When Diversification Disappears: Why Portfolios That Look Independent in Calm Markets Become One Bet Under Stress

A portfolio is diversified only to the extent that its positions fail together less often than they succeed apart. That is not the definition most investors carry around. The usual one is about counting: many positions, several asset classes, more than one strategy, managers who do not talk to each other. Counting is a fine proxy in calm markets, when the things that make positions different are the things that drive their returns. It stops being a proxy in stressed markets, when the thing that drives returns is whatever the positions have in common, and a portfolio that held a hundred apparently unrelated bets discovers that it held one.

This piece is about that discovery and why it keeps happening to careful people. We set out a way of counting the bets a portfolio actually contains, show how that count collapses in a stylized stress episode, and trace the collapse to a specific mechanism: exposures that are invisible in a covariance matrix estimated from calm data because they are only expressed under stress. The most important of these is liquidity, the ability to sell without moving the price, which almost every portfolio is short and few portfolios measure. The conclusion is not that diversification is an illusion. It is that diversification has to be measured in the regime in which it will be needed, and built from the factors that will drive returns there, not from the ones that drive returns now.

Counting the Bets You Actually Have

The number of positions in a portfolio is easy to count and nearly meaningless. Twenty positions with pairwise correlations near zero are twenty bets; twenty positions that all load on the same factor are one bet made twenty times. What matters is the number of independent sources of variance, and that is a property of the covariance matrix, not of the position list.

A convenient way to count them is through the eigenvalues of the correlation matrix. Each eigenvalue measures how much of the portfolio's variance is carried by one independent direction of co-movement. If the variance is spread evenly across many directions, the portfolio has many effective bets; if it is concentrated in one or two, it has few. The entropy of the eigenvalue distribution turns that intuition into a single number, the effective number of independent bets, which runs from one, when a single direction explains everything, up to the number of positions, when every direction explains an equal share.^[1]

Figure 1 tracks that number through a stylized stress episode for a portfolio of twenty equally weighted positions. The only thing that changes over the episode is the average pairwise correlation, which rises from a calm level to a stressed one over a few days and then decays back over a few months. Nothing about the positions changes. No one trades. The portfolio's nominal diversification is constant at twenty; in the model its effective diversification falls by roughly three-quarters within a week and takes the better part of a year to recover.

The dashed line is what a risk model calibrated on calm data would report throughout the episode, because that model does not know the correlations have moved until it has seen enough stressed days to re-estimate them. By then the drawdown has happened. This lag is the first reason diversification seems to disappear: it was never there in the amount that was being reported, and the report was the last thing to find out.

One Factor Where There Were Many

The same collapse can be seen from a different angle by asking how much of a portfolio's variance the single largest eigenvalue explains. In a principal-components decomposition that eigenvalue belongs to the first component, the direction along which the positions move together most, and its share of total variance is a direct measure of how much of the portfolio is, in effect, one trade.

Figure 2 compares that share in calm and stressed regimes for four stylized portfolio designs, each of which would be described as diversified in a marketing document. The designs get broader from left to right: a single equity market, several equity markets, several asset classes, and several distinct strategies. In the calm regime the first component explains a modest fraction of variance in all four, and the broader designs look somewhat better than the narrow ones. In the stressed regime the first component explains half or more of the variance in every design, and the ordering barely survives.

Two things in the figure deserve emphasis. The first is that the broadest designs are not immune; they are merely less exposed. Diversification across asset classes and strategies helps under stress, but by a fraction of what its calm-regime statistics promise. The second is that the calm-regime numbers are the ones that get written down. Risk budgets, position limits, and leverage are usually set against a covariance matrix estimated from recent, mostly calm, history, which is precisely the regime in which the first component looks smallest.^[2]

The Factor Nobody Priced

Why should correlations rise together across positions that share no obvious economic driver? Part of the answer is that stress raises the importance of drivers that were always there: discount rates, risk appetite, the dollar. But the part that matters most for a portfolio built to be diversified is a factor that is barely present in the return data in calm conditions, because it is barely expressed there. That factor is liquidity: the cost of turning a position into cash quickly.

In normal markets liquidity is cheap and roughly constant, so it contributes almost nothing to the variance of returns, and a factor model fitted to normal data assigns it almost no weight. Under stress it becomes the dominant driver. Leveraged holders receive margin calls and sell what they can sell rather than what they want to sell; the positions that are easiest to liquidate fall first and furthest regardless of their fundamentals; volatility-targeting and stop-loss rules across many unrelated portfolios fire at the same time, and the selling they trigger raises volatility further. The mechanism has been described formally as a liquidity spiral, in which market liquidity and funding liquidity deteriorate together and reinforce each other.^[3] Its portfolio-level signature is simple: positions that never moved together begin to, and the sign is negative.

Figure 3 shows the signature in a stylized simulation. Three hypothetical books, an equity long/short book, a credit relative-value book, and a cross-asset carry book, have idiosyncratic return streams that are unrelated to one another, plus a common exposure to a single liquidity factor. In the calm period the liquidity factor is quiet and the three drawdown paths look independent; a correlation estimated over those months would be close to zero. When the factor moves, all three draw down at once, and the equal-weighted blend of the three, which was supposed to be the diversified choice, draws down nearly as far as the average of its components and further than one of them.

The blend's calm-period drawdowns are less than half as deep as its components', which is exactly what diversification is supposed to do and exactly why the blend looks so good in a backtest. Its stress drawdown is nearly as deep as the average of theirs, because the thing that caused it was common to all three. Diversification did not fail; it was never applied to the factor that mattered. The factor was invisible in the data the portfolio was built from.

A Field Guide to Hidden Common Exposures

Liquidity is the most important hidden exposure but not the only one. Figure 4 lists the forms of diversification that portfolios most commonly claim, the exposure that tends to connect them under stress, and what the connection looks like when it is expressed. The pattern in every row is the same: the axis along which the positions were diversified is real, and the axis along which they converge is a different one, usually related to who holds them and how those holders are financed rather than to what the positions are.

The last two rows are worth pausing on. Diversifying across managers is often treated as the strongest form of diversification because it diversifies process as well as position. But managers who use the same data vendors, the same risk model, and the same liquidation rules will de-risk on the same day, and the day they do so is the day their positions become correlated whatever the positions are. Diversifying across time, by entering positions gradually, protects against the risk of a single bad entry point; it does nothing against a common exit trigger, and volatility targeting is a common exit trigger shared by a large part of the systematic industry.

Building for the Correlations You Will Get

The practical response is not to abandon diversification but to measure it in the regime in which it will be tested. Several practices follow, none of them novel and all of them easy to neglect.

• Estimate two covariance matrices, not one: a calm-regime matrix from recent data and a stressed-regime matrix from the episodes in which correlations converged, and size positions so that the portfolio is acceptable under both. The stressed matrix will be noisier, because there are fewer stressed days to estimate it from, and it should be shrunk heavily toward a simple structure; a noisy estimate of the right regime is more useful than a precise estimate of the wrong one.^[4]
• Count effective bets under the stressed matrix. A portfolio that has twenty effective bets in calm markets and four in stressed ones should be described, and leveraged, as a four-bet portfolio.
• Put liquidity in the risk model as a factor, with exposures measured by how each position would be sold under pressure rather than by what it is. Two positions with the same liquidation profile share a factor whether or not their returns have ever been correlated.
• Treat crowding and financing as exposures. Who else holds a position, how they are financed, and what would make them sell are properties of the position, and positions that share those properties share a risk.
• Test the exit as carefully as the entry. Any rule that reduces exposure when volatility rises is a rule that sells into a falling market alongside everyone else who uses it. The question is not whether to have such a rule but whether its trigger and its pace are set with that company in mind.

None of this makes correlations stop rising under stress. They will rise, and a portfolio that depends on their not rising is a portfolio with an unmeasured exposure. The goal is narrower and more achievable: to know, before the episode, how many bets the portfolio will contain once it starts, and to have sized the portfolio for that number rather than for the one on the calm-weather report.

Diversification, properly understood, is not a count of positions but a claim about the covariance matrix in a regime you have not yet seen. It has to be earned in that regime, by finding the exposures that are shared there and either hedging them, limiting them, or accepting them knowingly. The portfolios that hold up are not the ones with the most positions. They are the ones whose owners knew which of the positions were the same bet.