Pairs Trading: Betting on the Gap
Pairs trading means holding a long position in one instrument and a short position in a related one, sized so broad market direction roughly cancels. The trade is a bet on the relationship between the two, not on either price, and it profits only if the gap between them closes.
How it works
Long one instrument, short a related one, sized so market direction roughly cancels. What remains is a position in the spread, not in either price.
The trade wins when the gap closes and loses when it widens. Which leg rises does not matter, provided the long one beats the short one.
It removes broad market direction and keeps the comparison, which is why it can make money in a falling market and lose in a rising one.
It is a form of statistical arbitrage, a cousin of arbitrage proper — that one closes a difference that must close; this bets on one that usually has.
The short leg makes it a short selling position, and the hedging logic removes one risk by accepting another.
Choosing the pair
The pair needs a reason, not just a correlation. A high historical correlation is an observation, and across enough candidates some will show one by chance.
The reason has to be structural: the same inputs, the same customers, the same regulatory regime, naming why the two prices should track at all.
And the sizes must be matched by exposure, not by shares. Equal share counts are no hedge when one leg costs several times the other; match by beta.
A gap that should close can widen for a long time. No level exists at which a spread must stop moving, and nothing brings the prices back on any schedule.
And sometimes it widened because one of them changed. A gap that opens for a reason is not a gap that closes, and nothing in the price series tells you which kind you are looking at.
In practice
The short leg costs money every day you hold it, so time is a cost even when nothing moves.
Two legs, two round trips, twice the cost. One idea pays the entry cost twice and the exit cost twice.
And the thinner leg sets the position size. Whatever volume the less liquid instrument trades is the ceiling for both.
Convergence is measured in weeks, not sessions, which exposes the trade to borrow, to corporate events, and to a relationship that quietly ended.
An earnings report hits one leg and not the other. Dividends and index changes do the same, and an opening gap is the shape it takes.
So the stop loss belongs on the spread, not on either leg. Express it in units of the spread’s own standard deviation, and set it before entry.
A stop on one leg converts a hedged position into a directional one, in a market you chose to have no view on.
On this site’s shared history one round trip costs 2% of a median bar’s range, and a pair pays it on both legs.
Checking a candidate pair honestly
Start with the reason, written in a sentence, before any chart is opened. If you cannot say why the two instruments should move together, there is nothing worth testing yet.
Then look at the spread’s own history rather than at the two price lines. The question is whether the difference behaves like something that returns: how far it usually strays, how long it takes to come back, and whether it has ever simply stopped coming back.
Then ask the question almost nobody asks — how many pairs did you look at before this one? If a scanner examined every combination it could reach and handed back the best, the correlation you are admiring is partly a selection effect.
Its evidential value falls with every pair you rejected, which is overfitting in different clothes — the mechanism that flatters a backtesting run, arriving before the backtest has begun.
What pairs trading is not
It is not arbitrage. Nothing forces this gap to close.
It is not market-neutral because the legs are the same size. Neutrality is an estimate, and estimates drift.
It is not protective hedging. Both legs are opinions, and both can be wrong at once.
And it is not a gentler version of a directional trade. It swaps market risk for relationship risk.
When it fails
In a quiet range the pair just sits there costing. Nothing converges because nothing diverged, while borrow and commissions accrue against a spread doing what it should.
The second failure is the divergence that never returns. Mean reversion is the assumption under the whole structure, and when it is false the position has no natural exit.
A third is the pair chosen by scanner. Picking the highest correlation from a large set produces coincidences, which dissolve once money is committed.
A fourth is a corporate event. Results, an index reweighting or a dividend lands on one leg alone, and the spread jumps to a level it never comes back from.
A fifth is the short leg becoming expensive or unavailable. A recall closes half the position and hands back a directional trade nobody chose.
And a sixth is drift in the hedge ratio. Beta is estimated from the past; if it has moved and you have not re-measured, the position stopped being neutral some time ago.
The original data
Across the 31,760 videos in research/search-study-corpus.jsonl, four carry “pairs trading” in the
title and nine carry “correlation”. The four pairs-trading videos hold a median of 14,568 views
from three channels and a maximum of 98,599; the nine correlation videos manage a median of 1,162
from five channels. Statistical arbitrage appears once, at 13 views; arbitrage three times, at a
median of 2,143; hedging ten times, at a median of 9,853 across seven channels. The strategy is
better covered than the concept it is entirely built on, which is the wrong way round, and the
counts in research/broker-coverage.json explain why most published pairs are picked by scanning for
a high correlation rather than by naming a reason.
One round trip on this site’s shared 576-bar history costs 0.0098 price units — 2% of a median
bar’s range, 45% of the smallest bar, and more than 10% of a bar’s range on 15 of the 576 bars.
Measured by site/measure_series.py into research/series-measurements.json. A pair has two legs,
so one idea pays that twice going in and twice again coming out. Write down the reason the two
instruments are linked before you open a single chart; if the only reason you can write is a
correlation number, you do not have a pair.
Related
Correlation is the measurement every pair rests on, and why the number is weaker evidence than it looks. Mean reversion is the assumption that the spread comes back, tested on this site’s own series. And statistical arbitrage is the wider family, where the same logic runs across many pairs.
I held one of these far longer than I should have, on a pair I was certain moved together. The gap kept opening, and every week I told myself it had to snap back, while the short leg quietly cost me to carry. What had actually happened was that one of the businesses had changed and I was arguing with the news. I closed it for a reason I should have written down before I ever opened it.
— Michael Whitman
This page is educational, not financial advice. Test every idea on your own charts before risking money.