Status of this page: P0 draft. Methods are standard; empirical performance claims are limited to what the cited papers report. Validation of any pairs strategy should follow Backtest bias taxonomy.
1. Two ways to form pairs
Distance method (Gatev, Goetzmann and Rouwenhorst, 2006). Normalise price series, pick the pair with the smallest sum of squared deviations over a formation period, and trade divergences beyond a threshold during a trading period. The paper reports average annualised excess returns of up to about 11% on self-financing portfolios over 1962-2002, with profits typically above conservative cost estimates, and links profitability to a common factor. Later decay in profitability is documented in the literature; treat the 1962-2002 figure as historical. Cointegration method. Test whether a linear combination of two non-stationary prices is stationary (Engle and Granger, 1987). The residual is the spread.2. Hedge ratio estimation
Estimate beta only on the formation window and freeze it in the test window, otherwise the test is contaminated.
3. Ornstein-Uhlenbeck residual model
Model the spread X as dX = kappa (m - X) dt + sigma dW. Discretely, X_ = a + b X_t + e, with b = exp(-kappa dt). Key outputs:- Half-life = -ln 2 / ln b (in sampling units). Verified examples: b = 0.90 gives 6.6 periods, b = 0.95 gives 13.5, b = 0.99 gives 69.0.
- Equilibrium std = sigma / sqrt(2 kappa).
- s-score = (X - m) / sigma_eq, used as the entry/exit variable.
4. Practical filters
- Cointegration p-value is not a stability guarantee; re-test on rolling windows and require economic linkage (same industry, ETF vs constituents, ADR vs local line, share classes).
- Remove pairs with event risk: M&A, index inclusion, suspension.
- Size by spread volatility, cap by liquidity of the thinner leg.
- Cost model must include short-borrow fee and the crossing cost on both legs.
5. Failure modes
- Regime break: the pair stops being related; mean reversion becomes trend. Use a stop on the s-score and on time-in-trade.
- Crowding: many participants unwind together.
- Multiple testing: scanning thousands of pairs guarantees spurious cointegration. Apply a multiple-testing correction; Harvey, Liu and Zhu (2016) argue for a t-ratio hurdle above 3.0 for new factor claims, and the same logic applies to pair scanning.
Points to verify
- Exact filter threshold and sample used in Avellaneda and Lee (2010) before citing it numerically.
- Short-sale availability and uptick/designated-security rules for the HK names used.
Sources
- Gatev, Goetzmann and Rouwenhorst (2006), Pairs Trading: Performance of a Relative-Value Arbitrage Rule, Review of Financial Studies 19(3).
- Engle and Granger (1987), Co-integration and Error Correction, Econometrica 55(2). Link
- Avellaneda and Lee (2010), Statistical arbitrage in the US equities market, Quantitative Finance.
- Harvey, Liu and Zhu (2016), and the Cross-Section of Expected Returns, RFS 29(1). NBER

