Stock-Bond Correlations

How to invest when stocks and bonds are positively correlated
 

By: Greg Obenshain and Noah Kreutter

For the past 30 years, the correlation between stocks and bonds has been negative. But last year, the trailing three-year correlation turned positive for the first time since November 2000. This is something that many investors today have never experienced in their professional lives.

Figure 1: 3Y Trailing Monthly Stock-Bond Correlations, 1929–2023

Source: SBBI Ibbotson US Large Stocks and US LT Govt until 1989. Bloomberg thereafter.  
What caused this apparent shift? We believe the proximate cause was the spike in inflation after COVID stimulus, which forced the Federal Reserve to raise rates, which in turn triggered the biggest bond market sell-off ever and repriced US growth stocks, whose valuations depended on low discount rates.

But the specific details are part of a general pattern. According to a recent AQR study, a key determinant of the stock-bond correlation is the relative dominance of growth uncertainty and inflation uncertainty. Stocks and bonds react to growth shocks in opposite ways: stocks go up and bonds go down. But stocks and bonds react in the same direction to inflation shocks: stocks go down and bonds usually go down more.

Note that it’s not the level of inflation that matters; it’s the dispersion of changes. AQR found that inflation volatility is the most statistically significant variable in their model, while the level of inflation is statically insignificant. And last year saw a major spike in inflation volatility. Below we show the trailing one-year standard deviation of the year-over-year inflation rate and the trailing three-year average.

Figure 2: Inflation Volatility Is Rising

Source: Bloomberg, Verdad Analysis

Given these recent developments, what sort of correlations might we as investors project into the future? At least in the short term, we believe higher stock-bond correlations are likely to persist, for the simple reason that stock-bond correlations tend to exhibit some level of autocorrelation. Last month’s correlation between stocks and bonds is predictive of next month’s correlation, with next month's correlation tyically having the same sign as the previous month. This autocorrelation has been around 0.44 since 2002, with negtive correlations usually persisting from one month to the next, and positive correlations typically persisting from one month to the next, as shown below.

Figure 3: Stock-Bond Correlation, 30D Trailing vs 30D FWD, 2002–2022

Source: Bloomberg, Verdad Analysis
 
The drivers of correlations are also autocorrelated: inflation volatility and growth volatility tend to persist period over period. As a result, we believe that the most effective and parsimonious way to model this is simply to weight the recent past more heavily when computing one’s correlation matrix. This is for three main reasons. First, the fundamental drivers that AQR and others posit tend to be difficult to measure in real time. Second, correlation forecasting is a multidimensional problem—unlike return forecasting or variance forecasting—so there is not a robust, obvious way to specify a multifactor correlation model that forecasts a correlation matrix, as opposed to a single pairwise correlation. Finally, we believe that any relationship between external predictive features and a single pairwise correlation is itself nonstationary: the meta-rules that govern correlation regimes also change over time.

We believe exponentially weighting the recent past, with a longer half-life than one might use to compute a variance forecast, is the most parsimonious and elegant way to account for these factors. And in our internal testing, we have seen large improvements from switching to a time-weighted correlation matrix.

So, if we use the higher stock-bond correlation for our investing decisions, what are the implications for asset selection? Historically, higher stock-bond correlations have meant that bonds are less likely to offset stock declines and that long-duration bonds are less attractive. It also increases the likelihood of inflationary outcomes that favor commodities such as oil.

Below we show a table with returns to various asset classes in both high and low stock-bond correlation environments, normalized by historical volatility. We classify regimes by the standard deviation of the trailing stock-bond correlation, measured with a 30-day half-life, with an absolute deviation of 0.5 or greater needed to mark a positive or negative correlation regime. While the effect is not the strongest, it shows that neutral and positive-correlation environments tend to be good for equities and oil, while Treasurys and gold benefit from negative-correlation environments.

Figure 4: Sharpe Ratios by Asset Class and Regime, 1990–2023

Source: Bloomberg, Verdad Analysis

We cannot predict how assets will move relative to one another with any sort of certainty. But we can observe today’s environment and make shifts in our asset allocation to reflect recent changes. In today’s environment, stock-bond correlations have moved dramatically upward. That is enough to shift our view on optimal asset allocation at the margin.