Why does diversification fail exactly when you need it?
Diversification is the one genuinely free improvement in portfolio construction: combine assets that don't move together and portfolio volatility falls below the weighted average of the parts. The mathematics is solid.
It rests on a single input — correlation — and that input is not a constant. It is a statistic, estimated from a window, and it moves. Worse, it moves in a systematically inconvenient direction.
The empirical pattern
Across decades of market data, one regularity keeps appearing: correlations among risky assets rise in periods of market stress. Holdings that spent years behaving independently begin moving in lockstep exactly when their independence was the whole reason for owning them.
In 2008 the pattern was broad. Developed equities, emerging equities, corporate credit, commodities, real estate and many hedge fund strategies all fell together, despite years of history suggesting they were meaningfully distinct exposures. Portfolios that looked diversified on a correlation matrix estimated over 2004–2006 behaved, in the crisis, close to a single leveraged bet on "risky assets."
Why it happens
Three mechanisms, and none of them are statistical accidents:
A common factor dominates. In calm markets, asset prices are driven mostly by their own idiosyncratic news — this company's earnings, that country's growth. Under stress, one factor — risk appetite, or liquidity, or the price of borrowing — becomes so large that it overwhelms every idiosyncratic driver. When one factor moves everything, everything is correlated.
Forced selling ignores fundamentals. An investor facing a margin call sells what can be sold, not what should be sold. This mechanically transmits stress from the stressed asset into unrelated ones. Correlation is manufactured by the plumbing, not by the fundamentals.
The same players hold the same things. When many institutions run similar models and similar risk limits, they are pushed to reduce exposure at the same moment. Crowded positions unwind together regardless of what the assets have in common economically.
The complication that keeps you honest
It is tempting to summarise this as "in a crisis, all correlations go to one." That's a useful slogan and an incomplete fact.
In 2008, government bonds rose while equities fell — a flight to quality. The classic negative stock-bond correlation did exactly what its owners hoped.
In 2022, that same relationship reversed: equities and high-grade bonds fell together, because the shock originated in inflation and interest rates rather than in credit or growth. The hedge that worked in one crisis was an amplifier in the other.
So the accurate statement is harder and more useful: correlations are regime-dependent, and the regime that will apply to the next stress event is not knowable in advance. Sometimes they converge; sometimes a long-standing relationship simply inverts.
Measuring it yourself
The tool is a rolling correlation — compute the correlation of two assets' daily returns over a moving window (60 or 120 days) and plot how it evolves. What you'll typically see is not a flat line but a series of plateaus and jumps, with the sharpest jumps clustering around market stress.
Worked illustration: two equity regions with a full-period correlation of 0.65. Split by regime, that single number might decompose into roughly 0.55 in calm periods and 0.85 in stressed ones. The blended 0.65 describes neither regime — and a portfolio optimiser fed 0.65 has been told the diversification benefit will hold up in a crisis, which is precisely the claim the data refuses to support.
What professionals do about it
Not "solve it" — you cannot make correlations stable. The responses are all about not depending on the estimate: stress-testing with correlations forced toward one, examining conditional correlations in stressed sub-samples separately, and treating the diversification benefit reported by any single-number matrix as a calm-market figure.
This is the course's frame again. The correlation matrix is a map. It is drawn from calm terrain because most terrain is calm. It is least reliable on the days you consult it most urgently.
Try it now
- Three series are below. Navigate the first two to 2008 and Measure the year on each; then do the same for 2022. Note the sign in each case. That flip is the whole lesson in four numbers.
- Now do the harder comparison. The third chart is a different equity region, which ordinary correlation estimates treat as a genuine diversifier. Measure it against the first over a calm year, then over a crisis year.
- Write one sentence describing what the calm-year figure would have implied about crisis-year diversification. Observation only.
Unit done. Next: the risk that shows up in no correlation matrix at all.