Pairs Trading System

Production-Deployed Statistical Arbitrage Engine (US Equities)

Live Monitoring Dashboard:
http://uspairs.dr2112.com

📸 Screenshots:
See the /screenshots folder for a quick view of the dashboard.

---

Overview

This project is a fully deployed statistical arbitrage system that trades mean reversion across 50+ US equity pairs.

The system:

• Runs continuously on a VPS
• Evaluates signals every minute during NYSE hours
• Simulates execution via a stateful portfolio engine
• Exposes a real-time monitoring dashboard

It is designed as a production trading system

---

System Architecture

            Polygon.io
                 │
                 ▼
        Historical Data Loader
                 │
                 ▼
      ┌─────────────────────┐
      │   Analytics Layer   │
      │ (15m statistical)   │
      └─────────────────────┘
                 │
                 ▼
      ┌─────────────────────┐
      │   Strategy Engine   │
      │  (1m signal logic)  │
      └─────────────────────┘
                 │
                 ▼
      ┌─────────────────────┐
      │ Portfolio Simulator │
      │  Position Manager   │
      └─────────────────────┘
                 │
                 ▼
        SQLite Trading DB
          │            │
          ▼            ▼
   Streamlit Dashboard  Telegram Alerts

---

Strategy Design

Spread Model

Spread is defined as a price ratio:

spread = price_1 / price_2

Z-score:

z = (spread - rolling_30d_mean) / rolling_30d_std

Features:

• 30 trading day lookback
• Weekly rebalance of slow statistics (Fridays only)
• Statistics forward-filled intra-week
• Lookahead bias removed via shift(1)

---

Entry & Exit Rules

Entry

|z| ≥ 3.5

Additional signal filter:

• 2-of-3 consecutive 1-minute bars must exceed the threshold
• Reduces microstructure noise and false signals

Exit

|z| ≤ 1.0

Dynamic widening near events:

• Earnings multiplier
• Dividend multiplier
• Macro event multiplier

Maximum exit threshold cap:

|z| ≤ 1.5

Re-entry

• No cooldown
• New signals allowed immediately after exit

---

Hedge Ratio

Two-layer hedge system:

1️⃣ Direct Beta

• Instantaneous hedge ratio
• Updated every bar

2️⃣ Rolling Beta

• 30-day rolling regression
• Rebalanced weekly
• Forward-filled intra-week

Purpose:

• Reduce hedge noise
• Preserve statistical calibration

---

Position Sizing

Beta-adjusted dollar-neutral sizing:

capital_per_trade = equity * alloc_pct

leg1 = capital_per_trade / (1 + beta)
leg2 = leg1 * beta

Default parameters:

Parameter	Value
Starting Equity	$100,000
Allocation per trade	6.5%

---

Transaction Cost Model

A realistic four-component cost model is applied to every trade.

1. Broker Commission (IBKR Fixed)

$0.005 per share
$1.00 minimum per leg

Applied to all four executions in a round-trip pairs trade.

---

2. Regulatory Fees (Sell-Side Only)

FINRA TAF:

$0.000195 per share sold
Cap: $9.79 per trade

SEC Section 31:

$0.00 per dollar sold
(Current rate as of May 2025)

Applied to the two sell executions.

---

3. Short Borrow Cost

Default:

0.5% annual

Accrued daily on short-leg notional.

Calculation:

borrow_cost = short_notional * rate * days_held / 365

---

4. Slippage

Half-spread estimate:

1 basis point per execution

slippage = notional * 1 / 10000

---

Cost Impact

• All costs deducted at trade close
• Unrealized P&L reflects estimated round-trip costs

Trade records store:

• Commission
• Regulatory fees
• Borrow cost
• Slippage

for full cost attribution.

---

Risk Management

The system implements multi-layered risk protection:

• Beta drift entry filter (6.5%)
• Beta drift in-position stop (10%)
• Earnings event multiplier (1.5×)
• Dividend multiplier (1.3×)
• Macro event multiplier (1.2×)
• Exit threshold cap (1.5)
• Market-hours gating
• One position per pair
• Persistence filter
• Data staleness detection
• Engine heartbeat monitoring
• Telegram alert throttling

Rather than disabling trading around events, the system adapts risk dynamically.

---

Performance Tracking

Equity Curve

Equity snapshots:

• Every 15 minutes intraday
• End-of-day mark-to-market equity

Includes:

• Realized P&L
• Unrealized P&L

---

Sharpe Ratio

Computed from daily MTM equity returns:

Sharpe = mean(daily_returns) / std(daily_returns) * sqrt(252)

---

Dashboard Metrics

The monitoring dashboard displays:

• Total Trades
• Win Rate
• Sharpe Ratio
• Average P&L per Trade
• Average Hold Time
• Total Fees Paid

---

Telegram Notifications

Real-time alerts include:

Entry Alerts

• Pair
• Z-score
• Position sizing

Exit Alerts

• Net P&L
• Fee breakdown
• Holding period

End-of-Day Report

• Realized P&L
• Total fees paid
• Equity
• Win/loss count

---

Architecture

The engine uses a dual-layer signal architecture.

Slow Analytics Layer (15m Bars)

Responsible for:

• Rolling mean
• Rolling standard deviation
• Hedge ratio estimation

Characteristics:

• Weekly rebalance
• Cached computations

---

Fast Execution Layer (1m Bars)

Responsible for:

• Signal evaluation
• Risk filter application
• Trade simulation
• Portfolio updates

Signals evaluated every 60 seconds.

---

Production Infrastructure

Deployment stack:

Component	Description
VPS	DigitalOcean (Ubuntu)
Services	systemd-managed processes
Database	SQLite (WAL mode)
Dashboard	Streamlit terminal UI
Alerts	Telegram bot
Deployment	GitHub → VPS auto-pull

Features:

• Crash recovery via persistent state
• Concurrent engine/dashboard database access
• Event calendar integration
• Modular architecture

analytics → strategy → execution → persistence → monitoring

---

Data Sources

Market Data:

• Polygon.io
• 1-minute OHLCV
• 15-minute OHLCV

Additional datasets:

• Earnings calendar
• Dividend events

Historical warmup:

75 trading days

Storage:

• Append-only CSV history
• SQLite trading database

---

Scale & Performance

The engine currently supports:

• 50+ concurrent pairs
• 1-minute signal evaluation
• Medium-frequency stat arb

Design focus:

• Robustness
• Operational reliability
• Research reproducibility

(Not ultra-low-latency HFT)

---

What Differentiates This System

Key design features:

• Weekly hedge rebalance to reduce beta noise
• Explicit lookahead bias prevention
• Event-aware dynamic exit thresholds
• Beta drift enforcement during entry and holding
• Persistence filter to reduce microstructure noise
• Full four-component transaction cost model
• Mark-to-market equity with proper Sharpe calculation
• Real-time monitoring dashboard
• Production deployment infrastructure

This project bridges research logic and production trading systems.

---

Disclaimer

This project is provided for educational and research purposes only.

It does not constitute investment advice and should not be used for live trading without proper risk controls and regulatory compliance.