Asset Correlation: Why It Fails in a Crash

A 60/40 portfolio whose two sleeves carry a correlation of -0.3 has an annual volatility of about 9.2%. Move that correlation to +1 and the same holdings at the same weights produce 12.4%, a swing 35% larger, with nothing in the portfolio changed. Correlations tend to climb toward +1 during the selloffs the diversification was bought to survive, so the protection you measured in a calm market is at its smallest when it is called on.

Correlation is a reading, not a property

Correlation runs from -1 to +1 and describes how two assets have moved together over some window. Both halves of that sentence get dropped when the number is quoted, and both matter.

"Have moved" is past tense for a reason. Every coefficient you will see is computed backwards over a period somebody selected. A stock and bond correlation of -0.3 describes a sample. It carries no commitment about next March.

"Some window" is the larger problem. The same pair measured over 30 days, three years and twenty years gives three different numbers, and those numbers do not always share a sign. Long-run allocation tables are built from long windows because long windows are stable. Your portfolio lives in the short ones.

What a coefficient is worth in pounds

Volatility for a two-asset portfolio is:

sd = sqrt( (w1 x sd1)^2 + (w2 x sd2)^2 + 2 x w1 x w2 x rho x sd1 x sd2 )

Take £100,000 split 60/40. Assume the 60% growth sleeve has an annual standard deviation of 16% and the 40% defensive sleeve 7%. Those two inputs are assumptions chosen to be round rather than measured market figures, so substitute your own. The shape of the answer is what carries.

The first two terms are fixed by the weights. (0.6 x 0.16)^2 is 0.009216 and (0.4 x 0.07)^2 is 0.000784, and they sum to 0.01 exactly. Only the third term moves.

At +1 the result is 12.40%, which is simply the weighted average of the two volatilities (0.6 x 16% plus 0.4 x 7%). That is the ceiling, and it is what you hold when correlation breaks. The entire benefit of owning two things instead of one, in this arithmetic, is the distance between your actual figure and 12.40%.

Read the table as a crisis rather than a menu. You sized the portfolio expecting a typical year to swing roughly £9,158 either way. Correlation moves to +0.6 in the selloff and the swing becomes £11,500, a quarter larger than planned. At +1 it is £12,400.

The response curve is concave, which produces a result most people guess wrong. Half the total span between -0.3 and +1.0 is already spent by the time correlation reaches +0.3: the midpoint of 9.16% and 12.40% is 10.78%, and +0.3 gives 10.78%. A coefficient drifting from mildly negative to mildly positive costs you 1.62 points of volatility. The dramatic final lurch from +0.6 to +1.0 costs 0.90. The quiet drift is worth more than the crash you were watching for.

Why we stopped publishing the lookup table

An earlier EptaWealth draft carried a table of what each asset "usually" does when equities fall: gold up, bonds up. We pulled it, because 2022 is the case that breaks it. Equities and gilts fell in the same year, and gold in dollar terms finished roughly where it started, so the two assets most often named as equity hedges delivered a loss and a flat line inside the same twelve months.

That is not a freak result requiring an asterisk. Equities and bonds are both priced off the discount rate. When the shock is the discount rate itself, rather than a shock to earnings or a credit event, both respond the same way. The negative stock and bond correlation that investors had grown used to since 2000 was a feature of an era of falling rates, and it was measured entirely inside that era. A hedge that works against one kind of shock is not a hedge against every kind, and no table of asset behaviour can tell you which kind is coming.

Our position, and the case against it

We think the correlation coefficient is a weak input to portfolio design and a reasonable input to portfolio review. Size positions by what you can afford to watch fall, then use correlation afterwards to explain what happened. When we built the table above for our own allocation notes, the coefficient we had been working from was a twenty-year average, and the rolling three-month figure for the same pair had spent stretches of that period on the other side of zero. The long-run number was accurate and told us nothing about any quarter we actually sat through.

The counter-argument is fair and we would not wave it away. If your holding period is twenty years and you genuinely will not sell, the long-run correlation is the one that governs your outcome, and every short-window spike is noise you never realise. On that reading, designing around crisis correlations means paying a permanent cost in expected return to insure against a temporary event. It turns on whether you are confident you will hold, and most people overestimate that about themselves.

What to measure instead

Correlation is a summary of things you can observe directly, and the observations are more useful than the summary. Look at how much of your portfolio fell together on your five worst days, and at how much of your total value sits in positions that would all reprice on a single rate move. Both are countable from your own records; neither needs a coefficient.

That requires one return methodology applied across every holding. If your equities are measured on one basis and your metals or property on another, a cross-asset comparison is not measuring what you think it is. Property is the sharpest case, because it has no live price feed and is valued manually, which makes its measured correlation with anything largely an artefact of when you last updated the valuation. Tracking precious metals and real estate covers how that valuation gap distorts comparisons.

For the construction question behind all of this, see the multi-asset portfolio management guide and how to build a multi-asset portfolio from scratch.

The volatility figures above are arithmetic on assumed inputs, not a projection of any real portfolio, and this page sets out a method rather than personal advice.