Pairs Trading & Cointegration

In the 1980s a quant team at Morgan Stanley noticed something almost too simple to be a strategy: when two near-identical stocks — say two soda giants, or two oil majors — drifted apart, they almost always snapped back together. So they built a rule. Buy the cheap one, short the expensive one, wait for the gap to close, pocket the difference, repeat. No view on the market, no view on whether stocks go up or down — just a bet that the distance between two things that belong together won’t stay stretched forever. That trade is pairs trading, and it is the ancestor of the entire statistical-arbitrage industry.

The whole game lives or dies on one question: when is a gap a real rubber band that will snap back, versus two ships that merely happened to sail near each other before drifting off to opposite oceans? Get that wrong and you are not arbitraging — you are catching a falling knife with both hands. The mathematical tool that tells the two apart is cointegration, and as we’ll see, it is emphatically not the same thing as the correlation everyone reaches for first.

The pairs-trade idea

Analogy. Picture two dogs — Coke and Pepsi — being walked by the same owner across a field. The owner is “the soft-drink sector,” and as the owner strolls, both dogs roughly follow. The dogs are joined by an elastic leash: one can sprint ahead and the other lag, but the leash keeps yanking them back toward each other. A pairs trader doesn’t bet on where the owner is going (that’s market direction — irrelevant here). They bet on the leash: when one dog bolts too far ahead, short it and buy the laggard, and collect when the leash reins them back in.

Definition. A pairs trade picks two assets driven by the same underlying economic force — two soda companies, two oil majors, an ETF and the basket of stocks it holds, gold and gold miners — and trades the relative price, not the absolute price:

• Long the cheap leg (the one that has fallen behind).
• Short the rich leg (the one that has run ahead).

Because you hold one long and one short of comparable size, a market-wide move lifts (or sinks) both legs together and cancels out. The position is market-neutral by construction: you’ve hedged away “the owner” and kept only “the leash.”

When to use it

Pairs trading shines when you can name a genuine economic tether between two assets — same industry, same input costs, one is literally a basket containing the other. It struggles when the “pair” is a coincidence dredged up by scanning thousands of tickers for whatever lines up in-sample (data-mined pairs break the instant you trade them). Rule of thumb: if you can’t explain why these two should be tied together to a skeptical colleague in one sentence, you probably have a spurious pair, not a tradable one.

Correlation is not enough

Analogy. Two hikers leave the same trailhead and, for an afternoon, happen to walk in roughly the same compass direction — both drifting north-ish. Their step directions are highly correlated. But there’s no rope between them. By nightfall one is on a ridge and the other in a valley, miles apart, and nothing pulls them back. High correlation, zero anchor. A pairs trade on these two would have you betting the gap closes — and it never does.

The crux. Correlation measures whether two series’ returns (their step-to-step changes) move together. It is silent about levels — about whether the gap between the prices has any tendency to return to a fixed value. You can manufacture two random walks whose daily increments are correlated (push both with a shared shock plus a little private noise) and they will show a fat correlation over any given window — yet because each is a random walk, their difference is itself a random walk, free to wander off to infinity. Correlation: high. Reversion: none.

A stationary series has a fixed long-run mean and variance — it wiggles around a level and keeps coming home. A non-stationary series (the classic example: a random walk, also called integrated of order 1, or $I(1)$) has no such home; today’s level is just yesterday’s plus a fresh shock, so it drifts indefinitely. Prices are almost always $I(1)$. The question for a pairs trade is whether some combination of two $I(1)$ prices manages to be stationary.

The fingerprint of these two regimes shows up vividly in the autocorrelation function — how related a series is to its own past:

Cointegration — the real prize

Analogy. Back to the two dogs on the elastic leash. Individually, each dog’s position is a random walk — it wanders all over the field with no fixed spot it returns to (each is $I(1)$). But the leash length — the distance between them — is pinned: it stretches and contracts but always snaps back to its natural length. That leash is a stationary combination of two non-stationary things. That, precisely, is cointegration.

Definition. Two non-stationary ($I(1)$) price series $P_a$ and $P_b$ are cointegrated if there exists a constant $\beta$ — the hedge ratio — such that the linear combination

$S_t = P_{a,t} - \beta\, P_{b,t}$

is stationary ($I(0)$): it has a fixed mean $\mu$ and reverts to it. The vector $(1, -\beta)$ is called the cointegrating vector, and $S_t$ is your spread. The prices can each wander to the moon, but $\beta$ is tuned so their weighted difference stays leashed.

The hedge ratio $\beta$ answers “how many units of $B$ does it take to neutralize one unit of $A$?” If asset $A$ typically moves $1.50 for every $1.00 move in $B$, then $\beta \approx 1.5$, and a dollar-neutral-ish spread shorts 1.5 units of $B$ against each unit of $A$.

Estimating β — the Engle–Granger two-step

The classic recipe (Engle & Granger, 1987 — it won Granger a Nobel) is two moves:

1. Estimate the cointegrating vector. Regress $P_a$ on $P_b$ by ordinary least squares: $P_{a,t} = \alpha + \beta\,P_{b,t} + \varepsilon_t$. The slope $\hat\beta$ is your hedge ratio; the residual $\hat\varepsilon_t$ is the spread.
2. Test the residual for stationarity. Run an Augmented Dickey–Fuller (ADF) test on $\hat\varepsilon_t$. Its null hypothesis is “the residual has a unit root” (i.e. it’s a non-stationary random walk — no cointegration). A small ADF p-value (say < 0.05) lets you reject that null and conclude the spread is stationary: the pair is cointegrated, and you have a tradable rubber band.

Worked example. You regress KO on PEP and OLS returns $\hat\alpha = 4.0$, $\hat\beta = 0.80$. So your spread is

$S_t = P_{\text{KO}} - 0.80\,P_{\text{PEP}} - 4.0.$

You feed the residual series into an ADF test and get a p-value of 0.012. Since $0.012 < 0.05$, you reject the unit-root null: the spread is stationary. The pair is cointegrated with hedge ratio 0.80 — short 0.80 shares of PEP for every share of KO and the combination reverts. (If instead the p-value had been 0.34, you’d fail to reject, conclude no cointegration, and walk away — no rubber band.)

When to use it

Cointegration is the right lens whenever you suspect a structural tie that pins a long-run relationship: an ETF and its underlying basket (near-mechanical), two share classes of the same company, a commodity and its producers, cross-listed shares in two currencies. Re-estimate $\beta$ on a rolling window — the equilibrium relationship can slowly shift, and a stale hedge ratio quietly stops being market-neutral.

Building and trading the spread

You’ve established the pair is cointegrated and you have $\beta$. Now you have to turn the spread into signals. Raw spread values (“the gap is $3.20”) are hard to act on — is $3.20 a lot? It depends on the spread’s typical scale. So you standardize it into a z-score: how many standard deviations the spread currently sits from its own mean.

$z_t = \frac{S_t - \mu}{\sigma}$

where $\mu$ is the spread’s mean and $\sigma$ its standard deviation (estimated over a rolling lookback). Now every pair speaks the same language: $z = +2$ means “two sigma rich,” $z = -2$ means “two sigma cheap,” regardless of the asset’s price scale. The classic band strategy:

• Enter SHORT-the-spread when $z \ge +2$ (the spread is unusually rich → short $A$, long $\beta$ units of $B$).
• Enter LONG-the-spread when $z \le -2$ (unusually cheap → long $A$, short $\beta$ units of $B$).
• Exit when $z$ returns near 0 (the spread is back at equilibrium — collect).
• Optional stop at $z = \pm 3$: if it blows through three sigma instead of reverting, the cointegration may have broken — cut the trade rather than average down.

This is the centerpiece. Drag the entry band and resimulate to feel the trade-off between trading often (tight band, more noise) and waiting for conviction (wide band, rarer signals):

Worked example. Your KO–PEP spread has mean $\mu = 4.0$ and standard deviation $\sigma = 1.2$. Today the live spread is $S_t = 6.7$. Then

$z_t = \frac{6.7 - 4.0}{1.2} = \frac{2.7}{1.2} = 2.25.$

Since $z = 2.25 \ge +2$, the spread is rich: enter short-the-spread — short KO, long 0.80 units of PEP per KO share. You’ll exit when the spread falls back toward $4.0 ($z \approx 0$). If instead it kept climbing to $z = 3.1$, your $\pm 3$ stop would fire and you’d close at a loss, suspecting the rubber band has snapped.

Half-life of mean reversion

A stationary spread reverts — but how fast? A spread that closes its gap in two days is a great trade; one that takes eight months to revert ties up capital, racks up financing costs, and gives the relationship ample time to break before you ever collect. The speed of reversion is captured by the half-life.

The model. Model the spread as an Ornstein–Uhlenbeck process (the continuous-time cousin of an AR(1) autoregression): each period the spread is pulled back toward its mean by a fraction, plus a random shock. In discrete AR(1) form:

$S_t - \mu = \phi\,(S_{t-1} - \mu) + \epsilon_t,$

where $\phi$ (between 0 and 1) is how much of yesterday’s deviation survives into today. A $\phi$ near 1 means sluggish reversion (most of the gap persists); a small $\phi$ means snappy reversion. The continuous mean-reversion speed is $\kappa = -\ln\phi$, and the half-life — the time for the spread to close half its distance to the mean — is

$t_{1/2} = \frac{\ln 2}{\kappa} = \frac{\ln 2}{-\ln \phi}.$

Worked example. You fit AR(1) to the daily spread and get $\phi = 0.95$. Then:

$\kappa = -\ln(0.95) = 0.0513, \qquad t_{1/2} = \frac{\ln 2}{0.0513} = \frac{0.6931}{0.0513} \approx 13.5 \ \text{days}.$

So a typical dislocation takes about 13.5 trading days to close half its gap — call it roughly three weeks for a full round trip. That sets expectations: hold for days-to-weeks, not minutes; budget financing/borrow costs over that window; and if a trade is still open well past a couple of half-lives, treat it as a warning, not a bargain.

When to use it

Use the half-life to size your horizon and your costs, not as a go/no-go gate by itself. Match your data frequency and stop-out time to it (a 13-day half-life wants daily bars and a multi-week patience, not a 5-minute chart). And compare half-life across candidate pairs: given two cointegrated pairs, the shorter half-life recycles capital faster and spends less time exposed to the relationship breaking.

When it breaks

Here’s the uncomfortable truth: cointegration is a statistical observation about the past, not a law of nature. The leash can snap. A merger or acquisition fuses or severs the relationship overnight; a regime shift (new regulation, a technology that disrupts one firm, a commodity shock that hits one leg harder) permanently re-rates one side; one company’s fundamentals genuinely diverge — different growth, different debt, different fate. When that happens, the “spread” you’re trading is no longer a rubber band. It’s just two prices going their separate ways.

The killer failure mode. Your spread hits $z = -3$. Your model screams “historically cheap — load up!” So you double down. But the reason it’s at $-3$ is that the relationship broke — one firm got crushed by news the model can’t see — and the spread keeps right on going to $-4$, $-5$, never reverting. You’ve walked straight into a value trap: averaging into a position precisely because it looks cheap, when “cheap” is actually “permanently repriced.” This is how pairs books blow up. The mean-reversion logic that makes the strategy work is the same logic that destroys you when the mean has moved.

Analogy. The leash didn’t stretch — it broke. You keep pulling, expecting the dog to be yanked back, but there’s no dog on the other end anymore. Every tug just pulls slack.

Risk management — what actually keeps you alive:

• Hard stops (e.g. exit at $z = \pm 3$): cap the loss on any single pair instead of averaging into a broken one. A stop is you admitting “if it hasn’t reverted by here, my reversion thesis is probably wrong.”
• Position limits per pair and across pairs: no single rubber band should be able to sink the book. Diversify across many weakly-related pairs so one snapped leash is a scratch, not a wound.
• Monitor a rolling ADF p-value: re-run the cointegration test on a moving window. If the p-value creeps up (the spread is losing its stationarity), that’s an early warning the relationship is decaying — derisk before it’s a disaster.
• Watch for corporate events: flag mergers, spin-offs, index changes, earnings on either leg. These are the usual suspects behind a sudden break.

Putting it together

Pairs trading is the original statistical arbitrage: find two assets tethered by a shared economic force, go long the cheap leg and short the rich one, and harvest the gap as it closes — market-neutral by construction. The non-obvious key is that correlation isn’t the tether — correlation is about returns co-moving, with no anchor — whereas cointegration gives you a stationary spread with a fixed mean it reverts to. Estimate the hedge ratio $\beta$ (Engle–Granger: regress, then ADF-test the residual), standardize the spread to a z-score, trade the ±2σ bands, read the half-life ($\ln 2 / \kappa$) to size your horizon and costs — and never forget the leash can break, so keep a stop.

You now know how to find a tradable spread and when to bet on it reverting. But a sharper question lurks underneath: how do you even know, in general, whether a market wants to revert (snap back, like our spread) or trend (keep going, where doubling down is correct and fading is suicide)? Pairs trading is a pure mean-reversion bet — its mortal enemy is a regime where things keep diverging. The next lesson, Mean Reversion vs Momentum, zooms out from the single spread to the two opposing forces that govern all systematic strategies, and how to tell which one you’re standing in.