Diversification is only worth what the independence behind it is worth, and correlation is how that independence is measured - including the fact that it changes exactly when you need it not to.
12 min read · Level 2
Correlation is the number most likely to be quoted confidently and least likely to survive contact with the event it was quoted for. Every published figure is an average over a window someone chose, dominated by ordinary days, and the thing you are actually buying a diversifier for is the extraordinary ones - where the relationship routinely behaves differently. The useful discipline is therefore not to measure correlation more precisely but to stop relying on the measurement at all where you can: if you can name the different real-world forces that two holdings depend on, you have a reason for them to keep diverging that does not evaporate when the sample period ends. If the only evidence of independence is a historical coefficient, treat the diversification as provisional.
Correlation measures the degree to which two assets have moved together, on a scale from +1 to -1. At +1 they move in lockstep, at 0 their movements tell you nothing about each other, and at -1 one rises precisely when the other falls. Most real pairs sit somewhere between 0 and +1.
The practical consequence is that combining assets reduces the combined variability by an amount that depends on their correlation, and on nothing else about them. Two assets can have identical returns and identical volatility, and one pairing can be far more useful than the other purely because of how their movements line up.
The second thing to know is that a correlation is a measurement of the past over a chosen window. It is not a property of the assets, and it is not stable.
Because correlation is where diversification either delivers or quietly fails. A portfolio can look admirably spread across sectors, regions and asset types while every part of it is driven by the same underlying thing - the cost of borrowing, the price of energy, the health of one economy - in which case the spread is cosmetic.
It also explains an experience almost every investor eventually has: a portfolio that behaved as designed for years and then fell all at once. In periods of stress, correlations between risky assets tend to rise towards one, which means diversification is weakest at the precise moment it is being relied upon. Knowing that in advance is the difference between a disappointment and a surprise.
Educational, not personalised financial advice. Nothing in the Investor Academy is a recommendation to buy, sell or hold anything — what is right for you depends on your own circumstances and the rules where you live.
Take two assets, each with a 20% standard deviation of annual returns, held in equal weights. The portfolio's standard deviation depends only on the correlation between them:
At +1.0 it is 20.0% - no benefit whatsoever, because you effectively own one asset twice. At +0.5 it is 17.3%. At 0.0 it is 14.1%. At -0.5 it is 10.0%. At -1.0 it is 0% - the movements cancel exactly.
(The arithmetic: with equal weights and equal 20% volatilities, the portfolio variance is 0.02 plus 0.02 times the correlation, and the standard deviation is its square root. At a correlation of 0 that is the square root of 0.02, which is 0.141.)
Nothing about the assets changed across those five lines. Same expected returns, same individual volatility. The entire difference in the outcome is the relationship between them, which is why correlation deserves more attention than the holdings list usually gets.
The first mistake is treating a measured correlation as a fixed property. It is an average over whichever window was chosen, and different windows produce materially different numbers for the same pair. A near-zero correlation measured over a calm decade is a statement about that decade.
The second is hunting for negative correlation as if it were the goal. Assets that reliably move opposite to the things you own mostly cost something to hold, because someone has to be paid to take the other side. Low or unstable positive correlation, held cheaply, is usually the realistic objective.
The third is assuming different labels mean different drivers. Property, private credit, small companies and emerging markets have distinct names and can all be driven primarily by the availability and price of credit, which makes them one bet wearing four coats.
Write the one or two underlying factors each position actually depends on - rates, energy prices, consumer spending, one economy, one currency. Two holdings sharing a driver will be correlated regardless of what a historical calculation over a quiet period says.
Look at the relationship over a long period and over the most stressed period available. If the two numbers differ a lot, the long-run figure is not the one that will apply when it matters.
Ask how the pair behaved in the worst decline in the record, not on average. Average correlation is dominated by ordinary days, and ordinary days are not what you hold a diversifier for.
A relationship that is low because the two things genuinely depend on different forces is far more durable than one that is low because the numbers happened to move that way over the sample. Understanding beats calculation here, because the calculation is always backward-looking.
Assume that in a severe decline the risky things you own will fall together to a greater degree than history suggests. If the portfolio only survives on the assumption that they will not, it has a hidden dependency on a number nobody controls.
When assessing whether a portfolio's apparent spread is real, when evaluating something sold as a diversifier, and before relying on any pairing to behave differently in a downturn.
Do not turn this into a search for the perfect low-correlation combination. Correlations are unstable enough that optimising finely against measured history is a way of fitting a portfolio to a period that will not repeat.
A shape seen repeatedly across market history is an asset adopted specifically for its low or negative correlation to equities, which then declines alongside them in a severe episode. The mechanism is usually the same rather than a coincidence: in a genuine crisis investors sell what they can rather than what they would like to, and forced selling links assets that have no economic relationship at all. A holding's correlation in normal conditions describes its economics; its correlation in a crisis often describes who owns it and how much borrowed money is behind them.
Ownership structure, not economics, frequently determines how two assets behave on the worst day.
A correlation computed over twenty years is dominated by the thousands of unremarkable days it contains, because there are far more of them than there are crisis days. A pair can therefore show a genuinely low long-run correlation while having moved together in every one of the handful of episodes that actually mattered to a portfolio. Both facts can be true at once, and only one of them is about risk.
Averages describe the typical day. Portfolios are usually damaged on the atypical ones, so look at those separately rather than letting them be averaged away.
Create a free account to read the rest and apply it to your own investing.