How does stock pairs trading performance interact with lagged pair correlation and volatility? In her May 2016 paper entitled “Demystifying Pairs Trading: The Role of Volatility and Correlation”, Stephanie Riedinger investigates how stock pair correlation and summed volatilities influence pair selection, pair return and portfolio return. Her baseline is a conventional pairs trading method that each month: (1) computes sums of daily squared normalized price differences (SSD) for all possible stock pairs over the last 12 months and selects the 20 pairs with the smallest SSDs; (2) over the next six months, buys (sells) the undervalued (overvalued) member of each of these pairs whenever renormalized prices diverge by more than two selection phase standard deviations; and, (3) closes positions when prices completely converge, prices diverge beyond four standard deviations, the trading phase ends or a traded stock is delisted. A pair may open and close several times during the trading period. At any time, six pairs portfolios trade simultaneously. She modifies this strategy to investigate correlation and volatility effects by: (1) measuring also during the selection phase return correlations and sum of volatilities based on daily closing prices for each possible stock pair; (2) allocating each pair to a correlation quintile (ranked fifth) and to a summed volatility quintile; and, (3) randomly selecting 20 twenty pairs out of each of the 25 intersections of correlation and summed volatility quintiles. She accounts for bid-ask frictions by executing all buys (sells) at the ask (bid) and by calculating daily returns at the bid. Using daily bid, ask and closing prices for all stocks included in the S&P 1500 during January 1990 (supporting initial pair trades in January 1991) through December 2014, she finds that:
- On average over the sample period, a conventional pairs trading portfolio generates average monthly risk-adjusted (accounting for multiple factors) return 0.37%.
- 20 of 25 pair correlation/summed volatility-sorted portfolios outperform the conventional portfolio with average monthly risk-adjusted returns ranging from 0.18% to 2.09% (The highest value is not statistically reliable, but the next-highest return of 0.76% is reliable).
- Regarding interaction of pair correlations and summed volatilities with SSD:
- Pair correlations (positive relationship) and summed volatilities (negative relationship) explain 88% of the variation in SSD across pairs.
- About two thirds of all pairs selected based on SSD fall in the highest pair correlation quintile and the lowest summed volatilities quintile.
- Regarding building blocks of pairs trading profitability:
- Higher summed volatilities and lower correlations increase average return of converging pairs, but also increase average loss of diverging pairs.
- Lower summed volatilities and higher correlation raise the total number of trades.
- Across the 25 sort portfolios, performance generally (but not perfectly) increases with both pair correlation and summed volatilities. In other words:
- The positive summed volatilities effect on return per trade and the positive correlation effect on trading frequency dominate outcomes.
- SSD, which selects pairs with low summed volatilities, is suboptimal.
- Limiting the universe to highly liquid current S&P 100 or NASDAQ 100 stocks lowers pairs trading returns, suggesting that illiquidity explains part of the above results, but general findings still hold.
- Long side portfolio return contributions are below 50% for 23 of 25 sort portfolios, generally decreasing with summed volatilities, such that shorting constraints may limit exploitability.
In summary, evidence indicates that pairs trading portfolio performance relates positively to both pair correlation and pair summed volatilities during the selection phase.
Cautions regarding findings include:
- While the methodology accounts for the bid-ask spread, it does not account for transaction fees or impact of trading. These additional frictions would reduce reported returns.
- Testing many pair correlation/summed volatilities sorts introduces snooping bias, such that the best-performing portfolio overstates expectations.
- Calculations required to implement such a pairs trading strategy are likely beyond the reach of most investors, who would bear a fee for delegating to an investment manager.
