Correlation, covariance and why they betray you in a crisis

Every investor who has lived through a real drawdown has noticed the same thing: the portfolio behaved worse than the risk model said it would. Positions that were supposed to offset each other fell together. The "uncorrelated" allocation turned out to be correlated after all. The usual explanation offered is that "in a crisis, correlations go to one." That is a description, not a mechanism, and treating it as an explanation leaves you unable to predict where it will happen next.

There are four distinct reasons the relationships between your holdings change under stress, and they are not the same reason. Two of them are statistical illusions — the correlation did not change, only your measurement of it. Two of them are real economic mechanisms that genuinely re-link assets that were previously independent. Separating them is the difference between panic and preparation.

Covariance is the quantity; correlation is the packaging

Portfolio variance is built from covariances. The covariance of two assets is:

Cov(1,2) = ρ · σ₁ · σ₂

Correlation ρ is covariance divided by the two volatilities, which rescales it onto the interval from −1 to +1. That rescaling is convenient for intuition and dangerous for risk management, because it discards the two σ terms — and in a crisis, those σ terms move violently.

Work an example. In calm conditions, two assets each have 15% volatility and a correlation of 0.2. Their covariance is 0.2 × 0.15 × 0.15 = 0.0045. In a crisis, volatility on each rises to 35% and correlation rises to 0.8. Covariance becomes 0.8 × 0.35 × 0.35 = 0.098 — roughly twenty-two times larger. Correlation quadrupled; covariance rose by a factor of more than twenty, because volatility entered twice.

Now the portfolio. Held 50/50 in calm conditions:

Var = 0.25(0.0225) + 0.25(0.0225) + 2(0.25)(0.0045) = 0.0135 → volatility 11.6%

Each asset alone is 15% volatile; the pair is 11.6%. Diversification cut risk by about 23%. In crisis conditions:

Var = 0.25(0.1225) + 0.25(0.1225) + 2(0.25)(0.098) = 0.11025 → volatility 33.2%

Each asset alone is now 35% volatile; the pair is 33.2%. Diversification cut risk by about 5%. The benefit did not vanish, but it shrank to roughly a fifth of what it was — and it shrank in the one regime where the reduction had real value. That is the betrayal, stated precisely.

Illusion one: measuring correlation inside the tail biases it upward

If you compute the correlation of two assets using only the observations where markets fell sharply, you will get a higher number than the full-sample correlation — even when the true relationship between them has not changed at all. This is a selection effect on the conditioning variable, and it is a mathematical property of the bivariate distribution rather than evidence of contagion. Boyer, Gibson and Loretan documented the point formally in the late 1990s, and Forbes and Rigobon made it the centrepiece of their 2002 paper "No Contagion, Only Interdependence," which showed that a large part of the apparent correlation spike during emerging-market crises disappears once the estimates are corrected for the increase in volatility.

The practical implication is uncomfortable but useful: some of the "correlation breakdown" you observe is your own measurement procedure, and comparing a stress-period correlation with a calm-period correlation is not an apples-to-apples comparison. When you look at conditional correlations, correct for volatility or expect to be misled.

Illusion two: you have far less data than parameters

A covariance matrix over N assets contains N(N+1)/2 independent parameters. For fifty assets that is 1,275 numbers. Ten years of monthly returns gives you 120 observations. You are estimating more than a thousand parameters from a hundred-odd data points, and the resulting matrix is not merely imprecise — it is structurally unstable, with small sampling errors producing large swings in the implied relationships.

This is why sample covariance matrices fed directly into an optimiser produce nonsense, a point developed in the next lesson. The standard remedy is shrinkage: pulling the noisy sample estimates toward a simpler, more stable structure such as a constant-correlation model. Olivier Ledoit and Michael Wolf, both at the University of Zurich, formalised this approach in a series of papers from the early 2000s, including the memorably titled "Honey, I Shrunk the Sample Covariance Matrix." Shrinkage does not make you right; it makes you less confidently wrong, which is usually the better trade.

Mechanism one: leverage and forced selling create a common factor

When a leveraged holder faces a margin call, they do not sell the position they like least. They sell what they can sell — the liquid, the profitable, the things that still have a bid. That imposes a common selling factor across assets that share nothing economically except being owned by the same stressed balance sheet.

Brunnermeier and Pedersen described the self-reinforcing version of this in 2009: falling prices tighten funding constraints, tighter funding forces more selling, more selling drives prices down further. In a funding spiral, ownership structure becomes a risk factor. Assets held predominantly by leveraged investors move together regardless of their fundamentals, and the correlation is real, not an artefact.

This mechanism has a clear tell. Ask who holds the asset and with what leverage. Two positions held by the same over-extended pool of capital are one position for crisis purposes, however different their cash flows.

Mechanism two: a regime change re-links assets through a shared driver

The other real mechanism is a change in the macro variable that drives both assets. Equities and nominal government bonds are the canonical pair. When the dominant shock hitting the economy is a growth shock, the two move in opposite directions: bad growth news hurts earnings and lowers policy rates, so bonds rally as equities fall, and the bond allocation does its job. When the dominant shock is an inflation shock, both fall together: higher inflation raises discount rates, which hurts bond prices directly and equity valuations at the same time.

The equity–bond correlation is therefore not a property of the two asset classes. It is a property of the prevailing macro regime, and it has flipped sign over long historical periods. Investors who built portfolios during a long stretch of growth-shock dominance learned a rule that stopped applying when inflation returned as the driving shock — as it did in 2022, when equities and bonds fell together over the year.

March 2020 illustrated a third variant: a dash for cash in which even assets that normally act as havens, including gold, sold off for a period, because the binding constraint was the need for immediate liquidity rather than any view about the assets themselves.

The Swiss franc adds a specific twist

For a Swiss-resident investor, currency is not a side issue in this analysis. The franc has a documented tendency to appreciate during global risk-off episodes — capital seeks jurisdictions with credible institutions, a large current-account surplus and low inflation. That safe-haven behaviour means the CHF return on an unhedged foreign asset can be worse than the local-currency return exactly when foreign assets are falling.

Concretely: a global equity portfolio held unhedged by a Swiss investor takes two hits in a crisis — the fall in the assets and the appreciation of the franc against the currencies those assets are priced in. The correlation between "foreign asset returns" and "foreign currency returns" is negative from a CHF perspective in stress, which amplifies drawdowns rather than dampening them. In normal times the same mechanism runs in reverse and is easy to forget.

There is a more general version of this warning, and it is the one worth carrying. Through the early 1990s the currencies inside Europe's Exchange Rate Mechanism traded within narrow official bands, and any volatility or correlation estimated from that period described the bands rather than the market. When sterling and the Italian lira were forced out of the ERM in September 1992, the moves that followed were of a size the prior history ruled out — not because the market had changed character, but because the constraint that had shaped the data was removed. Any estimated relationship that exists only because a policy holds it in place carries the risk of that policy changing, and no amount of historical data will reveal it. The lesson on currency exposure works through the Swiss instance of this directly.

How to use this without pretending to precision

Treat measured correlations as a description of a regime, not a constant. Run your portfolio through explicit stress assumptions rather than trusting the historical matrix: ask what happens if every risk asset you own moves with correlation 0.9 and volatility double its long-run level, and check whether the resulting drawdown is one you could hold through. That is a cruder tool than an optimiser, and it fails more honestly.

Build diversification from economically distinct drivers rather than statistically distinct histories. Two assets with a low measured correlation but the same underlying dependence on global growth, global liquidity or the same leveraged holder base will re-link. Two assets with genuinely different cash-flow drivers are more likely to keep some independence.

And keep unlevered exposure to genuinely liquid assets — the ones you would be willing to sell into a bad market. Diversification that only works if you never have to trade is not diversification; it is a hope.

What to take away

• Correlation is covariance divided by two volatilities, so a portfolio can become dramatically more concentrated in risk even when correlation rises only modestly — a move from ρ = 0.2, σ = 15% to ρ = 0.8, σ = 35% multiplies covariance by more than twenty.
• Part of the observed "correlation spike" in crises is a statistical artefact of conditioning on large moves, established by Forbes and Rigobon in 2002; comparing tail correlations to full-sample correlations without correcting for volatility will mislead you.
• A covariance matrix over fifty assets has 1,275 parameters and ten years of monthly data provides 120 observations, which is why raw sample estimates are unstable and why shrinkage methods exist.
• Forced deleveraging creates real correlation between economically unrelated assets, so the question "who owns this and with how much leverage" is a risk question, not a curiosity.
• For a Swiss investor the franc's safe-haven behaviour means unhedged foreign holdings tend to lose twice in a global risk-off episode, and the 1992 ERM breakup is the standing reminder that any relationship held in place by policy can break without warning, leaving no trace in the prior data.

Sources

• Kristin Forbes and Roberto Rigobon, "No Contagion, Only Interdependence: Measuring Stock Market Comovements," Journal of Finance, 2002.
• Brian Boyer, Michael Gibson and Mico Loretan, "Pitfalls in Tests for Changes in Correlations," Federal Reserve Board International Finance Discussion Papers, 1999.
• Markus Brunnermeier and Lasse Heje Pedersen, "Market Liquidity and Funding Liquidity," Review of Financial Studies, 2009.
• Olivier Ledoit and Michael Wolf, "Honey, I Shrunk the Sample Covariance Matrix," Journal of Portfolio Management, 2004.
• Barry Eichengreen and Charles Wyplosz, "The Unstable EMS," Brookings Papers on Economic Activity, 1993 — the September 1992 ERM crisis and the collapse of the official bands.
• Bank for International Settlements, Quarterly Review — recurring analysis of cross-asset comovement, funding stress and market liquidity during the March 2020 dislocation.
• Andrew Ang, Asset Management: A Systematic Approach to Factor Investing, Oxford University Press, 2014, on regime dependence in asset-class relationships.