Pairs Trading — The Market-Neutral Strategy on Correlated Assets

Most strategies require answering the market's hardest question: will it go up or down? Pairs trading sidesteps that question. You take two assets that have historically moved together — two stocks in the same sector, two coins from the same category — and trade not the direction, but the gap between them: buy the one that's lagged behind, short the one that's run ahead, and wait for the relationship to revert to normal. The market can rise, fall or go nowhere in the meantime — in theory, it doesn't matter.

The strategy traces back to the 1980s, to the desks of Morgan Stanley's quants, and has earned a rare honor in trading: solid academic research over decades of data. The results are interesting — and instructive about the fading of market edges. This article covers the mechanics, the numbers from the research (with the caveats they deserve), and the places where a market-neutral strategy can be very un-neutral to your account.

What Pairs Trading Is

The foundation of the strategy is the spread — the difference (or ratio) between the prices of two related assets. If the assets share a real common denominator (sector, business model, shared capital flows), their price relationship oscillates around some norm. Deviations from that norm happen constantly — for liquidity, news, or fund-flow reasons — and those deviations are the opportunity:

1. Pair selection. You look for strongly related assets. A standard preliminary filter: return correlation of at least 0.80 over a historical window. But correlation isn't enough — the key test is cointegration (e.g. Engle-Granger), confirming that the spread is stationary: it has a stable mean it reverts to, rather than drifting off indefinitely.
2. Spread normalization. The spread is converted to a z-score: how many standard deviations the current value is from the mean of the window (e.g. 60 days). Z-score of 0 = normal; ±2 = a rare deviation.
3. Entering on the gap. The z-score crosses a threshold (classically ±2): short the relatively expensive asset, long the relatively cheap one, with the legs sized so their values (or betas) balance out. Net exposure to the market: roughly zero.
4. Exiting on the reversion. The spread reverts to the mean (z-score ~0) — you close both legs. You profit from the gap closing regardless of whether the market rose or fell in the meantime.

This is a classic strategy from the mean reversion family — except you're not expecting reversion from a price (which can trend for years), but from a relationship between two prices, which is statistically much more stable.

It's worth immediately distinguishing two ways of balancing the legs. The dollar-neutral variant matches notionals ($5,000 per leg) — simple, but blind to the fact that one asset might be twice as volatile as the other. The beta-neutral variant scales the legs inversely to their volatility or market beta, so both react to the shared factor with similar force. On pairs with similar volatility the difference is cosmetic; on mismatched pairs, it decides whether the position is actually neutral or just looks that way on the spreadsheet.

What the Numbers Say

The most-cited study is the paper by Gatev, Goetzmann and Rouwenhorst (Pairs Trading: Performance of a Relative-Value Arbitrage Rule), analyzing a simple, mechanical version of the strategy on US stocks over four decades (1962–2002). Result: self-financing pairs portfolios (long one leg, short the other — net capital ~zero) generated up to about 11% of annualized excess return, with volatility clearly lower than the market and low correlation to it.

Caveats without which that number is marketing, not information:

• This is a historical result from a specific era. The authors themselves noted that the strategy's profitability declined in the later years of the sample — as statistical arbitrage became more widespread and hedge-fund capital crowded the space. Follow-up research after 2002 points to further erosion of the simple variants. A published edge is an exploited edge.
• Gross vs net. The excess return is calculated before the real costs a retail trader faces: commissions on two legs ×2 (entry and exit), the cost of borrowing stock for the short (borrow fee), and slippage. With hundreds of trades a year, costs can eat half the theoretical result or more.
• Methodology matters. The tested variant (pairs picked by minimum distance of normalized prices, entry at 2 standard deviations, rigid formation and trading windows) is one specific machine. "Pairs trading" at your broker with an arbitrarily chosen pair is a different machine — the study's results don't automatically transfer to it.

A second important finding is qualitative: in the studies, the strategy's profit did not come from a simple reflection of known risk factors — suggesting it rewarded a genuine service: providing liquidity and correcting relative mispricing. That's good news (the profit source is real) and bad news at the same time (algorithms with near-zero transaction costs now compete for that service).

How It Works Step by Step (a Numeric Example)

An illustrative example on a crypto pair from the same sector, numbers simplified:

1. Selection and testing. Candidates: two large layer-1 coins, A and B. Correlation of daily returns over the past year: 0.86 — preliminary filter passed. Engle-Granger cointegration test on (log) prices: statistically significant relationship, stationary spread. The pair makes the list.
2. Spread parameters. You define the spread as the price ratio A/B. 60-day mean: 2.50; standard deviation: 0.10. Thresholds: entry at z-score ±2.0, exit at 0, emergency stop at ±3.5 — because a spread that has run that far off usually signals a regime change more often than an opportunity, statistically speaking.
3. Entry. After a strong week for A, the ratio rises to 2.72: z-score +2.2. You open: short A for $5,000, long B for $5,000 (legs equal by notional). Exposure to market direction: ~zero. You're risking the relationship, not the market.
4. The reversion scenario. Over two weeks the whole market drops 10%, but A drops harder and the ratio reverts to 2.52 (z-score ~0). You close both legs. The short leg earned more than the long leg lost: result ~+7% on the spread, i.e. roughly $350 on $10,000 of notional engaged — despite the falling market. From that you subtract 4 commissions and two weeks of funding cost on the short leg.
5. The breakdown scenario (equally important). Alternatively: B's fundamentals deteriorate (an exploit, developer exodus), and instead of reverting the spread runs to a z-score of 3.5. The emergency stop closes the position at a loss of roughly 4–5% of notional. Without it, waiting for a "reversion to the mean" that has stopped existing has no floor.

You size the notional of both legs relative to your portfolio the same way you always do, with position-sizing rules — with the adjustment that a "neutral" position can still lose on both legs at once when the relationship breaks down.

Risks — When the Strategy Does NOT Work

• Cointegration breakdown. Risk number one. The relationship between two assets holds only as long as the shared fundamental holds. A merger, a business-model change, a protocol exploit, a delisting — and the spread drifts off for good. A mean-reversion strategy with no hard stop on the spread then turns into a double loss: on the long leg and on the growing short.
• Correlation mistaken for cointegration. The most common workshop mistake: a pair picked on return correlation alone. In a crypto bull market, everything is correlated with everything through a shared speculative factor — that doesn't mean any given pair has a stable price relationship. Without a formal cointegration test, you're trading a dependency that might just be an artifact of the sample period.
• Parameter overfitting. The averaging window, entry thresholds, pair selection — all of it can be "tuned" to history until the backtest shines. Honest testing standards (out-of-sample data, transaction costs in the simulation, no look-ahead) apply doubly here, because the strategy inherently generates a lot of low-margin trades.
• Cost of the short leg. In stocks: borrow fees and the risk of a recall at the worst possible moment. In crypto: the funding rate on the perpetual — when it's positive, your short side of the spread pays it day after day, quietly eating the theoretical profit.
• A crowded field. Statistical arbitrage is one of the most heavily exploited quant strategies in the world. Simple variants on liquid stock pairs are today run by machines with costs retail can't match; the more realistic niches are markets and pairs off the radar of the big players — with their own liquidity risks.
• Neutrality can be an illusion. Equal notionals don't mean equal risk: if one leg has twice the volatility (beta) of the other, the position has hidden directional exposure. In market stress, correlations and betas spike — a "neutral" position can behave like plain leverage in a crash.

The no-hype verdict: pairs trading is a rare case of a strategy with solid academic documentation and a quantifiable profit mechanism — a historical ~11% of annualized excess return on self-financing portfolios is a result most internet "systems" have never come close to. But the same literature honestly shows what happened next: the edge faded as it got exploited, and the retail version pays costs the studies never counted. If you trade pairs: a cointegration test instead of correlation alone, a stop on the spread instead of faith in reversion, costs in the backtest instead of gross results. Market neutrality doesn't mean neutrality toward workshop mistakes — those, a double position punishes twice as hard.

FAQ

Rafał — Krypto Bez Ściemy

Trader and founder of Strefa Tradingu. He’s been breaking crypto down on YouTube for years — no hype, no signals, with a focus on market structure and risk management.

kryptobezsciemy.plInstagramYouTubeKlub Zero