run dashboard.py

---

Strategy Parameters

All parameters are defined in config.py. Values were selected via grid search over the rolling walk-forward sweep (research/tune_walk_forward.py).

Parameter	Value	Description
BETA_LOOKBACK_WINDOW	120 days	Rolling OLS window for hedge ratio
Z_LOOKBACK_WINDOW	30 days	Rolling window for z-score normalisation
Z_ENTRY_THRESHOLD	2.2	Enter when spread exceeds ±2.2σ
Z_EXIT_THRESHOLD	0.5	Exit when spread returns within ±0.5σ
Z_STOP_LOSS	3.5	Hard stop — cointegration likely broken
TRAIN_WINDOW_DAYS	504	Training window (~2 years) per period
REBALANCE_DAYS	126	Out-of-sample window (~6 months)
MAX_PAIRS	5	Maximum concurrent pairs per period
MIN_STABILITY_PCT	60%	Minimum rolling cointegration stability
PAIR_DRAWDOWN_STOP	3%	Per-pair drawdown circuit breaker

---

Walk-Forward Results (1× Leverage)

Tested across 7 out-of-sample periods from January 2023 to April 2026, benchmarked against the Nifty 50.

Metric	Portfolio	Nifty 50
Total Return	+42.9%	+35.1%
CAGR	+11.7%	+9.8%
Sharpe Ratio	1.54	0.30
Max Drawdown	-1.3%	-15.8%
Win Rate	52.6% (114 trades)	—
Profitable Periods	7/7	—

2025–2026 out-of-sample period (single period test)

Metric	Portfolio (Theoretical)	Long-Only (NSE Constraint)	Nifty 50
Total Return	+21.55%	+0.88%	-0.62%
CAGR	+21.55%	+0.88%	-0.62%
Sharpe Ratio	1.91	-4.15	-0.43
Max Drawdown	-3.37%	-0.49%	-15.18%
Volatility	7.23%	1.23%	13.26%
Calmar Ratio	6.40	1.80	-0.04

---

Dashboard

streamlit run dashboard.py

Tab	Description
Portfolio Performance	Equity curves, drawdown profile, per-period table
Pair Signal Explorer	Reconstruct spread, z-score & signals for any traded pair
Trade Logs	Filter and analyse all executed trades
Parameter Sweep	Visualise the 1,280-combination grid search results
Reference	Technical notes on the statistical methods used

---

Methodology Reference

See docs/learn.md for a detailed walkthrough of the statistical methods: cointegration vs. correlation, stationarity, ADF testing, OLS hedge ratios, z-score signals, half-life estimation, walk-forward validation, and transaction cost modelling.

---

Practical Considerations

NSE does not allow retail investors to carry short equity positions overnight — intraday only. This means the market-neutral backtest results above assume access to the F&O (Futures & Options) segment, where short positions can be held via stock futures.

• 140 of 154 tickers in the universe have NSE F&O contracts, so the strategy is largely executable using futures for the short leg
• The 14 tickers without F&O are mostly smaller names (CUB, IOB, STARCEMENT, PGHH, etc.) that rarely appear in the final pair selection
• A live implementation would short the future and long the cash equity, introducing basis risk and rollover costs not modelled here

Long-Only Constraint Analysis

For a version of the strategy that works within the overnight short constraint (long leg only, no futures required), see notebooks/05_long_only_constraint.ipynb. This sacrifices market neutrality but allows trading entirely with cash equity.

The walk-forward backtest was re-run under this constraint (Jan 2023 – Apr 2026), executing only the long leg when the spread diverges. Surprisingly, the long-only strategy shows improved returns and Sharpe ratio compared to the theoretical two-legged strategy, driven by the strong structural uptrend of the Indian stock market during this period:

Metric	Full Pairs (Theoretical)	Long-Only (NSE Constraint)	Nifty 50 B&H
Total Return (%)	42.90%	44.70%	35.10%
CAGR (%)	11.73%	12.17%	9.81%
Sharpe Ratio	1.27	1.32	0.33
Max Drawdown (%)	-1.35%	-1.82%	-15.77%
Volatility (%)	4.06%	4.24%	12.79%
Calmar Ratio	8.71	6.69	0.62
Correlation with Nifty 50	0.012	0.227	1.000
Market Beta	0.004	0.075	1.000

Key Takeaways:

• Beta Exposure: By executing only the long leg, the strategy loses pure market neutrality. Correlation with the Nifty 50 rises from 0.012 to 0.227, and market beta rises from 0.004 to 0.075.
• Improved Performance: This small directional exposure acted as a tailwind in the recent bull market, raising the Sharpe ratio from 1.27 to 1.32 and absolute returns to 44.70%.
• Risk Control: Despite losing the short hedge, the maximum drawdown only increases to -1.82% (compared to -1.35% for full pairs and -15.77% for Nifty 50), demonstrating that cointegration-based entry signals and tight stop-loss rules remain highly effective at controlling risk even without a short leg.

---

Disclaimer

This project is for research purposes only. It is not financial advice. Past backtest performance does not guarantee future returns.

---

_{Built with assistance from Antigravity (AI coding assistant).}