lev-os Structuring Pairs Trading Strategies
Designs statistical arbitrage pairs with cointegration analysis, spread dynamics, and entry/exit signal calibration.
When To Use
- Building a new pairs trading strategy from candidate universe screening through live signal generation
- Evaluating whether two or more instruments exhibit a stable, tradeable cointegrating relationship
- Calibrating entry/exit thresholds and position sizing for an existing spread
- Diagnosing strategy decay — determining whether a pair's statistical relationship has broken down
- Comparing multiple candidate pairs to select the highest-conviction opportunities
Inputs To Gather
- Instrument universe: Tickers, asset class (equities, ETFs, futures, FX), and exchange/venue
- Historical price data: Adjusted close prices; minimum 2–5 years daily or equivalent intraday bars; confirm corporate action adjustments for equities
- Fundamental linkage rationale: Sector/industry overlap, supply-chain relationship, macro factor exposure, or structural reason the pair should mean-revert
- Trading constraints: Account size, margin requirements, borrow availability/cost, maximum holding period, commission/slippage assumptions
- Risk parameters: Maximum drawdown tolerance, per-trade loss limit, gross/net exposure caps, correlation budget within broader portfolio
- Regime context: Current volatility regime (VIX level, realized vol percentile), recent structural breaks in the sector, pending catalysts (earnings, M&A, index rebalance) [VERIFY against live market data]
Workflow
Screen candidate pairs
- Filter universe by sector, market cap, liquidity (minimum ADV), and borrow availability
- Compute rolling pairwise correlations (60d, 120d, 252d) and rank by stability
- Apply Engle-Granger or Johansen cointegration tests on log-price series; retain pairs with p-value < 0.05 across multiple lookback windows [VERIFY test assumptions: stationarity of residuals, no structural break in sample]
Estimate spread dynamics
- Fit the cointegrating regression: log(P_A) = β · log(P_B) + μ + ε; record hedge ratio β and intercept
- Test residual series for stationarity (ADF, KPSS); estimate half-life of mean reversion via Ornstein-Uhlenbeck calibration
- Compute rolling z-score of the spread; assess distribution properties (skew, kurtosis, fat tails)
- If half-life exceeds maximum holding period constraint, flag pair as unsuitable
Calibrate entry/exit signals
- Set entry thresholds: typically ±1.5–2.5σ from spread mean; optimize via walk-forward backtest, not in-sample curve-fitting
- Set exit thresholds: mean reversion target (0σ) and/or profit-take level; define stop-loss at ±3–4σ or dollar-based max loss
- Evaluate asymmetric entry (long-spread vs. short-spread) if spread distribution is skewed
- Determine position sizing: equal-dollar, beta-neutral, or volatility-weighted; compute notional per leg
Backtest and stress-test
- Run walk-forward backtest with realistic transaction costs (commissions, bid-ask spread, borrow cost, market impact)
- Report: Sharpe ratio, Sortino ratio, max drawdown, win rate, average holding period, profit factor
- Stress-test against regime changes: 2008 credit crisis, 2020 COVID dislocation, sector rotation events
- Test sensitivity to hedge ratio drift — re-estimate β on rolling windows and measure P&L degradation
- Confirm no survivorship bias or look-ahead bias in data
Define execution and monitoring plan
- Specify order types (limit vs. MOC), leg sequencing (simultaneous vs. legged), and execution venue preferences
- Set re-hedge frequency for β drift (e.g., weekly recalibration if β moves > 5%)
- Define kill criteria: pair is closed and removed if cointegration test fails on trailing 6-month window or if cumulative loss exceeds stop threshold
- Document escalation triggers for manual review (spread hitting 4σ+, sudden liquidity drop, corporate event on either leg)
Output
Deliver a Pairs Trade Strategy Report containing:
- Pair summary table: Ticker pair, sector, hedge ratio (β), spread half-life, cointegration p-value, correlation
- Spread chart: Historical spread with z-score overlay, entry/exit bands, and marked trade signals
- Signal parameters: Entry z-score, exit z-score, stop-loss z-score, position sizing method, notional per leg
- Backtest results: Performance metrics table (Sharpe, Sortino, max DD, win rate, avg hold, profit factor), equity curve, drawdown chart
- Risk summary: Max concurrent exposure, worst-case scenario P&L, margin requirement estimate, borrow cost impact
- Execution plan: Order type, rebalancing schedule, kill criteria, monitoring dashboard requirements
- Assumptions and limitations log: All [VERIFY] items, data quality notes, model limitations
Quality Checks
- Cointegration holds across at least two independent lookback windows (e.g., 2-year and 5-year)
- Hedge ratio is economically plausible (not extreme values suggesting spurious fit)
- Backtest Sharpe > 1.0 after realistic transaction costs; if below, flag as marginal
- No single trade accounts for >25% of total backtest P&L (guards against curve-fitting)
- Half-life is within feasible holding period (typically 5–60 trading days for daily strategies)
- Walk-forward out-of-sample results do not degrade >30% vs. in-sample
- All data is survivorship-bias-free and adjusted for splits, dividends, and delistings [VERIFY data vendor methodology]
- Borrow availability confirmed for short leg; cost incorporated into P&L estimates [VERIFY with prime broker or locate desk]