Week 14: Pair Trading — Relative Value, Mean Reversion, and the Simplest Equity-Hedge-Fund Strategy

---

1. Why This Is Important

Week 13 introduced long/short — the architecture. This week applies that architecture to its narrowest, most teachable form: **two stocks, one long, one short, equal dollar, sized to the spread between them.** Pair trading. The simplest equity-hedge-fund strategy in the book, and the cleanest place for a retail investor to learn the mechanics of relative value before scaling up to anything more complicated.

You need this lesson for four reasons.

**It is the cleanest test of the long/short toolkit you just

learned.** Long Coke / short Pepsi at equal dollar is dollar- neutral by construction, roughly beta-neutral by virtue of the two stocks having near-identical sensitivities to the consumer- staples factor, and the entire P&L is the spread alpha — the difference between two close substitutes. There is nowhere for a directional view to hide. If the trade works, you traded the spread; if it doesn't, you didn't. Pair trading is the laboratory in which you find out whether you have any relative-value edge at all.

**It is the entry point to a real, durable alpha source — not

the made-up kind.** The structural alphas that actually compound after costs are liquidity, sector rotation, long-term trends, and buying what the passive machinery has abandoned. Pair trading sits inside the structural-alpha class because the pair trader is *supplying liquidity to the temporarily dislocated leg*. When PEP gaps down on a weak quarter and KO sits still, the pair trader who buys the dislocation is on the other side of forced sellers (mutual-fund redemptions, factor unwinds, index rebalances). The reversion of the spread back to its long-run relationship is the structural payment for that liquidity.

**It teaches you why "real-looking" pairs decay — the regime

change problem.** Regime change is the recurring bell of this course: regimes change, often without warning, and a strategy that worked for a decade can stop working in a quarter. Pair trading is the cleanest case study of that principle in retail-accessible form. Two stocks that have moved together for fifteen years can permanently de-couple — KO and PEP did, partially, after PEP's snack-foods business compounded faster than KO's beverage-only one. A pair trade that ignored that fundamental divergence and kept fading the widening spread back to its 10-year mean would have been wrong for six straight years. **Cointegration is a conditional, regime-dependent property, not a permanent one.**

**It frames the August 2007 quant quake — the canonical case

study for crowded factor risk.** When too many quants run essentially the same pair-trading and stat-arb strategies, the pairs themselves become a factor — a factor that can unwind in correlated panic. Over four trading days in August 2007, a forced unwind by one or two large multi-strat funds drove a simultaneous reversal across the standard pair-trading factor, and most of the field lost two-to-four sigma on what should have been independent trades. The lesson — your alpha is only alpha when the people running the same strategy aren't all forced to exit at the same time — is one of the most important risk- management lessons in modern finance, and pair trading is where it was first learned.

This lesson sits in the opportunistic slot of the four-tranche structure and on the alpha end of the barbell. Most of your wealth is in the safety end (cash, Treasuries, gold, deep-ITM long-dated calls on names you'd own anyway). A small sleeve at the alpha end is where the pair-trading book lives — the pinwheel, in Horace's phrase, that spins when the rest of the portfolio is sitting still.

---

2. What You Need to Know

2.1 Relative Value — the Core Reframe

Long-only investing asks an absolute question: "Is KO cheap?" The answer requires a view on KO's earnings, multiple, growth, balance sheet, and the broader equity-risk premium. Get any of those wrong and the trade can be wrong even if KO is "cheap" by your model.

Pair trading replaces the absolute question with a relative one: "Is KO cheap relative to PEP?" The two beverage incumbents share a customer base, a regulatory regime, an exposure to the commodity-price cycle in sugar and aluminium, an exposure to the consumer-staples factor, an exposure to USD strength, and roughly the same multiple regime. A bad consumer-staples macro hits both; a sugar-tax scare hits both; a strong-dollar quarter hits both. What it does not hit symmetrically is the company-specific operational variance — KO's bottler restructuring, PEP's snack-foods margin, an FX shock specific to one country, a recall that lands on one brand. The spread isolates that company-specific variance from everything else.

The directional risk you eliminate is large. The 60/40 holder of 2,000 shares of KO has roughly half a million dollars exposed to "the consumer staples factor wakes up tomorrow and falls 10% on a risk-off day." The dollar-neutral pair trader long 1,000 KO and short 1,000 PEP has approximately zero dollars exposed to that factor. What the pair trader has, instead, is a $0 net but $200 gross book that is exposed to **the difference between two stories that look similar**. That difference is what they get paid for — or punished for, when the difference turns out to be permanent rather than mean-reverting.

This is why pair trading is a clean teaching example: it strips off market beta entirely and lets you see, in isolation, whether the relative-value bet has any signal at all.

2.2 Spread Construction and the Z-Score

Once you've picked a pair, you need a way to say how dislocated it is. The standard approach is the **rolling-Z-score on the log-spread**.

Construct it in three steps.

Form the log-spread as s_t = log(P^A_t) - log(P^B_t).

Working in log-space is the convention because it makes the spread invariant to dollar normalisation and gives a multiplicative interpretation: a constant s_t means a constant ratio of prices, regardless of level.

Compute the rolling mean and stdev over a lookback window —

60 trading days is a common starting point for daily data. Call them mu_t and sigma_t.

Z-score the spread: z_t = (s_t - mu_t) / sigma_t. By

construction, in a stable cointegrated relationship, z_t is approximately distributed mean-zero, unit-variance, with most observations falling between -2 and +2.

The image below shows what this looks like for KO and PEP from 2015 through 2024 — the top panel is normalised prices, the bottom panel is the rolling-60-day Z-score with +/-2 sigma bands.

A pair trader looks at the bottom panel and asks: *when the line hits +2, do I trust that it'll come back to zero, and how long will it take?* The answer comes from cointegration, not correlation.

2.3 Correlation vs Cointegration — Why the Distinction Matters

Two random walks can be highly correlated and still diverge to infinity. That is the trap that kills naive pair trades.

• Correlation measures whether two return series move together

over short horizons. Two stocks with monthly correlation of 0.85 just means their monthly returns line up; it says nothing about whether the spread between their levels mean-reverts.

• Cointegration is stronger. Two price series are cointegrated

if there exists a linear combination of them that is stationary — i.e., the spread itself has a stable mean and variance and reverts when it strays. Cointegration is the mathematical formalisation of "they share a long-run equilibrium relationship."

In practice, cointegration is tested with the Engle-Granger two-step procedure or a Johansen test on level series; the details are second-order. The intuition is the part you must internalise: **you can trade pairs only when the spread is mean-reverting, and correlation alone does not guarantee that.** The textbook counter-example is two assets in different secular trends — both might rally together for years (high correlation) while the gap between them widens monotonically (no cointegration). A pair trader who fades that widening gap loses every quarter for a decade.

A practical rule: before you trade a pair, plot the spread Z-score over a 5-to-10-year window and ask "does this thing visibly come home, repeatedly, after each excursion?" If yes, the cointegration test is mostly confirming what your eye already saw. If no, no amount of statistical machinery will rescue the trade.

2.4 Classic Pairs and Why They Work

Four canonical retail-accessible US-listed pairs, with the structural reason each cointegrates:

Pair	Sector	Why the spread mean-reverts
KO / PEP	Consumer staples — beverages	Two near-identical global beverage incumbents; same customer base, same commodity cost structure, same defensive multiple regime. Spread breaks only when one company makes a strategic-mix change (PEP's snack-foods build-out is the classic example).
V / MA	Financials — payments duopoly	Visa and Mastercard are a regulatory duopoly with identical economics: card interchange, network effects, oligopolistic pricing power. Spread breaks only on company-specific litigation or geographic-mix exposure (China, Russia).
XOM / CVX	Energy — integrated supermajors	Both run integrated upstream-downstream books with similar reserves, similar geography, similar capital-allocation discipline. Spread breaks on idiosyncratic operational events — refinery accidents, project-cost overruns.
GLD / SLV	Precious metals ETFs	Gold and silver share a monetary-debasement story, but silver is half industrial, half monetary; the spread (the "gold-silver ratio") cycles between 50 and 90 with cyclical mean-reversion driven by the relative weight of those two demand drivers.

The common feature is that the two legs share a structural driver and differ on a contained idiosyncratic dimension. The trade is paid for the contained dimension (the spread mean-reverts) and protected against the structural one (it nets out).

2.5 Entry/Exit Rules and a Worked Backtest

The textbook rule is **+/-2 sigma entry, 0 sigma exit, with a hard stop at +/-3 to +/-4**.

• When z falls below -2: the spread is two standard deviations

cheap, A is unusually cheap relative to B. Go long A, short B at equal dollar.

• When z rises above +2: the spread is two standard deviations

rich. Go short A, long B at equal dollar.

• When z crosses zero: the spread is back at its rolling mean.

Exit. Do not wait for the opposite tail; the marginal expected return from running another sigma is small and the regime-change risk grows with holding period.

• If z continues past +/-3 or +/-4 in the wrong direction: **cut

the position and re-evaluate the cointegration assumption** — more often than not, the relationship has broken structurally rather than dislocated stochastically.

The chart below shows that ruleset applied to KO/PEP from 2015 through 2024. The top panel is the Z-score with up-arrows on long- spread entries and down-arrows on short-spread entries; the bottom panel is the cumulative log-spread P&L net of a 5 bp round-trip cost per side change.

What the chart teaches: **most pair trades are small wins, with occasional regime breaks that cost more than several wins put together.** This is exactly the shape of a structural-alpha return stream. It is also why pair trading is run as a *book of many pairs* rather than one — the law of large numbers is what makes the small-edge-per-pair model viable. A single-pair retail implementation is more education than alpha; the production version diversifies across thirty to fifty pairs simultaneously.

2.6 The 2007 Quant Quake — Crowding Risk Is Real

In the first week of August 2007, the standard equity-stat-arb book — roughly defined as "long the cheap leg of a pair, short the rich leg, across many pairs simultaneously" — lost between 4 and 12 sigma of its expected daily return for four straight days. The unwind cascaded across most of the major multi-strat hedge funds running similar models: AQR, Renaissance Equity Market-Neutral, Goldman Global Equity Opportunities. Some funds halved in a week.

The post-mortem is now well-documented. One of the largest multi- strats — most of the public reconstruction blames Goldman's quant book, though the precise actor is less important than the mechanism — was forced to deleverage suddenly, likely due to a margin call originating from an unrelated subprime-mortgage book. Closing the equity book required selling its long stat-arb leg (buying back the short side) en masse. Because every other major quant ran the same pairs from the same factor data, those same pairs simultaneously moved adversely against everyone else's book. As losses mounted, more funds were forced to deleverage, which moved the pairs further, which forced more funds to deleverage. **Standard pair-trading factor exposure had become a crowded trade — and crowded trades, when they unwind, do so together.**

The lesson is foundational and is the cleanest illustration in finance of the regime-change principle. **An alpha source remains alpha only as long as the population running it is small enough that no forced exit can simultaneously move all the trades against you.** The day too many people run the same strategy, the strategy itself becomes the risk factor. This is why modern stat arb runs a continuous "crowdedness factor" overlay — measuring how many other funds are likely in the same trade — and dials gross exposure down as crowdedness rises. The retail pair trader inherits the same discipline in miniature: do not over-size a canonical, well-known pair, because every other dollar in that trade is implicitly correlated with yours.

2.7 Where Pair Trading Sits in the Barbell

The barbell is most of your wealth in safety, a small sleeve in genuine asymmetric edge. **Pair trading lives at the edge end.** Sized correctly, a pair-trading sleeve runs at single- digit percentage of total portfolio, in dollar-neutral pairs where the gross exposure is two-to-three times sleeve size and the net is approximately zero. The sleeve's expected contribution to total portfolio return is small (3-8% annualised on the sleeve, which translates to 0.3-0.8% on a 10% sleeve allocation); the expected contribution to total portfolio Sharpe is potentially much larger because the sleeve's returns are roughly orthogonal to the long-only book.

This is the right way to think about pair trading for a retail investor: **not as a way to compound wealth, but as a way to reduce the variance of the wealth path.** It is a Sharpe-ratio play, not a return play. Anyone who promises you a pair-trading strategy with double-digit gross returns and low drawdowns is either over-fitted, levered, or selling a course.

---

3. Common Misconceptions

Misconception 1: "High correlation means a good pair."

Correlation measures co-movement of returns; cointegration measures mean-reversion of levels. High correlation with no cointegration is the canonical wide-and-keep-widening trap.

**Misconception 2: "If a pair has worked for ten years, it'll keep working."**

KO/PEP's relationship has shifted twice in the last twenty years on PEP's snack-foods build. Cointegration is conditional on the underlying business mix staying stable. When the mix changes, the spread re-prices to a new equilibrium, and the trader fading the old equilibrium loses for years. The regime has shifted.

Misconception 3: "Z-score of +/-2 is a hard rule."

The thresholds are calibration choices, not laws. In a low-vol regime, +/-1.5 sigma fires more trades and may be the right cut; in a high-vol regime, +/-2.5 reduces false positives. Calibrate to the pair, the lookback, and the realised vol regime — not to the textbook number.

Misconception 4: "I'll just trade one pair to learn."

A single-pair book is not pair trading; it is one trade with high operational overhead. Real pair-trading strategies run thirty-plus pairs to make the law of large numbers work. The single-pair exercise is education; do not confuse it with alpha.

**Misconception 5: "Pair trading is risk-free because it's market- neutral."**

Pair trading has spread risk, regime-break risk, crowding risk (2007), borrow-cost risk on the short leg, and operational risk on margin and dividends. Net dollars zero is not net risk zero — LTCM was market-neutral and still lost 90%.

**Misconception 6: "If the spread keeps widening, my position gets bigger and the eventual reversion is more profitable."**

This is the martingale fallacy applied to pairs. As the spread widens, your mark-to-market loss grows, the broker raises your margin requirement, and the very mechanism that should make the trade more profitable in theory becomes the mechanism that forces you out before the reversion arrives. Stay solvent first — the market can stay irrational longer than you can — and that discipline is the anchor here.

Misconception 7: "Borrow cost is small enough to ignore."

For mega-cap pairs (KO, PEP, V, MA) it usually is — single-digit basis points. For smaller-cap or specialty pairs it isn't, and the borrow rate can flip from 1% to 30% on a press release. Always check the borrow on the short leg before sizing the trade.

Misconception 8: "The 2007 quant quake is ancient history."

The same mechanism repeats. February 2018 had a vol-target factor unwind. March 2020 had a multi-strat de-leveraging week. Every few years, factor crowding produces another four-day correlated unwind. The quake was the cleanest case study, not the only one.

**Misconception 9: "If I use options to express the pair, I avoid the borrow problem."**

You also avoid the linear-spread P&L. An options-based pair has non-linear exposure to volatility and a defined-cost premium that eats into the spread alpha. It is not a free upgrade — it is a different trade with different greeks.

Misconception 10: "Pair trading is alpha, full stop."

Pair trading is alpha *as long as the spread cointegrates and the strategy isn't crowded*. Both conditions can fail. It is structural alpha conditional on regime and crowding — not a permanent free lunch.

**Misconception 11: "The strategy works because I picked the right pair."**

The strategy works because many pairs simultaneously revert slightly, and the small per-pair edge aggregates into a usable return stream at the book level. Single-pair returns are mostly noise; book-level returns are where the alpha is recognisable.

---

4. Q&A

**Q1: Should a retail investor with a $100k portfolio actually run a pair-trading book?**

A: Probably not as a serious P&L source. The minimum scale for a pair-trading book to be statistically meaningful is twenty-to- thirty pairs simultaneously, with operational discipline on borrow, margin, and dividends. Below that scale, you have a high-overhead toy. *Run one or two pairs in small size for education* — a few percent of your portfolio — and treat the output as tuition. If you want the alpha without the overhead, buy a market-neutral mutual fund or ETF and let someone else run the book.

Q2: What lookback window should I use for the Z-score?

A: 60 trading days is a sensible starting point for daily data; 40 is more responsive but noisier; 120 is more stable but slower to detect regime breaks. The right answer is a function of how fast the cointegration of your specific pair tends to reset. For mega-cap pairs (KO/PEP), 60 to 90 days. For more volatile pairs (small-cap, biotech), 30 to 60. Calibrate, do not adopt by default.

Q3: How big a position can I take in one pair?

A: A common institutional rule is 2-5% of the sleeve per pair, with a hard 10% cap. For a retail investor, on a $100k portfolio with a $10k pair-trading sleeve, that's $200-500 of gross exposure per pair (i.e., $200 long + $200 short = $400 gross at the small end). It feels small. It is small. Pair trading is a per-pair-tiny, many-pair-aggregated strategy by design.

Q4: When do I cut a pair and admit the cointegration broke?

A: When z extends past +/-3 to +/-4 and shows no signs of returning, and you can identify a fundamental reason for the break — a strategic-mix change in one of the companies, a regulatory event that hits one and not the other, a takeover. A spread that goes to -3.5 with no fundamental story is more likely to revert than one that goes to -2.5 with a clear story. Trade the fundamental, not just the Z-score.

Q5: How does pair trading interact with taxes?

A: Poorly, in a US taxable account. Pair trades have short holding periods on average (weeks to a few months), so realised gains are short-term — taxed at ordinary income rates, not long-term capital. The tax drag on a pair-trading book is one of the largest hidden costs. This is one reason institutional pair traders run inside tax-deferred wrappers (pension money, insurance separate accounts) where the structural tax disadvantage is absent.

Q6: How do dividends work for the short leg?

A: You owe them. If you are short PEP across the ex-dividend date, the dividend is debited from your account on pay-date and credited to the lender. There is no offset. For dividend-rich pairs (utilities, REITs), this is a meaningful drag and must be included in the per-trade cost calculation.

**Q7: What's the difference between pair trading and statistical arbitrage?**

A: Stat arb is pair trading scaled up: the same Z-score logic, but applied across hundreds-to-thousands of pairs simultaneously, with multi-factor decomposition (sector neutralisation, factor neutralisation), with optimised execution (VWAP/TWAP, dark pools), and held for shorter horizons (intraday to a few days). Pair trading is the educational unit; stat arb is the production implementation.

Q8: Why did the 2007 quant quake hit specifically pair traders?

A: Because the universe of "pairs" identified by a standard cointegration screen is largely the same across funds — the same sector pairs, the same factor neutralisations, the same optimisation. When one large player has to liquidate, they unwind the same trades that everyone else is on. Diversification in trade count does not help when the trades themselves are factor- correlated.

Q9: Can I use options instead of stock for the short leg?

A: Yes, and for retail it's often cleaner — buy a put on the expensive leg instead of shorting the stock. You avoid borrow, locate, and dividend obligations. You pick up theta decay and a volatility-dependent premium, and the linear-spread P&L is distorted into something more option-like. For a small retail sleeve where operational simplicity dominates, options-based pairs are a reasonable choice. The barbell logic naturally points here.

Q10: How does pair trading fit with the four tranches?

A: Pair trading is opportunistic — it sits in the third tranche alongside specific structural-alpha trades. It is not bedrock (too volatile, too operationally demanding), not core (it doesn't target a long-run risk premium), and not asymmetric speculation (no convex payoff). It earns its place by being roughly orthogonal to the rest of the portfolio.

Q11: What does a typical pair-trading sleeve P&L look like?

A: Many small wins (50-150 bps each on the gross), occasional regime breaks that cost 200-500 bps, and a flat-line during periods of compressed cross-sectional vol. Annualised return on the sleeve is typically in the 4-8% range gross, with realised vol of 5-10%. After fees and taxes for retail, considerably less. The chart shape is "many small steps up, occasional cliffs down," and the Sharpe ratio is the metric that matters — not the absolute return.

Q12: What should I read next if I want to actually build this?

A: Three things. (a) The original Gatev-Goetzmann-Rouwenhorst paper (2006) on pair trading — the academic baseline. (b) Khandani-Lo (2007), the post-mortem on the August 2007 quant quake. (c) The chapters in Lopez de Prado's *Advances in Financial Machine Learning* on cointegration and triple-barrier labelling. Read in that order — academic foundation, the war story, then the modern toolkit.