4. Correlations that go to 1 just when it matters

Macro 5: Debt, credit and crises

Every diversified portfolio rests on an assumption that is almost never stated: that the relationships between assets measured in the past will continue to hold. Those relationships are estimated from data dominated by calm periods, because calm periods make up most days. And diversification does not have to work on ordinary days: it has to work on the five worst days of the decade, precisely the ones on which the estimated correlations stop describing reality.

What a correlation measures and what it does not

Correlation sums up in a single number between -1 and 1 how far two assets have moved together. It is the raw material of any portfolio risk calculation: the volatility of a collection is not the average of the volatilities but something smaller, thanks to its components not moving in step. That discount is, literally, diversification.

Three limitations to keep in mind:

• It only captures linear relationships. Two assets can show a correlation close to zero for years and then collapse together in response to the same extreme event.
• It is an average over the chosen period. A correlation of 0.2 can be the average of years at 0.05 and weeks at 0.9, and that average describes neither of the two regimes.
• It is not stable, and its instability is not random: it shifts systematically in the same direction when stress arrives.

Why calm-period correlations do not hold in a crisis

The explanation is not statistical, it is mechanical, and it links straight back to the previous module:

• In calm conditions each asset follows its own fundamentals. A pharmaceutical company and a copper miner respond to different things, which is why they move with little relation to each other.
• Under stress what matters is who holds them and how they are financed. If the same leveraged participants hold both assets and receive margin calls, both get sold on the same day for the same reason, which has nothing to do with pharmaceuticals or with copper.
• Forced selling is indiscriminate by construction. What gets sold is what has a buyer, so the selection criterion stops being value and becomes liquidity.
• Risk models closely resemble one another. When volatility rises, almost all of them order a cut in exposure at the same time, and that simultaneous cut creates the fall the models were measuring.

Put differently: in normal times assets behave according to what they are; in a crisis they behave according to who holds them and how much is owed. Saying that correlations go to 1 is therefore not rhetorical exaggeration but a description of a change of regime in the mechanism that sets prices.

Apparent diversification against real diversification

From this follows a practical distinction that separates a robust portfolio from one that only looks robust.

Apparent diversification consists of holding a lot of different lines: twenty funds, five sectors, three countries, cryptocurrencies, commodities. If they all depend on the same underlying factor, whether global risk appetite, the level of rates or the availability of funding, you have a single bet repeated twenty times. Counting positions is not measuring risk.

Real diversification demands assets that respond to different causes and that sit with different holders running different funding structures. Even then it is worth accepting the limit: some macro events are large enough that no refuge exists within the market other than cash.

An example with figures: the correlation between US equities and Treasuries was negative through much of the 2000s and 2010s, which is why the 60/40 portfolio worked so well. In the 1970s it was positive, and it was positive again in 2022, when both legs fell together. The sign depends on whether the dominant shock is one of growth or one of inflation: it is not a constant of nature, it is a property of the prevailing macro regime.

March 2020: the day even Treasuries were sold

The most instructive episode of recent years. Between 9 and March 18, 2020, with equities collapsing, something happened that no correlation model built on recent history allowed for: long-dated US Treasuries, the safe haven par excellence, began to fall at the same time as equities. The ten-year yield, which had hit record lows near 0.5%, rebounded sharply in the middle of a stock market catastrophe. Gold fell too. The spreads between newly issued Treasuries and equivalent older issues widened anomalously, unmistakable evidence that the problem was the plumbing of the market and not the credit of the issuer.

The reason was the one this module describes. Leveraged funds exploiting small differences between the cash bond and its future had to unwind. Asset managers all over the world needed dollars to meet redemptions and margin, and sold the only thing they could shift in size: government debt. The dealers who normally absorb that flow had full balance sheets and stopped doing so. The most liquid asset on the planet stopped behaving like one for a few days, and only normalized when the Federal Reserve began buying on the order of tens of billions of dollars a day.

What you take away from this level

• There is no debt ratio marking the frontier of the unsustainable: what matters is the currency, the maturity, the holders and the relationship between interest rates and nominal growth.
• The credit spread gives warning before equities do, and a minimal level signals not the absence of risk but the absence of compensation for it.
• Leverage does not multiply symmetrically: zero is absorbing, and the ruin threshold is the inverse of the leverage.
• In a crisis you sell what you can, not what you want to, which is why assets with nothing wrong with them fall as well.
• Diversification measured in calm conditions is an optimistic estimate, and the only asset that never correlates with anything else is cash.