Diversification

The single most important quantitative fact in portfolio construction is that portfolio risk is driven by covariance, not by the riskiness of individual holdings. Intuition first, then the formula, then the result that decides how many stocks to own.

Why does correlation drive diversification?

Intuition. Imagine two equally-sized holdings, each of which swings a lot on its own. If they always move together, same direction and same time, combining them changes nothing; you have one big swing. If they move independently, their swings partly cancel: on the days one is down, the other is as likely up as down, so the combined line is smoother than either. If they moved in perfect opposition, they would cancel almost entirely and the combined line would be nearly flat. The magic of diversification is entirely in that cancellation, and the cancellation is entirely governed by correlation: how much the holdings move together. Owning more names does nothing by itself; owning more independent names is the whole game.

What is the two-asset portfolio variance formula?

The two-asset formula. For a portfolio of two assets with weights w₁, w₂ (summing to 1), standard deviations σ₁, σ₂, and correlation ρ₁₂:

σ_p² = w₁²σ₁² + w₂²σ₂² + 2·w₁·w₂·ρ₁₂·σ₁·σ₂

where σ_p² is portfolio variance and σ_p (its square root) is portfolio standard deviation, our stand-in for "risk" in what follows. (Standard deviation measures volatility, not permanent loss, and that caveat matters. The mathematics of diversification is about the dispersion of outcomes, while the deeper risk you learned in M9.01 is permanent impairment. Both matter; they are not the same thing.) The only term that carries the correlation is the last one. Set ρ₁₂ = 1 and the formula collapses to (w₁σ₁ + w₂σ₂)²: the portfolio SD is just the weighted average of the two SDs, no benefit at all. Drop ρ₁₂ below 1 and that last term shrinks, pulling σ_p below the weighted average. That gap between the weighted-average SD and the actual σ_p is the diversification benefit, and it widens as correlation falls.

How low can diversification take portfolio risk?

The general case and the result that matters. For N holdings the variance is the sum of every variance and every pairwise covariance. Take the clean, revealing special case: N equally-weighted holdings (w = 1/N each), each with the same variance σ², and an average pairwise correlation ρ̄ (so the average covariance is ρ̄σ²). There are N variance terms on the diagonal and N(N−1) covariance terms off it, so:

σ_p² = (1/N²)·[N·σ² + N(N−1)·ρ̄σ²] = (1/N)·σ² + (1 − 1/N)·ρ̄σ²

Rearranged, this is σ_p² = σ²·[ρ̄ + (1 − ρ̄)/N]. Now read what it says as N grows:

• The first piece, (1/N)·σ², is the idiosyncratic (diversifiable) risk, the part unique to each holding. It shrinks as 1/N and heads to zero. This is the risk you are paid nothing extra to bear, because you can remove it for free by holding more names (M3.02's "one free lunch").
• The second piece, (1 − 1/N)·ρ̄σ², is the systematic (undiversifiable) risk, the shared movement no amount of diversification removes. As N → ∞ it converges to ρ̄σ². So the portfolio variance floor is the average covariance, and the portfolio SD floor is σ√ρ̄.

That floor is the punchline. If the average correlation among your holdings is ρ̄ = 0.3 and each has SD σ = 30%, then no matter how many stocks you own, your portfolio SD cannot fall below 30% × √0.3 = 30% × 0.5477 = 16.4%. Diversification is not a road to zero risk; it is a road to the systematic floor, and where that floor sits is set entirely by how correlated your holdings are. Lower the average correlation and you lower the floor. That, not adding more of the same, is how a portfolio manager actually reduces risk.

How many stocks do you need to diversify?

Where the marginal benefit of the Nth name goes to zero. Put numbers on it. With σ = 30% and ρ̄ = 0.3, using σ_p = σ·√[ρ̄ + (1−ρ̄)/N]:

The shape is the whole lesson. Going from 1 name to 10 removes about 90% of the diversifiable risk and takes SD from 30% to 18.2%. Going from 10 to 30, tripling your holdings and tripling your research burden, buys you only another 1.1 points of SD (18.2% → 17.1%). Beyond 30, you are shaving basis points. This is the empirical backbone of the classic finding (Evans & Archer 1968; Statman 1987) that most diversifiable risk is gone by roughly 20–30 stocks. Modern studies, with higher single-stock idiosyncratic volatility, honestly put the "well-diversified" number higher (40–50) for someone who genuinely wants the floor. The exact number depends on ρ̄ and on how equal your weights are. The shape, steep benefit early and near-flat quickly, is universal.

What is diworsification?

"Diworsification": the cost side the variance math hides. Peter Lynch coined diworsification (M6.02) for adding holdings that make a portfolio worse. The variance table shows risk falling (a little) with each added name; it says nothing about return. But the expected return of an equally-weighted portfolio is simply the average expected return of its holdings, so every name you add pulls the portfolio's expected return toward that name's expected return. Add your 30th-best idea and you are averaging in a below-average expected return in exchange for a trivial risk cut. Concretely: suppose your ten best ideas have an average expected forward return of 15%, and ideas 11 through 30 average 9% (they are worse ideas, which is why they rank 11–30). Going from 10 equally-weighted names to 30 moves your portfolio expected return from 15% to (10×15 + 20×9)/30 = 330/30 = 11%, a 4-point haircut, while cutting SD by only ~1.1 points (18.2% → 17.1%).

You paid four points of return to remove one point of volatility.

That is diworsification in numbers, and it is the mathematical heart of the concentration case: past the early names, diversification's benefit is nearly exhausted while its cost, diluting your best ideas with your worse ones, keeps rising.

The honest counter, which the concentration case will weigh, is that this arithmetic assumes you can actually tell your ideas apart: that your "15% ideas" really will beat your "9% ideas." If your ability to rank is weak or your expected-return estimates are noise, the return "cost" of diversifying is illusory and the risk benefit is real, so you should diversify. The whole debate turns on that single honesty.