What Survives a Crisis?: Stress-Testing the Assumptions a Portfolio Rests On

Every portfolio is an argument about what will stay true. The argument is rarely written down. It lives in the covariance matrix, in the choice of which assets are called hedges, in the assumption that a position can be reduced at something close to the last price. In ordinary markets the argument is never tested, because the relationships it relies on hold well enough that nobody asks. A crisis is the moment the argument is read aloud, and it is often the first time its author hears what it says.

This piece asks which of those relationships survive, and which collapse, when the crisis arrives. We do it in a stylized model rather than with history, partly because history offers too few crises to separate their types cleanly, and partly because the point is the mechanism rather than the magnitude. We take nine assumptions a diversified portfolio commonly rests on and expose each to four archetypal crises: a credit crisis, a liquidity shock, an inflation shock, and a rate shock. The conclusion is uncomfortable in a useful way. Almost nothing survives every archetype, and what does survive is not the set of relationships portfolios are usually built on.

Four Archetypes, Four Channels

Crises are not interchangeable. The financial crisis of 2008 was a credit crisis: it began in the quality of collateral, and once the balance sheets holding it were forced to deleverage, the acute phase ran over weeks and faded over months. March 2020 was a liquidity shock: the same kinds of balance sheets sold whatever could be sold, over weeks, and the selling stopped as soon as funding was restored. The inflation of 2021 and 2022 was an inflation shock: no single event, but a slow repricing of what a nominal cash flow is worth, over many months. The bond market episodes of 1994 and 2013 were rate shocks: the price of money was repriced over months, slowly enough that positions built on the old price were unwound into it rather than out from under it. Each is a different disease with a different tempo, and each acts on a portfolio through a different channel.

We model the four channels directly. A credit crisis widens spreads; a liquidity shock withdraws funding; an inflation shock raises the price level unexpectedly; a rate shock raises real rates. Each archetype delivers a full dose of its own channel and smaller doses of the channels it spills into, because credit fear becomes liquidity fear and inflation fear becomes rate fear. Each assumption is then characterized by how sensitive it is to each channel, and its survival score under an archetype is the exponential of minus the sensitivity-weighted dose it receives.^[1] Figure 1 shows the result: nine assumptions down the side, four archetypes across the top, and a score from zero, meaning the relationship collapsed, to one, meaning it held.

The pattern is more informative than any cell. Reading across, four assumptions fail in every archetype, two hold in every archetype, and three hold in some and fail in others; it is the last group that decides whether a portfolio built for one crisis survives a different one. Reading down, the columns pair up. Credit and liquidity crises weaken the same things, diversification across regions and sectors, stable trading costs, and carry, and spare the same things, the bond hedge and the dollar. Inflation and rate shocks are gentler on the first group and turn on the second. The most consequential division in the table is therefore not between assumptions but between fast crises and slow ones, and a portfolio built for one kind is, in this model, built against the other.

The Haven Depends on the Question

The clearest example is the relationship on which a great deal of portfolio construction rests: that government bonds rise when equities fall. In the model it is among the most reliable relationships in a credit or liquidity crisis, because both are episodes of falling growth expectations and falling policy rates, and among the least reliable in an inflation or rate shock, because both are episodes in which the discount rate rises for every asset at once. The relationship does not weaken in the second case; it inverts. Figure 2 traces a stylized rolling correlation between equities and bonds through each archetype, in trading days from onset.

Three features of the paths deserve attention. In the credit crisis the correlation becomes more negative than in calm markets, which is the hedge working as intended, and it fades slowly. In the liquidity shock it jumps toward zero and briefly above it, because for a few days everything liquid is sold to raise cash, and then snaps back once funding returns; the hedge fails exactly when it is needed most and recovers before most investors have decided what to do about it. In the inflation and rate shocks the correlation climbs into positive territory and stays there for months; in the inflation shock the drift begins well before onset, which is how slow shocks announce themselves. A bond position sized as a hedge on the strength of the first two paths is, in the second two, a second source of the same loss.^[2]

Gold behaves in something like the mirror image, which is why the first two rows of the heatmap look so different. Gold holds up in a credit crisis, is sold in a liquidity shock along with everything else that can be sold, and loses its footing whenever real rates rise, which is what an inflation shock eventually forces and what a rate shock does immediately. The dollar is the nearest thing in the model to an all-weather haven: it is bid in both fast crises and in a rate shock, and it wobbles only when the inflation is domestic. The lesson we draw is not that one haven is better than another but that the word haven is incomplete without a clause saying from what.

What the Construction Choices Did

Assumptions are not held in the abstract; they are held through construction choices, and it is the choices that produce the loss. Figure 3 takes eight common choices and asks how much of a baseline book's stylized stress loss each one avoided, or added, averaged over the two fast archetypes and over the two slow ones. Each choice is tied to the assumption in Figure 1 that it relies on and scaled by the share of the book's risk it moves. Two of them, a leverage cap and a liquidity buffer, are defensive choices that pay off in proportion to how badly the assumption they insure against fails.

The chart divides into three groups. The defensive choices, the leverage cap and the liquidity buffer, help in every archetype and help most in the fast ones, because they do not rely on any relationship holding; they rely only on the crisis being a crisis. Volatility targeting helps in both regimes and more in the slow ones, because volatility clusters, but a fast crisis arrives as a jump before the cluster has formed, so the targeting is late. The reliant choices are the interesting group. The duration hedge is the most useful choice in the fast archetypes and the most harmful in the slow ones; the trend overlay is the reverse, cut by a fast reversal and paid by a slow repricing; carry and cross-region diversification hurt in both, since the assumptions they rely on hold in no archetype.

What stands out is how large the swings are relative to the levels. The average effect of the duration hedge across the four archetypes is close to zero; its effect within either group is the largest in the chart. Averaging across crises, which is what a long-sample backtest does, produces a number that describes no crisis anyone will live through. This is the practical reason we prefer to evaluate a construction choice archetype by archetype, and to ask of each hedge not whether it works but in which crisis it works and in which it becomes exposure.

How Each Assumption Fails

A survival score records that a relationship broke; it does not say how. The mechanism matters because it determines what can be seen in advance. Figure 4 lists, for each of the nine assumptions, the mechanism by which it fails in the model, the archetypes in which the failure is severe, and the observable that tends to move first.

Two things in the table are worth stating plainly. First, the failures that matter most are the ones that arrive before the crisis is named. The bond hedge does not break on the day inflation is announced; it breaks over the months in which the realized correlation drifts upward while the assumption in the risk model stays where it was. Second, the observables in the last column are all cheap. None requires a forecast of the crisis, only a willingness to measure the relationship the portfolio depends on, continuously, and to act when the measurement disagrees with the assumption. A good deal of what survives a crisis, in our experience, is the habit of doing that.

What Survives

Figure 5 summarizes the model's scoreboard. Of nine assumptions, two hold in all four archetypes: the dollar's tendency to strengthen and the tendency of volatility to cluster. Four fail in all four: diversification across regions, diversification across sectors, stable trading costs, and carry. Three, the bond hedge, gold, and trend persistence, hold in some archetypes and fail in others, and those three are the ones that matter most, because they are the ones a portfolio can be built either for or against.

The relationships that survive share a property: they are statements about the behavior of stress itself rather than about which asset will be bid. Volatility clusters because forced selling begets forced selling, whatever the cause. The dollar is bid because so much of the world's borrowing is denominated in it, whatever the cause. Neither is a forecast of the crisis; each is a description of what crises do. The relationships that fail are the ones that bet, usually without saying so, on the cause.

Three practices follow. The first is to write the assumptions down, one per hedge, and to test each against the archetypes separately rather than against a pooled history in which the archetypes cancel. The second is to favor construction choices that do not rely on a relationship holding, a leverage cap, a liquidity reserve, position limits set against the failure case, and to accept that they cost something in calm markets, since that cost is the premium on the only insurance whose payout does not depend on diagnosing the crisis correctly. The third is to treat the conditional hedges, bonds, gold, trend, as conditional: sized for the archetype they protect against, monitored for the ones they do not, and reduced when the measured relationship leaves the range the sizing assumed.^[3]

None of this makes a crisis survivable by design. It makes the portfolio's argument explicit, so that when it is finally read aloud, nothing in it comes as a surprise.