Correlation, and when it fails

Chapter 5 said diversification works when holdings do not fail together. Correlation is the measure of how much they move together, and it is the assumption underneath every diversified portfolio.

What it measures

A number between −1 and +1.

+1 — they move together, exactly. 0 — no relationship. −1 — they move in opposite directions, exactly.

Two holdings with a correlation near 1 are one holding. Two with a correlation near 0 genuinely spread risk. A correlation below 0 is rare, valuable and usually temporary.

The practical consequence: a portfolio's risk depends on the correlations between its parts, not just on the risk of each part. Ten volatile assets with low correlation can produce a calmer portfolio than five stable ones that move together.

The problem

Correlation is computed from history, and the history is mostly ordinary.

Two assets that behaved independently across ten years of normal markets can fall together in the one month that matters. And this is not bad luck — there is a mechanism.

In a crisis, people sell what they can. Not what they want to sell. Somebody facing a margin call, a redemption or a cash need sells whatever is liquid and saleable, which means selling pressure arrives in assets that have nothing to do with the problem.

Leverage unwinds everywhere at once. Positions funded with borrowed money are closed simultaneously across unrelated markets, because the funding dried up rather than because the assets changed.

Confidence is a single factor. When the dominant question becomes "is everything all right", the usual distinctions between assets matter less than that one question, and everything correlated to it moves together.

So the uncomfortable summary: correlations rise when you most need them to stay low. Diversification delivers least in the environment it was bought for.

What this does and does not mean

It does not mean diversification fails. It still removes specific risk, which is most of the risk in a concentrated portfolio, and in ordinary years it works exactly as described. A crisis is a bad few weeks, and portfolios are held for decades.

It does mean the benefit should not be over-estimated. A model saying a portfolio can fall at most 18% because of its correlation structure has used correlations from the wrong months.

It means stress-test rather than optimise. The useful question is not "what does the historical correlation matrix say" but "what happens if everything risky falls 40% together?" If the answer is intolerable, the allocation is wrong regardless of what the correlations said.

What actually stays uncorrelated

Short of the ideal, three things hold up better than most.

Cash and very short-duration instruments. Chapter 4. Dull, which is the property.

Government bonds, often — though chapter 3 of the fixed income subject's seesaw means a crisis driven by inflation and rising rates can hurt both bonds and equities at once, which is the case that breaks the classic pairing.

Your own income, if it is stable and unrelated to markets. The most under-appreciated diversifier most people own.

Gold has a reputation here and a mixed record: often helpful in confidence shocks, not reliably so.

The practical rules

Assume higher correlation than history shows, in the scenarios you care about.

Count exposures, not holdings. Chapter 5's point: four assets with one driver is one exposure.

Hold something genuinely uncorrelated, and accept that it earns less. That is the emergency fund of chapter 4, and the fixed income allocation of chapter 12 of that subject.

Never rely on a hedge you have not tested. A relationship that worked for ten years is a historical fact, not a contract.

The point

Correlation measures how much holdings move together, and portfolio risk depends on it rather than on each holding alone. Correlations are measured in calm markets and tend to converge in crises, because people sell what they can and leverage unwinds everywhere at once — so stress-test the portfolio against everything risky falling together rather than trusting the matrix.