Pairs Trading and Spread Analysis
Pairs trading takes two instruments that move together and bets on the moments they stop doing so, and it is the cleanest way to trade a relationship instead of a direction. Spread analysis is the craft underneath: pricing that relationship, timing its excursions, and abandoning it when the statistics die. The position is one trade built from two legs: long the cheaper member of the pair, short the richer one, sized so the market's direction largely cancels and what remains is the spread between them. The cross-market correlation map surveys which assets travel together; the correlated-pair divergence read treats a split between correlated instruments as a signal about the stronger one; and the dollar's grip on both legs explains why currency swings break relationships that looked permanent. Pairs trading goes a step further than any of those: it converts the relationship itself into a tradable instrument with its own price, its own statistics, and its own discipline for entry, exit, and abandonment.
One Position From Two Instruments
The logic starts with why two things move together. Two oil producers share the commodity price, the cost structure, and the investor base, so their stocks co-move most days and their differences reflect company-specific news rather than sector direction. When one leg gaps on idiosyncratic news and the other does not, the pair's spread widens, and the bet is that the gap closes once the news is absorbed. The hedge ratio is what turns two stocks into one position: for every 100 shares of the cheaper leg, the trader shorts a calculated number of the richer leg, chosen so that a one-percent move in the sector moves both legs by the same dollar amount. Done right, the combined position earns nearly nothing when the sector rallies or sinks together and earns when the two legs converge or diverge, which is why practitioners describe it as trading the spread, not the stocks.
Building the Spread and Its Statistics
The spread is a series with its own history, and three statistics from that history set the rules. The mean is the spread's gravitational center, the level it revisits when nothing unusual is happening. The standard deviation measures how far the spread normally wanders before returning, and dividing the current spread's distance from the mean by that deviation produces the z-score, the single number most pair traders act on. The half-life, estimated from how quickly the spread decays back after each excursion, converts the statistics into time: a 15-day half-life means the average divergence has historically healed by half in roughly two weeks, which sets a natural clock for the trade. Entries live in the tails, typically beyond two standard deviations, exits live near the middle, and the stop lives beyond the level where the story changes from unusual to broken.
Plot the two prices on one panel and the spread with its mean and bands beneath; the two prices travel together while the spread does the oscillating.
A Worked Example: One Spread Round-Trip
Take two large oil producers, leg A trading at 112.40 and leg B at 148.75, with a hedge ratio of 0.76 estimated over two years of daily data. The spread, defined as the price of A minus 0.76 times the price of B, averages 1.10 with a standard deviation of 1.30 over the trailing 60 sessions, and the estimated half-life of deviations is 18 days. Early in the month the spread begins widening in B's favor: B climbs on a drilling-rights story while A drifts, and by day 12 the spread reads minus 1.55, which is a z-score of minus 2.04. The trade goes on: long 10,000 shares of A at 110.47, short 7,600 shares of B at 147.40, dollar-neutral to within rounding.
| Day | Leg A | Leg B | Spread | Z-score | Action |
|---|---|---|---|---|---|
| 1 | 112.05 | 148.05 | 0.95 | -0.12 | watch |
| 8 | 111.30 | 147.90 | -0.85 | -1.50 | alert |
| 12 | 110.47 | 147.40 | -1.55 | -2.04 | enter |
| 24 | 110.90 | 146.90 | 0.42 | -0.52 | hold |
| 35 | 111.80 | 145.76 | 1.02 | -0.06 | exit |
The convergence takes 23 days, longer than the 18-day half-life estimate but inside its own uncertainty, and the exit fills at a z-score of minus 0.06 with the spread back at 1.02. The arithmetic of the round-trip: A rises 1.33 per share for 13,300 dollars on the long leg, B falls 1.64 for 12,464 dollars on the short leg, a combined 25,764 dollars before costs on roughly 2.2 million dollars of gross exposure, about 1.0 percent net for the hold. The trade paid for the wait inside its own half-life.
What Breaks a Pairs Trade
Every pair carries a silent assumption: the relationship is structural, not incidental. The break case teaches why the stop exists. Four weeks into a routine trade, leg B announces an acquisition at a premium and jumps 8.4 percent; the spread blows through the minus 3.2 level while the trailing correlation collapses from 0.87 to 0.41. The stop was set at minus 3.0 on the z-scale, a level chosen at entry as the point where the divergence stopped looking statistical, and the fill came at minus 3.34. The model died quietly, one widening week at a time.
Three failure modes cover most of the graveyard. Structural breaks, mergers, spinoffs, regulatory shocks, change the reason the pair existed, and no z-score survives that. Regime breaks shift the parameters: a spread whose mean and deviation were stable for years can re-anchor after a commodity cycle turns, which is why parameters get re-estimated on trailing windows rather than frozen. And crowding breaks the exit: when many books hold the same mean-reversion trade, the convergence everyone waits for becomes a stampede for the same door at the first sign of trouble. The discipline that survives all three is the same: size for the stop, respect the half-life clock, and treat the thesis as a hypothesis with an expiration date.
Costs sit closer to the P&L in this structure than in any directional trade, because the edge per round-trip is modest by construction. The short leg owes borrow on its full notional for the life of the trade, and a hard-to-borrow leg can turn a 1.0 percent gross win into a scratch. Dividends run the same direction: the short leg pays them out, and ex-dividend dates inside the hold are a known drag to schedule around. Slippage doubles, since every entry and exit is two executions, and the exit is usually hurried, which makes the convergence-day spread worth more than any indicator tweak. The practical rule: estimate the all-in cost per round-trip before entry, and require the expected spread move net of costs to clear the hurdle with room. A pair whose average excursion is barely triple the cost structure is a market-making operation, not a trading strategy, and the half-life clock should be short enough that time decay does not eat the remainder.
Pairs Trading Questions
Four questions cover most of what traders ask about pairs.
Do I need cointegration or is correlation enough?
Correlation tells you the two legs co-move; cointegration tells you their spread tends to return to a stable level, which is the property the trade actually monetizes. Two series can correlate strongly while their spread drifts without bound, and trading such a spread is catching a falling knife with extra steps. The practical screen requires a mean-reverting spread over the estimation window, and the half-life estimate doubles as the test: if deviations never decay, the statistics will say so.
How is the hedge ratio chosen?
The commonest estimate is a regression of one leg's returns on the other over the trailing window, which yields a ratio that equalizes sensitivity to the shared factor. Ratios drift as the businesses drift, so practice re-estimates on a rolling window and avoids re-hedging mid-trade on small changes. A ratio that needs weekly surgery is a sign the pair itself is unstable. Re-estimation windows matter as much as the method: a ratio fit over five years describes a different pair than one fit over sixty sessions, and the estimation window should roughly match the horizon of the trades, which keeps the statistics honest about how fast the relationship actually moves.
What position size makes sense?
Size follows from the stop, not the conviction. A common construction risks a fixed fraction of capital between entry and stop, with the dollar-neutral hedge equalizing sector exposure, which makes the trade's outcome depend on convergence alone. Gross exposure runs high by design since the legs offset, so exposure discipline matters even though directional risk is small. The stop belongs in the same arithmetic: risk per trade equals the distance from entry to the abandonment level, times the spread's dollar value per point, and pairs sized this way lose predictably when wrong, which is the property that keeps a book of them survivable through the break cases.
How does this differ from trading a single stock?
The single-stock trade carries market risk and company risk; the pair strips most of the market component and isolates the relative call. That isolation cuts both ways: the trade can be right about the relationship and still lose to a break, while the single-stock trader can be wrong about the market and still win on idiosyncratic news. Pairs trade cleaner precisely because they strip the story down to one question, and that question has a stop.