Excerpted from a whitepaper by One River Asset Management's Chase Muller and Patrick Kazley titled: Regime Change Resilience - Rebooting Risk Mitigation with Structural Correlation.
After decades, you recognize patterns. The biggest winners and losers in each cycle tend to be younger. Unburdened by the past, open to change, they often lack fear. Older folks who remain standing are either lucky or attained some wisdom, acquired at great cost. Pain. The most honest of those live in the fear that they have gotten lucky, and it will run out. Having recognized the impossibility of knowing the future, and knowing each cycle contains some new surprise, they surround themselves with younger people, blending the strengths of young and old.
Beginning in 1962 when the daily bond time series is available and going through today, the correlation between stocks and bonds is slightly negative (-0.1 correlation). The t-statistic, or level of reliability of that full sample observation, is highly statistically significant with a -7 t-stat, where a t-statistic of approximately +/- 2.5 or larger is typically considered statistically significant. The t-stat being much larger than that makes it very unlikely to be a spurious finding over the sample period.
However, if you divide this timeframe into different periods, the apparent consistency and reliability of this observation changes drastically. From 1962-1981, when US interest rates went from historic norms to record highs, the correlation between bonds and equities inverts and is positive (+0.2 correlation). Thus, in Oct 1981 when interest rates reached their secular peak, if you had used a backward-looking risk model to estimate cross-asset correlations or build a risk mitigation portfolio, you would have assumed equities and bonds were positively correlated, and indeed the significance of that relationship would have been entirely supported through a statistical lens (+13 t-stat).
Naturally, you might be tempted to look at these results and conclude the relationship between equities and fixed income is indeed reliable, as long as you control for the rising or falling rate environment. However, the relationship and changes to it are not as easily predicted by a single factor such as the general drift of interest rates over time. To illustrate this, from Oct 1981 – Oct 1998 when rates collapsed from highs, the relationship between stocks and bonds was also positive with a higher level of consistency (+0.2 correlation, with a +16 t-stat).
Lastly, the 1998-present period resulted in a -0.4 correlation between stocks and bonds, with a highly significant -30 t-stat. What we have not explored here, but is also worth highlighting at least in passing, is the potentially undesirable conditional correlation that can accompany transitory relationships. Even an assumed relationship that holds on average over longer time frames can break down in extreme risk-off events and lead to deeper drawdowns and more short-term pain. March 2020 was a such a case of risk assets concurrently declining and transitory correlations breaking down when they were needed most.
Using backward-looking returns to justify cross-asset correlation expectations might yield convincing statistics, but ultimately this approach has not proven to be a fully reliable method of sourcing correlation estimates essential for proper risk mitigation and diversification. Indeed, without properly matching a statistical observation with an intuitive linkage, you run the risk of relying on ephemeral relationships for stability. This raises a question: If forward-looking allocation models based on historical returns are only valid in a world of relatively static cross-asset relationships, how does an allocator find reliable sources of diversification in the face of regime changes?