Stress-testing 'uncorrelated': what we look for before calling an asset truly diversifying

"Uncorrelated" is one of the most overused and undertested words in alternative investing. Almost every manager presenting a non-traditional strategy claims some version of it. Most of those claims rest on a single number — a historical correlation coefficient versus the S&P 500 — that is doing far more work than it can actually support.

This article describes what a more serious test of uncorrelation looks like. The intent is not to argue that any particular asset class fails such a test, but to give an allocator a framework for asking better questions and to be explicit about how we think about the question ourselves.

What a correlation coefficient measures, and what it misses

A correlation coefficient is a single number summarizing the average linear relationship between two return series over the sample period over which it is calculated. It is a useful first-order statistic. It is also, for any genuinely interesting question about diversification, an incomplete answer.

Three things a correlation coefficient does not tell you.

First, it does not tell you whether the relationship is stable across regimes. A correlation that averages to a low number over a full market cycle can hide periods of much higher correlation during specific kinds of conditions — rate spikes, credit selloffs, equity drawdowns. Two return series can have an unconditional correlation near zero and still co-move sharply during the periods that matter most to risk management.

Second, it does not tell you whether the relationship is linear. A correlation coefficient is built around the assumption of an approximately linear joint distribution. Real markets have asymmetric tail behavior: assets that look uncorrelated in normal times can become tightly correlated in extreme moves. Tail dependence is a different statistical object from correlation, and a low correlation coefficient is fully compatible with high tail dependence.

Third, it does not tell you whether the relationship is causal or whether it reflects a hidden common factor. Two assets can have low unconditional correlation while both being driven, in part, by the same underlying factor. When that factor moves, the correlation between the assets spikes.

Regime-conditional correlation

The most basic refinement of unconditional correlation is to split the sample into regimes and compute correlations within each. The choice of regime depends on the question. Common choices include rate regimes (rising, stable, falling), credit regimes (tightening, widening), equity regimes (drawdown, bull market, sideways), and volatility regimes (high VIX, low VIX).

A well-tested uncorrelated asset will exhibit low correlation in each of these regimes, not just on average. The interesting question for any manager claiming uncorrelated returns is what happens to the relationship during the worst quintile of equity returns, the worst quintile of credit moves, the periods when rates rise sharply, and the moments when the unconditional volatility regime shifts. A pleasant unconditional correlation is not a substitute for performance in those specific conditions.

Tail correlation and liquidity correlation

Two refinements deserve particular attention because they are where most diversification claims fail under stress.

Tail correlation

Tail correlation is the co-movement of assets specifically in their joint tails — the events that most affect portfolio outcomes during stress. The published academic literature on tail dependence, including work by researchers focused on extreme value theory and copula-based modeling, has consistently shown that assets which appear low-correlation under normal conditions can exhibit much higher dependence in their joint tails. This is the structural reason why portfolios that look well diversified on a covariance-matrix basis can experience much larger drawdowns than expected during severe market events.

Practical implication: any uncorrelated-asset claim should be stress-tested against equity drawdowns of 30 percent or more, credit-spread widenings of several hundred basis points, and persistent shifts in the equity-bond correlation. A serious manager will have a view on how their strategy performs in each of these scenarios.

Liquidity correlation

Liquidity correlation is a related but distinct phenomenon: the tendency for marketability across assets to collapse together during broad selloffs. The pattern is well documented across crises in 1998, 2008, and 2020. Many alternative strategies that show low marked-to-market correlation in normal times become functionally correlated under stress because their owners need cash and force-sell whatever can be sold. The asset that cannot be sold at modeled value during a crisis is, for that crisis, correlated with the assets that triggered the forced selling.

Allocators should distinguish two questions. First, would the asset's fundamental cash flows continue to be uncorrelated with equity beta during a crisis? Second, would the asset retain its marketability during a crisis, or would forced selling create a co-movement that did not previously exist? These are different questions with potentially different answers.

Factor decomposition: hidden exposures in supposedly uncorrelated assets

The factor-investing literature, much of it developed by researchers including Cliff Asness, Antti Ilmanen, Andrew Ang, and their colleagues, has demonstrated repeatedly that many alternative strategies marketed as uncorrelated turn out, on examination, to load heavily on conventional factors. Long-short equity strategies often have substantial market beta. Many hedge fund strategies exhibit credit beta, volatility beta, or value/momentum factor loadings. The relevant question is not "what is the unconditional correlation?" but "what factors does the strategy actually load on, and how does that loading change with conditions?"

Asness in particular has written for decades about the importance of looking through asset class labels to the underlying factor exposures. His 2001 work on hedge fund correlation showed that average hedge fund returns exhibited a correlation to long-only equities of over 0.8 — a result that does not square with the marketing description of the category as uncorrelated. Ilmanen's book Expected Returns, published in 2011, developed a comprehensive framework for thinking about alternative assets as bundles of compensated factor exposures rather than as monolithic categories.

The discipline that follows from this literature is to decompose any alleged uncorrelated return stream into its factor components and ask which factors it loads on. A truly diversifying asset will load on factors that the rest of the portfolio does not already own. An asset that loads on equity beta dressed up in alternative clothing is not diversifying; it is the same exposure with a higher fee.

A working framework: six questions for any uncorrelated claim

Drawing these threads together, a practical framework for evaluating any uncorrelated-asset claim looks something like the following.

• •What is the underlying risk factor? Is it genuinely distinct from the macroeconomic and market factors that drive the rest of the portfolio, or is it a rebranding of an existing factor?
• •How does the relationship behave across regimes? Is the correlation low across rate regimes, credit regimes, equity regimes, and volatility regimes, or is the low average concealing high correlation during specific conditions?
• •What is the tail behavior? When equities draw down sharply, when credit spreads widen, when volatility spikes, how does the strategy behave?
• •What is the liquidity behavior? Does marketability hold up under stress, or do the assets become non-tradable precisely when liquidity matters most?
• •What are the factor exposures? When the strategy is decomposed into systematic factor loadings, what comes out, and is the residual large enough to warrant the description uncorrelated?
• •Is the uncorrelation structural or temporary? Does it rest on a permanent feature of the underlying risk (e.g., mortality outcomes are not driven by interest rates) or on a market regime that may not persist?

A manager who has thought carefully about these questions will be able to answer them. A manager who has not will tend to deflect.

How insurance-linked assets stand up to this framework

It is worth applying the framework to the asset class we work in, because no fair argument exempts insurance-linked investing from the same scrutiny. The analysis is structural; we are not making numerical claims.

On the first question — the nature of the underlying risk factor — life-insurance-linked assets stand on relatively strong ground. The principal driver of cash flow in life settlements is the timing of individual mortality outcomes. Mortality is a biological process whose distribution is governed by age, sex, health, and lifestyle, and whose timing for any individual is not, in any material way, a function of interest rates, equity markets, or credit spreads. The structural argument for low correlation rests on this fact, not on any historical regression.

On the second question — regime conditioning — the structural argument continues to hold. A recession does not change when a given insured will pass away. A rate shock does not directly affect mortality outcomes. The cash flows are governed by contracts with regulated insurance carriers and triggered by mortality events that occur on their own timetable. Where the structural argument needs qualification is the discount rate at which those cash flows are valued: marked valuations of a life settlement portfolio do depend on the discount rate applied, which in turn does depend on broader financial conditions. The realized cash flows are uncorrelated with macro variables; the marked-to-model valuation has some exposure to the rate environment used in valuation.

On the third question — tail behavior — the same structural logic applies. The mortality outcomes that determine ultimate cash flows are not concentrated in market crises. There is no obvious mechanism by which equity drawdowns of 30 percent would accelerate or delay mortality in a way that produces tail dependence with the equity market. The marked-to-model valuations may exhibit some dependence on the discount-rate regime, but the underlying cash flows do not.

On the fourth question — liquidity behavior — life-insurance-linked assets fail in a particular way that deserves honest acknowledgment. The asset class is illiquid in all environments, and more so in stress environments. Secondary market pricing during a broad selloff would likely be poor relative to modeled hold-to-maturity value, because the universe of forced sellers expands and the universe of patient buyers contracts. Holders who are not forced to sell will see modeled values continue to develop along their actuarial trajectory; holders who are forced to sell will experience a correlation with equity drawdowns through the liquidity channel that does not exist in the underlying cash flows. The asset class is not a liquid hedge against market stress, and presenting it as one would misrepresent the structure.

On the fifth question — factor exposures — the most honest answer is that life-insurance-linked assets carry exposure to a distinctive longevity/mortality factor and to a carrier credit factor (the obligation of the insurance carrier to pay the death benefit when the insured passes away). Neither of these factors is heavily owned by typical institutional portfolios, which is the structural source of the diversification benefit. Discount-rate sensitivity in marked valuations is a real but secondary exposure.

On the sixth question — whether the uncorrelation is structural or temporary — we believe the answer is structural. The argument rests on the nature of the underlying risk, not on any particular sample period or market regime. We do not believe the realized cash flows of life-insurance-linked assets will become correlated with equities in any future regime, because the link to mortality outcomes is biological rather than financial.

A note on humility

It is worth a brief, explicit note on humility. The history of alternative investing is full of strategies that looked uncorrelated until they were not. Convertible arbitrage in 1998. Mortgage credit in 2007. Risk parity at various moments. In each case, low historical correlations had been built up during periods that did not include the regime in which the correlations later changed. A patient observer of the alternative-investments industry will be appropriately skeptical of any new claim that a particular asset is structurally different.

Our posture is that life-insurance-linked assets are structurally different in a specific, narrow, defensible sense: the underlying cash flow risk depends on biological mortality outcomes, which are independent of financial markets. We are also clear that this structural argument is not a guarantee against marked-to-model volatility, illiquidity correlation under stress, or the surprise we should always expect from history. The discipline is to hold the structural view and the humility together.

A short closing

The word "uncorrelated" is overused. A more useful conversation between an allocator and a manager involves regime-conditional behavior, tail dependence, liquidity behavior under stress, and factor decomposition. The framework above is one we apply to the asset classes we work in; we encourage allocators to apply it everywhere, including to our own strategies.

For a fuller treatment of how the three insurance-linked asset types we focus on actually generate cash flows, see our overviews at seapoint.capital/strategy/life-settlements, seapoint.capital/strategy/collateralized-loans, and seapoint.capital/strategy/payout-annuities.