Correlation and What Diversification Actually Removes
7 steps · one page
In short
The previous article established that a portfolio's risk depends on how its holdings move relative to one another. Correlation is the number that measures it, and almost everything difficult about portfolio construction is difficult because of the properties of that number.
Scope. This article explains what a correlation measures, how it governs the benefit of diversification, and the three ways the number misleads. It recommends no allocation, suggests no weights and sets no target for any measure. All figures use the illustrative teaching parameters fixed on this pillar's hub and are not forecasts of any market.
It runs from −1 to +1. At +1 two holdings move in lockstep; at 0 their movements are unrelated; at −1 one rises exactly as the other falls. What matters is that the benefit of combining holdings is entirely determined by where in that range they sit.
What correlation buys
Taking equal amounts of the two assets on the hub parameters — equities with a 16.0% standard deviation, bonds with 6.0% — the weighted average of their individual variabilities is 11.00%. What the combination actually delivers depends on the correlation.
| Correlation | Standard deviation of the 50/50 combination | Reduction against the weighted average of 11.00% |
|---|---|---|
| −1.0 | 5.00% | 6.00 pp |
| −0.5 | 7.00% | 4.00 pp |
| 0.0 | 8.54% | 2.46 pp |
| +0.2 | 9.09% | 1.91 pp |
| +0.5 | 9.85% | 1.15 pp |
| +0.8 | 10.55% | 0.45 pp |
| +1.0 | 11.00% | 0.00 pp |
Worked example
Worked example — diversification is the gap between the average and the combination, and correlation is the whole of it. At a correlation of +1.0 the combination is exactly the weighted average and the benefit is precisely nothing — holdings that move together are, for risk purposes, one holding. At −1.0 the benefit is 6.00 percentage points and the combination is less variable than either asset alone. Note the shape: the reduction is not linear in correlation. Moving from +1.0 to +0.5 buys 1.15 points; moving from 0.0 to −0.5 buys a further 1.54. The benefit accelerates as correlation falls, which means the difference between two holdings correlating at 0.8 and at 0.5 matters far less than the difference between 0.2 and −0.1. This portal sets no target correlation and recommends no combination.
The floor, restated as a function of correlation
The previous article established that a portfolio of many equally weighted holdings converges to a floor no number of holdings can breach. That floor is set by correlation alone.
| Average pairwise correlation | Irreducible standard deviation |
|---|---|
| 0.00 | 0.00% |
| 0.05 | 6.71% |
| 0.10 | 9.49% |
| 0.20 | 13.42% |
| 0.40 | 18.97% |
| 0.60 | 23.24% |
| 1.00 | 30.00% |
Two readings of that table are worth holding together. At zero average correlation the floor is zero — with enough genuinely unrelated holdings, variability could in principle be eliminated entirely. At an average correlation of 0.20 the floor is already 13.42%, or 45% of a single holding's variability. A small amount of shared movement destroys most of what unlimited diversification could otherwise achieve.
Three ways the number misleads
First: correlation measures linear co-movement and nothing else. Two holdings can be strongly related and correlate at zero — if one rises when the other makes either a large move up or a large move down, the linear relationship cancels. A correlation of zero does not mean independence, and treating it as such is the most common technical error in this area.
Second: it is an estimate from a window, and the window is a choice. A correlation computed over three years, five years or twenty will differ, as will one computed on daily against monthly data. A correlation quoted without its period and frequency is not a fact; it is one of several available numbers.
Third, and most consequential: it is not stable, and it moves in the direction that hurts.
Why the reported number understates the one that matters
Markets behave differently in calm and in stress, and holdings that move independently most of the time frequently move together when a great deal is falling at once. The consequence for a single reported correlation can be computed.
The simulation below draws 250,000 periods from a mixture: 85% of periods from a calm regime where the two holdings correlate at 0.10 with modest variability, and 15% from a stress regime where they correlate at 0.85 with roughly three times the variability.
| Measurement | Correlation |
|---|---|
| Calm periods, 85% of the sample | 0.10 |
| Stress periods, 15% of the sample | 0.85 |
| The single full-sample figure a reader would be quoted | 0.56 |
Worked example — the reported correlation describes neither regime. The full-sample figure of 0.56 is far above the 0.10 that applies on most days, and 0.29 below the 0.85 that applies when it matters most. It is not wrong — it is a correctly computed average of two different worlds, and it describes neither. Note the direction of the error that matters: the number a reader is given understates co-movement precisely in the conditions where diversification is being relied upon. A second subtlety is visible in the arithmetic: the full-sample figure sits much closer to the stress value than the 85/15 split would suggest, because the stress periods are far more variable and therefore dominate the covariance. Even so, it remains materially below the stress correlation. This is a simulation with stated parameters illustrating a property of mixtures, not a measurement of any market.
What follows
Diversification is a claim about the future behaviour of a relationship, not about the past. A portfolio assembled on historical correlations is assembled on the assumption that those relationships persist, and that assumption is doing more work than the assembly.
Shared exposures matter more than counts. Holdings can be numerous and superficially varied while sharing an underlying driver — the same customers, the same input costs, the same funding conditions — and the correlation will reveal that only after it has mattered.
And the honest position is that the most important number in portfolio construction is the least reliable one. Expected returns are contested and variability is fairly stable, but correlation is both central and unstable, which is a poor combination and is not improved by ignoring it.
Frequently asked
8 questions
What does correlation measure?
The degree to which two holdings move together, on a scale from −1 to +1. It is the quantity that determines how much combining holdings reduces variability.
How much does correlation change the benefit?
Entirely. On the hub parameters a 50/50 combination has a standard deviation of 11.00% at a correlation of +1.0 — exactly the weighted average, so no benefit at all — and 5.00% at −1.0. The benefit is the gap, and correlation is the whole of it.
Is the benefit proportional to correlation?
No, it accelerates as correlation falls. Moving from +1.0 to +0.5 buys 1.15 percentage points; moving from 0.0 to −0.5 buys a further 1.54. So the difference between 0.2 and −0.1 matters far more than the difference between 0.8 and 0.5.
Why does a small correlation destroy so much of the benefit?
Because the floor is the single-holding variability times the square root of the average correlation. At zero the floor is zero; at 0.20 it is already 13.42%, which is 45% of a single holding's variability.
Does a correlation of zero mean two holdings are independent?
No. Correlation measures linear co-movement only. Two holdings can be strongly related and correlate at zero — for instance if one rises whenever the other makes a large move in either direction. Treating zero as independence is the commonest technical error here.
Why does the window matter?
Because a correlation computed over three years differs from one over twenty, and daily data differs from monthly. A correlation quoted without its period and frequency is one of several available numbers rather than a fact.
Do correlations rise in a crisis?
Holdings that move independently in calm conditions frequently move together when a great deal is falling at once. In the simulation, a calm correlation of 0.10 and a stress correlation of 0.85 produce a single reported figure of 0.56 — above what applies on most days and well below what applies when it matters.
What is the practical consequence?
That the number a reader is given understates co-movement precisely in the conditions where diversification is being relied upon, and that diversification is a claim about the future behaviour of a relationship rather than a fact about the past.
References
- Markowitz (1952) — Portfolio Selection, Journal of Finance 7(1) —
- Longin and Solnik (2001) — Extreme Correlation of International Equity Markets, Journal of Finance 56(2) (correlation in stress versus calm) —
- Investor.gov (SEC) — Diversification —
Educational and informational only — not investment advice, a recommendation, or an offer to buy or sell any security. Investing involves risk, including the possible loss of principal. Worked examples use fictional companies and figures.