Correlation Is Not Constant: What Happens to Diversification When the Regime Changes?

The number that makes a portfolio look diversified is the one least likely to hold when diversification is needed. A correlation coefficient summarizes how two assets moved together over some window of the past, and portfolio construction treats that summary as if it described a stable property of the assets themselves. It does not. Correlation is the outcome of a contest between the different kinds of news that move prices, and when the balance of that contest shifts, so does the sign. The negative stock-bond correlation that anchored balanced portfolios for two decades was not a law; it was a regime.

This piece is about how cross-asset correlation structures change, and why they change in the particular ways they do. We work through four kinds of transition: volatility shocks, monetary-policy transitions, liquidity events, and crises. All of the figures are stylized simulations built from a small factor model whose parameters are stated under each chart, so the shapes are ours to choose; the point is the mechanism, not the magnitudes. The conclusion is that a portfolio process has to treat correlation as a state variable, something to be estimated conditionally, stressed deliberately, and diversified against, rather than as a constant to be plugged in.

Two Assets, One Number, Five Regimes

The cleanest place to see the problem is the correlation between equities and government bonds, because it is the single number most balanced portfolios rely on most heavily. When the news that moves markets is about growth, the two assets move in opposite directions: a weaker economy lowers expected earnings and lowers expected policy rates, so equities fall and bonds rally. When the news is about inflation and the policy response to it, both assets are discounted at the same rising rate and fall together. The sign of the correlation is therefore a statement about which kind of news currently dominates, and nothing more.

Figure 1 makes this concrete in a stylized simulation. Weekly stock and bond returns are generated from three common factors, growth, rates, and liquidity, with loadings that stay fixed throughout; what changes across the shaded regimes is only the volatility of each factor. The solid line is the trailing six-month correlation an observer would measure from the simulated returns; the dashed step is the correlation implied by the factor volatilities in each regime. The observer sees the truth late, with noise, and rarely at the level the model implies, which is the condition every portfolio manager works under.

Three features of the picture carry over to real markets. The measured correlation lags every transition by roughly half its window, so a rolling estimate is always describing the regime that is ending. The sign is stable for long stretches and then reverses within months, which is why an unconditional average across the whole sample describes no period the portfolio actually lived through. And the excursions that matter most for a diversified portfolio, the moves toward positive correlation in the policy and liquidity regimes, arrive precisely when the equity allocation is also under the most pressure.

What Moves Under the Surface

The stock-bond pair is one cell in a matrix, and the same mechanism operates on all of the cells at once. Every asset loads on the same handful of common factors with a different set of weights: equities load positively on growth and negatively on rates; government bonds load negatively on both; credit sits between the two; commodities load on growth and, unlike most risk assets, positively on rates; the dollar loads negatively on growth and positively on the demand for liquidity. In calm conditions the growth factor supplies most of the variance, and the correlation matrix reflects the pattern of growth loadings. Figure 2 shows the matrix our stylized model implies for a calm, growth-led regime.

Now hold every loading fixed and raise the volatility of a single factor: the demand for liquidity, the factor that describes what happens when many participants need to raise cash at the same time and sell whatever can be sold. Figure 3 shows the result. The matrix has not been re-estimated with different assets or a different sample; it is the same model with one factor's variance raised. Yet nearly every off-diagonal entry has moved toward the same value, the safe assets have joined the risk assets, and the only entries that remain negative are the ones involving the asset that liquidity demand flows into, the dollar.

This is the mechanism behind the observation, common to every account of financial stress, that correlations go to one in a crisis. They do not literally go to one, and they do not rise because the assets have changed. They compress because a factor that is normally quiet has become loud, and every asset that loads on it, which is nearly all of them and with the same sign, inherits its movement. The diversification that existed in the calm matrix was diversification across growth exposures. It was never diversification against a liquidity shock, because almost nothing in a conventional portfolio is short that factor.^[1]

Correlation Rises With Volatility

If stress raises the variance of a factor that most assets share, then the average level of correlation across the matrix should rise with the level of volatility, and it should do so nonlinearly, because the shared factor's contribution to variance grows relative to the idiosyncratic component. Figure 4 plots that relationship in a stylized set of monthly observations: each point is a month, its horizontal position is the realized volatility of equities in that month, and its vertical position is the average pairwise correlation across the eight assets. The dashed line is a straight fit through the points; the curvature the points show around it is the part the fit misses.

Two cautions apply to any chart of this shape, simulated or measured. The first is that correlation estimated over a high-volatility window is mechanically biased upward even when the true relationship is unchanged, because large observations dominate the sum of cross-products; part of the slope in any empirical version of Figure 4 is estimation, not economics.^[2] The second is that the relationship is not symmetric in time. Correlations rise faster on the way into stress than they fall on the way out, since the deleveraging that raises them happens quickly and the rebuilding of diversified positions happens slowly. A model that treats correlation as a smooth function of volatility will be late in both directions.

Neither caution changes the conclusion that matters for portfolio construction: the correlation matrix a portfolio faces in a drawdown is not the one it was optimized on. An allocation whose risk is computed from a calm-regime matrix understates its exposure to the shared factor by construction, and the understatement is largest for the portfolios that appear most diversified, because those are the ones whose apparent safety depends most on the off-diagonal entries.

Four Ways a Regime Turns

The transitions in Figure 1 were drawn as a sequence, but in practice they are distinct mechanisms with different speeds, different signatures, and different implications for what still diversifies. Figure 5 sets out the four that recur most often in accounts of market stress, together with the calm regime they depart from.

A volatility shock is a repricing of risk itself. Something, an unexpected data release, a large position being unwound, a break in a widely used hedge, raises implied and realized volatility sharply, and levered holders of risk are forced to cut exposure regardless of the fundamental news. The stock-bond correlation moves toward zero and can briefly turn positive as bonds are sold to meet equity margin calls, then returns to its prior sign within weeks. Episodes such as the May 2010 flash crash are the extreme case: a shock that unfolds and largely reverses within a single session, with correlation structures that are meaningless at any horizon longer than the event itself.

A monetary-policy transition is slower and more durable. When inflation replaces growth as the dominant uncertainty, the policy rate becomes the variable every asset is discounted by, and the correlation between equities and bonds turns positive and stays there for as long as the regime lasts. The tightening cycle that began in 2022 was broadly described this way: the two halves of the balanced portfolio fell together for an extended stretch, and the diversification that had been assumed for two decades was not there to be had. In the stylized model, what still diversifies in a policy regime is exposure to the rates factor with the opposite sign; beyond the model, the natural candidates are trend across asset classes and relative-value positions that do not depend on the level of rates at all.

A liquidity event is the fastest and the least discriminating. When enough participants need cash at once, the question of what an asset is worth is replaced by the question of what it can be sold for today, and the answer is worse for everything at the same time. The distinguishing signature is that the safe assets sell too: government bonds, gold, and high-grade credit fall alongside equities, and the only thing that reliably rises is the currency the cash is needed in. March 2020 is commonly described as unfolding this way over a period of weeks before central-bank intervention restored the ordinary pattern. Figure 3 is a stylized picture of that kind of episode.

A crisis proper, a collapse in growth expectations with rising default risk, looks different again. The growth factor becomes loud rather than the liquidity factor, and the stock-bond correlation becomes strongly negative: equities and credit fall, government bonds rally. This is the regime in which the balanced portfolio works as designed, and the reason it was designed that way. The failure mode is the transition, not the destination. Most crises begin with a liquidity phase in which the bond hedge fails, and only afterward settle into the flight-to-quality pattern in which it works.^[3]

What This Means for How We Build Portfolios

Taking correlation seriously as a state rather than a constant changes several things about the portfolio process. The first is estimation. We do not think a single covariance matrix, however carefully shrunk toward a structured target, is the right object to hand an optimizer. The object is a set of matrices conditioned on the regime, together with an estimate of which regime the market is in and how likely it is to change, and the portfolio that results is the one that is acceptable across the set rather than optimal under one member of it.^[4]

The second is stress testing. The stress that matters for a diversified portfolio is not a larger draw from the calm matrix; it is a draw from the stressed one, in which the off-diagonal entries have moved. Scaling volatility up while holding correlations fixed produces a comfortingly linear loss estimate that has little to do with what happens. Scaling the liquidity factor's variance, as Figure 3 does, is closer to the experience, and it tends to show that the portfolio's true exposure is not the sum of its asset exposures but its net loading on a single factor it never chose to own.

The third is what counts as diversification at all. Diversifying across assets that share a factor loading is diversification only in the regime where that factor is quiet. Diversifying across mechanisms, across positions that respond to growth, to rates, and to liquidity with different signs, is intended to hold up across regimes, and it is the reason a systematic portfolio tends to hold positions that look unrelated to one another. They are related, through the factor structure, in a way designed to survive the moment the structure shifts.

None of this makes correlation predictable. It makes the portfolio's dependence on any particular correlation explicit, which is the most that can honestly be asked. A balanced portfolio that knows it is a bet on the sign of the stock-bond correlation is in a better position than one that believes it is not making a bet at all.