Count the Pairs,
Not the People

Twenty-three people in a room. Two hundred fifty-three pairs. The same arithmetic governs a portfolio, and it explains why diversification stops working long before anyone notices it has stopped.

Put twenty-three people in a room and the odds that two of them share a birthday are 50.7 percent. Most people guess something closer to five percent. The gap between those two numbers is not a failure of arithmetic. It is a failure to notice which question is being asked.

The intuitive guess answers a different question: what are the odds that someone shares my birthday. That number is small, and it stays small. The actual question is whether any two people match. The math is indifferent to the individual. It is counting the system.

Twenty-three people generate 253 pairs. Each pair is a separate opportunity for a match, and the pairs accumulate faster than the guests.

The party guest is doing arithmetic on the wrong question. The arithmetic is fine. The question is wrong.

01Where this stops being an analogy

Most writing that connects the birthday paradox to markets does so by metaphor. There is one place where no metaphor is required, because it is the same formula operating on the same kind of object.

A portfolio of n holdings has n variances, one for each position. It also has n(n − 1)/2 distinct covariances, one for each pair. Ten holdings produce 45 pairs. Twenty-three produce 253. Fifty produce 1,225. That is not an analogy to the birthday problem. It is the identical combinatorial fact applied to a different set of objects.

Portfolio variance is not the average of the individual variances. For an equally weighted portfolio it decomposes cleanly:

Read the weights. At 23 holdings, 4.3 percent of portfolio variance traces to the average variance of the individual positions. The remaining 95.7 percent traces to how they move against one another.

Three holdings already put two thirds of the weight on covariance. Ten holdings put ninety percent there. Past that point, adding positions barely changes what the portfolio does, because the term being diluted was already the smaller one.

This is the part that matters in a crisis. Diversification does not converge toward zero risk. It converges toward the average covariance of the things being held. When correlations rise, the floor rises with them, and every holding rises together because the floor was always the shared term.

A portfolio of thirty positions that all draw on the same underlying condition is not thirty bets. It is closer to one bet, expressed thirty ways, with a longer statement.

02Both curves flatten

The birthday problem is usually described as accelerating: the curve said to start flat and finish vertically, with the last additions doing most of the work. It is a natural thing to assume and it is the reverse of what happens.

The steepest single addition is the twentieth person, who moves the probability by 3.23 points. The fiftieth person moves it by 0.46. The seventieth moves it by 0.02. The curve is bounded above by certainty, and everything bounded above eventually flattens.

Set Figure 3 next to Figure 2. They are the same shape. Both climb quickly, both decelerate, and both settle against a bound they cannot pass. The birthday curve is bounded by certainty. The diversification curve is bounded by average covariance.

That shared shape is the actual lesson, and it has a direct portfolio reading. The benefit of an additional holding is largest when there are few holdings, and approaches nothing well before most portfolios stop adding. The twentieth position does very little. The fiftieth does almost nothing measurable. What continues to grow after the benefit has flattened is the count of pairs, which is to say the count of things that can move together at the wrong moment.

03Two places the analogy is usually stretched

The birthday paradox gets applied to two other investing claims. Both conclusions are defensible. Neither follows from this arithmetic, and borrowing the arithmetic to support them weakens arguments that stand fine on their own.

Rare market events

The claim is that because surprising coincidences are common in large samples, extreme market moves should also be expected. The conclusion is right. The reasoning does not transfer.

Birthdays are independent draws from a roughly uniform distribution over 365 outcomes. The paradox is a statement about how many comparisons exist, not about how any single draw behaves. Market returns are not independent, not uniform, and not thin tailed. Extreme moves are common because the distribution has fat tails and because the draws share common causes, which is a different mechanism entirely.

The correct version of the claim requires no borrowed support. Returns are not normally distributed, the tails are heavier than a normal distribution implies, and position sizing that assumes otherwise fails at the moment it is being relied upon. A portfolio is built to survive those moves rather than to anticipate them.

Compounding

Compounding and the birthday problem both defeat linear intuition, which is where the association comes from. They are opposite shapes.

• The birthday curve is bounded and decelerating. It cannot exceed 1, so its gains must shrink.
• Pairs grow quadratically and without bound. Doubling the holdings roughly quadruples the pairs.
• Compound value grows exponentially and without bound. Doubling the horizon squares the growth factor.

Three different functions, routinely collected under the single word exponential. The distinction is not pedantry. It determines which one you are allowed to extrapolate. Compounding rewards extending the horizon. Diversification does not reward extending the holdings list, because that curve stopped paying somewhere around the tenth position.

04The version that applies to households

The pairs formula does not stop at the portfolio. It applies to anything that has to agree with something else.

Count what a household actually holds. Not the balances. The objects that must remain consistent with one another: accounts, entities, trusts, beneficiary designations, titling, state jurisdictions, professional relationships. Each pair is a place where two things can disagree.

Between the first row and the third, the objects grew by a factor of 4.6. The pairs grew by a factor of 22.8. In most households the balance sheet over that same span grows by considerably less than 22.8.

This is why an eighth account rarely feels like a decision. It arrives as paperwork. It leaves behind forty new relationships that someone is now responsible for keeping consistent, and the responsibility is usually unassigned because it was never visible enough to assign.

It is also why the failures cluster where they do. A trust that was drafted and never funded. A beneficiary designation that survived a remarriage. A titling arrangement that made sense under one state's rules and is now being read under another's. None of these is a failure of any single object. Each one is a pair that stopped agreeing.

05What the arithmetic actually recommends

Nothing about counting pairs argues for holding fewer things. It argues for measuring the right quantity.

• A holdings count describes the diagonal. It says nothing about the 253 entries that determine behavior.
• Correlations measured in calm conditions describe the floor in calm conditions. The relevant floor is the one that applies when it matters, and it is higher.
• Positions added past the point where the benefit curve flattened are not diversification. They are additional pairs, carrying the coordination cost of a decision without the risk reduction that was supposed to justify it.
• The largest single-name exposure in most affluent households is not in the portfolio. It is the employer, appearing simultaneously as salary, equity compensation, deferred compensation, and frequently real property in the same metropolitan tax jurisdiction. Correlation measured inside the portfolio will not detect it, because the correlated assets are outside the portfolio.

The last point is the one that generalizes. Every version of this problem has the same shape. The object being measured is smaller than the system it sits inside, and the interactions live in the part that was not measured.

Most advice optimizes individual decisions. The harder and less visible problem is the connections between them.

Twenty-three people is a small room. It is also 253 chances for something to line up.