Twins That Drift Apart: The Academic Test of Pairs Trading

Pairs trading is one of those strategies that everyone on a trading floor had heard of and nobody had ever properly tested. The story goes that a quant group at Morgan Stanley in the 1980s built systems to spot pairs of stocks that normally moved together, and to bet on them converging whenever they drifted apart. It made money. It became legend. It stayed folklore.

Evan Gatev, William Goetzmann and Geert Rouwenhorst decided to check the legend. They ran the simplest possible version of the strategy on forty years of US stock data and reported, without embellishment, what happened.

The problem: a famous strategy nobody had audited

Wall Street has no shortage of confident claims. "This works" is easy to say and expensive to verify. Pairs trading in particular suffered from a specific credibility issue: the people who claimed it worked were the people running it, and they were not showing their books.

The academic question is sharper than "does it make money." It is: does a simple, mechanical, publicly-describable pairs rule generate returns that cannot be explained by known risks? If yes, that is a challenge to market efficiency, and it tells you something real about how prices behave. If no, then pairs trading is either a myth or it depends entirely on secret sauce that cannot be tested.

The trap in testing it is obvious to anyone who has ever backtested anything. There are thousands of stocks, so there are millions of possible pairs. If you go hunting through history for the pairs that happened to converge nicely, you will find plenty, and you will have discovered nothing but your own ability to sift noise. The test has to be designed so that the strategy is choosing pairs using only information available at the time, and then trading them going forward.

The key idea via analogy: two runners who normally stay in step

Picture two runners who have trained together for years and always run shoulder to shoulder. One day, for no obvious reason, one of them gets twenty metres ahead. You do not know which runner is going faster or slower, and you do not care. You simply bet that the gap will close.

That is a pairs trade. Two stocks that historically move together drift apart. You buy the one that fell behind, you short the one that ran ahead, and you wait for them to reconverge. You do not need a view on the market, on the sector, or on the economy. You are betting on the gap, not the direction. If the whole market crashes, both legs crash, and your position is roughly unaffected. That is what makes the trade market-neutral, and it is why the strategy is attractive to anyone who wants a return stream that does not just ride the index.

The authors' implementation is almost aggressively simple, which is the point. They split time into two phases:

• A formation period. Look back over the past year of daily prices. Normalise each stock's price path so you are comparing shapes rather than levels. For every stock, find the partner whose price path was closest to it, using the plainest possible measure: the sum of squared differences between the two normalised paths. That is it. No cointegration test, no factor model, no industry classification, no machine learning. Just: which two lines were closest.
• A trading period. Over the following six months, watch the matched pairs. If a pair's two normalised prices diverge by more than two standard deviations of their historical spread, open the trade: long the laggard, short the leader. Close it when the prices cross again, or when the six months run out.

Run this on the whole US market, month after month, from 1962 to 2002. Build a self-financing portfolio of the top pairs.

It made money. The best-performing versions produced average annualised excess returns of up to roughly eleven percent, before the usual caveats. More importantly for the academic question, the returns were not explained away by exposure to the market, to size, to value or to momentum. The profits survived reasonable transaction cost assumptions. And the strategy's returns had very low correlation with the market, which is exactly what a market-neutral strategy is supposed to deliver.

They also ran a crucial sanity check. If the profits came from genuinely finding related companies, then randomly matched pairs should earn nothing. They tested that, and randomly-formed pairs did not produce the same profits. The matching was doing real work.

Why it mattered

• It made a piece of trading floor folklore into a documented anomaly. After this paper, "pairs trading works" was no longer a claim, it was a finding in a top journal, replicable by anyone with a price database.
• It established the template for statistical arbitrage research. The formation-period-then-trading-period design, with strict avoidance of look-ahead bias, became the standard experimental setup that later stat-arb papers copied and refined.
• It showed that relative-value returns are a distinct thing. The profits did not load on the known factors. That means pairs trading was not a repackaged value or momentum bet. It was capturing something else: short-horizon relative mispricing between close substitutes, and the compensation for standing ready to correct it.
• It quietly described a liquidity-provision business. The authors noted that the pairs trader is systematically buying whatever just fell and selling whatever just rose. That is the behaviour of a market maker. A large part of what the strategy earns is plausibly the reward for supplying liquidity to investors who wanted out in a hurry, which is a much more economically comfortable explanation than "the market is dumb."

The honest limitations

• The profits shrank over time. The paper itself shows the returns decaying in the later part of the sample. Follow-up work has found that after the early 2000s, with the rise of decimalisation, faster execution and a crowd of quant funds running the same idea, the plain-vanilla version of this strategy earns very little. Publishing an anomaly is a good way to kill it.
• It is short volatility in disguise. The trade profits when things converge and loses when they diverge. Most of the time, things converge. Occasionally a pair diverges because one company is genuinely broken (a fraud, a failed drug trial, a takeover of only one of the two), and then the "gap will close" thesis is simply wrong, and it does not close, and the loss on that pair can wipe out many small wins. The return distribution has a long left tail. Look at the Sharpe ratio alone and you will not see it.
• Shorting is not free and not always possible. The strategy requires shorting the outperformer. In practice that means borrow costs, recall risk, and the possibility that exactly the stock you most want to short is the hardest and most expensive to borrow.
• The matching rule is crude by design. "Closest historical price path" is a deliberately naive way to find economically related companies. It will happily pair two firms that briefly moved together by accident. More careful methods exist, but the paper's willingness to use the crude version is what makes the result convincing rather than data-mined.
• Crowding is the killer. When many funds hold the same convergence bets with leverage, a shock that forces one of them to unwind pushes the spreads further apart, hurting everyone else and triggering more unwinding. That is not hypothetical. It is exactly what happened to quant equity funds in August 2007.

The one-line takeaway

Gatev, Goetzmann and Rouwenhorst took the oldest piece of quant folklore on Wall Street, tested it with a rule so simple it could be described in a paragraph, and found that betting on the reconvergence of stocks that normally move together earned real, market-neutral profits for decades, most likely as payment for providing liquidity to people in a hurry.