Statistical Arbitrage: How Pairs Trading and Mean Reversion Work

In the week of August 6, 2007, many of the most sophisticated quantitative equity hedge funds in the world lost money at the same time, in a market where the S&P 500 barely registered a crisis. Amir Khandani and Andrew Lo later traced the damage, in NBER Working Paper 14465, to an unwind of similar long/short portfolios that had started in July. That episode, the “quant quake,” is the best short lesson in statistical arbitrage: a strategy built on many small, independent-looking bets can turn into one large bet when everyone runs the same model. If the idea of a long book and a short book is new, start with long/short equity.

What is statistical arbitrage?

Statistical arbitrage is a market-neutral strategy that uses a model to buy securities that look cheap relative to related securities and short the ones that look expensive, betting that the relative mispricing will close.

It is “arbitrage” only loosely. A true arbitrage locks in a riskless profit, such as buying and selling the same asset in two places at different prices. Stat arb relies on probabilities: on average, a gap of this size between these securities has closed, so a large number of such bets should earn a positive return. Any one bet can lose. The edge comes from running hundreds or thousands of them at once, sized so that market, sector and factor exposures cancel, leaving only the relative moves.

Three ingredients define the strategy: a way to decide what “related” means (a pair, an industry basket, or statistical factors), a measure of how stretched the relationship is (usually a z-score), and rules for entry, exit and when to give up. Everything else is execution: costs, borrow, speed and risk limits.

How does pairs trading work?

Pairs trading tracks the price spread between two closely related stocks and trades it when the spread moves unusually far from its average: short the stock that has become expensive, buy the one that has become cheap, and close both when the gap returns to normal.

Gatev, Goetzmann and Rouwenhorst’s study of pairs trading tested the simplest possible version. They matched each stock with the partner whose normalized price had tracked it most closely over a year, then traded pairs over the following six months whenever the prices diverged by more than two historical standard deviations. Using daily data from 1962 to 1997, self-financing portfolios of the top pairs earned average annualized excess returns of up to 12%, and the profits exceeded a conservative estimate of transaction costs for most of the period. The authors noted that part of the profit may have come from market microstructure effects, such as bid-ask bounce.

A pair is market neutral by construction. If both stocks fall 10% together, the long loses what the short gains. What you are betting on is the relationship, which is why gross exposure, not net, is the measure of risk in a stat arb book.

What is cointegration, and why does it matter more than correlation?

Two prices are cointegrated when some weighted combination of them, the spread, is stationary: it moves around but keeps returning to a fixed mean. That anchoring is what a pair trade needs. Correlation alone doesn’t provide it.

Correlation measures whether two stocks’ daily returns move in the same direction. Two stocks can be highly correlated day to day and still drift apart for years, because one grows faster. Their price gap never comes back, and a pair trade on them loses steadily. Cointegration tests the price levels instead. The usual approach is to regress one stock’s log price on the other’s to get a hedge ratio, then test whether the residual, the spread, is stationary. If it is, the spread has a mean, a typical size of deviation and a speed of return, measured by its half-life.

The last row matters most. Two oil refiners or two railroads share inputs, customers and regulation, so a statistical relationship between them has a reason to persist. Two unrelated stocks that happened to track each other for a year usually don’t.

How do z-score entry and exit rules work?

The z-score is the spread’s distance from its mean measured in standard deviations. A common rule set is: enter when the z-score passes +2 or −2, exit when it returns to about zero, and stop out if it keeps widening past about 4.

Chart 1 runs those rules on a simulated spread built to mean-revert, with a half-life of about 8.7 trading days. Over 500 days the z-score crossed the ±2 bands five times. Each time the rules sold the spread when it was rich or bought it when it was cheap, and closed when it came back to zero. All five round trips made money after an allowance for costs, the average holding period was 20.6 days, and the book was in a trade on about 21% of days. The largest z-score in the sample was 3.21, so the ±4 stop was never hit.

That result is the best case on purpose. The simulated spread is mean-reverting by construction, and its mean and volatility are known exactly. Real traders estimate both from history, and the estimates shift. The two most common ways real pairs lose money are both missing from this chart: the relationship breaks (a takeover, a lawsuit, a new product), so the spread never comes back; or the spread keeps widening for weeks before it reverts, and the trader is stopped out or forced out near the worst point.

What does a stat arb pair trade look like step by step?

Here is the first simulated trade from Chart 1, translated into dollars for an illustrative pair of related stocks, A and B, with $1 million on each leg.

1. Estimate the relationship. Regress A’s log price on B’s over the past year: hedge ratio 0.9, spread standard deviation 3%.
2. Wait for the stretch. On day 165 the spread is 2.49 standard deviations rich: A is about 7.5% expensive relative to B.
3. Enter. Short $1 million of A and buy $0.9 million of B, matching the hedge ratio, so the market exposure roughly cancels.
4. Set the exits. Take profit when the z-score returns to 0; stop out at +4 (a further 4.5% against you).
5. Hold. Twenty trading days later the z-score is −0.05. Close both legs.
6. Count the result. The spread closed 2.54 standard deviations, about 7.6%, on $1 million: roughly $76,000, minus about $3,000 of trading costs, borrow fees and financing in this example.

Now the losing version. If A’s premium had been the market pricing in a takeover bid, the spread would have jumped to +6 instead, the stop at +4 would have filled near +6 after the overnight gap, and the trade would have lost about 10.5%, or $105,000: more than the gain on the winning trade. That asymmetry, many small wins and occasional large losses, is the shape of every mean-reversion strategy, and it is why stop rules and position sizes matter more than the entry signal. Our guide to fat tails explains why the large losses come more often than a normal distribution suggests.

How do stat arb hedge funds run the strategy at scale?

Stat arb hedge funds run thousands of small positions at once, built from statistical factors rather than single pairs, with market, sector and style exposures neutralized and turnover measured in days.

Instead of finding pairs one at a time, a quant fund models each stock’s return as exposure to a set of common factors (the market, sectors, size, value, momentum) plus a residual. The residual is treated as a mean-reverting spread, exactly like a pair’s, and each stock is bought or sold depending on how stretched it is. Avellaneda and Lee’s study of US equities built signals this way, using principal component analysis (PCA) and sector ETFs as the factors.

Their results show both the appeal and the decay. PCA-based strategies had an average annual Sharpe ratio of 1.44 over 1997 to 2007 after transaction costs, but only about 0.9 over 2003 to 2007. ETF-based signals earned about 1.1 over the full period, and adding trading-volume information lifted the ETF version to 1.51 over 2003 to 2007. These are published backtests, not live fund returns, but the direction matches what practitioners describe: the simple signals got crowded, and the edge moved to faster data and more inputs.

Because the residual moves are small, stat arb books run high gross exposure at close to zero net. In the SEC Form PF data compiled by the Office of Financial Research, equity-strategy hedge funds averaged about 2.7 times gross assets to net assets in the second quarter of 2026 and relative-value funds about 23 times, so modest losses per position can become large losses on capital. Many stat arb teams now sit inside multi-strategy platforms, which add strict drawdown limits on top.

What happened to stat arb funds in August 2007?

In the week of August 6, 2007, quantitative long/short equity funds suffered unusually large losses as similar portfolios were unwound at the same time, according to Khandani and Lo’s analysis.

Using simulated returns from five valuation factors, Khandani and Lo found that the unwinding began in July 2007 and continued to year end, with a first unwind on August 1 and a larger one on August 6 that started in financial-sector stocks. Transactions data showed that a simple market-making strategy, which buys losers and sells winners in the short run, lost heavily that week while profiting before and after it, which they linked to a temporary withdrawal of market-making capital beginning August 8. Their conclusion was that the episode reflected both portfolio deleveraging and a pullback of liquidity, with implications for systemic risk in the hedge fund industry.

The mechanism is general. When many funds hold the same longs and shorts, one fund’s forced selling pushes the spreads the others are betting on wider, which triggers their own risk limits. Daniel and Moskowitz document a related pattern in momentum crashes, which cluster in panic periods after market declines. Our guide to hedge fund crowded trades follows the loop from the first loss to the forced sale.

Can individual investors use statistical arbitrage?

Individuals can trade a few pairs in a margin account, but a handful of pairs is a concentrated bet on a few relationships, not a statistical edge, and costs fall harder on small accounts.

The practical frictions add up. Every pair needs a short leg, which means a locate under Regulation SHO, a borrow fee, dividends owed and the risk of a recall. Holding periods measured in weeks generate short-term gains taxed at ordinary rates. And a fund’s statistical edge depends on running hundreds of independent bets so that the law of large numbers works; with five pairs, one broken relationship decides the year.

What does transfer is the discipline. Before any trade that depends on something reverting, write down what “stretched” means in numbers, where you take profit, and the level that proves the relationship has broken. The same thinking applies to selling options after a large move: our guide to implied volatility rank covers how to judge when option prices are stretched relative to their own history.

What does stat arb thinking mean for a concentrated stock position?

For an executive or tech employee with a large single-stock position, the useful idea is relative value: measure your stock against its peers and the index, not just against its own price.

If your company’s stock has run far ahead of its sector, a stat arb desk would see a stretched spread and bet on it narrowing. You probably can’t short a peer basket against restricted stock, and you may not want to. But you can act on the same read in ways that suit your position: sell covered calls when the stock is stretched and implied volatility is high, add a put spread or collar when the gap to peers is extreme, or speed up a planned diversification. Each of those cuts exposure to the relative move without a borrow, recall or squeeze risk.

Proflex uses this lens with concentrated holders: compare the position with its peers and the index, judge whether the move is stretched or a real change in the business, and choose an options structure with a known cost for either case.

What is a checklist for a mean-reversion or pairs trade?

1. Name the economic reason the two securities should stay related.
2. Test cointegration on prices, not just correlation of returns, over more than one period.
3. Estimate the spread’s half-life and make sure it is shorter than your holding horizon.
4. Set entry (for example ±2), exit (about 0) and stop (for example ±4) before trading.
5. Check borrow availability, fee and dividends on the short leg.
6. Size so that a gap through your stop still costs an amount you can accept.
7. Remove any pair that hits its stop; don’t re-enter at a “better” z-score.
8. Cut size when many spreads widen at once; that is crowding, not opportunity.

Frequently asked questions

Key takeaways

1. Statistical arbitrage bets that stretched price relationships between related securities will revert.
2. Pairs trading is the simplest form: trade the spread when its z-score passes ±2, exit near zero, stop if it keeps widening.
3. Cointegration, not correlation, is what makes a spread mean-revert.
4. Published returns were strong but have decayed: up to 12% a year for pairs in 1962–1997; PCA stat arb Sharpe fell from 1.44 to 0.9.
5. Stat arb runs high gross at near-zero net, so crowding and leverage are its main risks, as the August 2007 quant meltdown showed.
6. Small, frequent wins and rare, clustered losses are the normal shape of the strategy.
7. For individuals, the useful lesson is the discipline: define the stretch, the exit and the break point before you trade.

Sources and method

1. Gatev, Goetzmann & Rouwenhorst, Pairs Trading: Performance of a Relative Value Arbitrage Rule (NBER w7032) — pairs trading returns, 1962–1997
2. Khandani & Lo, What Happened to the Quants in August 2007? (NBER w14465) — the August 2007 quant unwind
3. Avellaneda & Lee, Statistical Arbitrage in the U.S. Equities Market (NYU Courant) — PCA and ETF stat arb Sharpe ratios
4. Office of Financial Research, Hedge Fund Monitor (SEC Form PF) — relative-value and equity hedge fund leverage, Q2 2026
5. SEC, Key Points About Regulation SHO — short sale locate and close-out rules
6. Daniel & Moskowitz, Momentum Crashes (NBER w20439) — how crowded factor strategies crash in panic states

Chart 1 and the simulation statistics are a Proflex simulation, not a backtest of real stocks: a seeded (20261029) AR(1)/Ornstein–Uhlenbeck spread with daily mean-reversion speed 0.08 (half-life about 8.7 days) over 500 trading days, converted to a z-score with its stationary standard deviation. Rules: sell the spread at +2, buy at −2, exit at 0, stop at ±4, cost 0.1 standard deviations per round trip. The spread parameters are known in advance, which real traders never have. Chart 2 restates published Sharpe ratios from Avellaneda and Lee. Chart 3 dates come from Khandani and Lo. Leverage figures are OFR Form PF series for June 30, 2026. The worked pair example is illustrative.