Stock-bond correlation: regimes, drivers and what it means for 60/40, risk parity and hedging

Advanced12 min read · updated 21 October 2026 · 8 sources

The stock-bond correlation is the single most important number for the risk of multi-asset portfolios. From about 2000 to 2021 it was mostly negative: bonds rose when equities fell, making mixed portfolios calmer than the individual risks would suggest. In the 20th century, by contrast, it was mostly positive, and with the inflation of 2022 it rose again. Here you will learn what drives the sign, how much it changes 60/40, risk parity and hedging, and how uncertain a rolling correlation is.

Why this one number determines so much

The basics are explained in the article correlation and diversification. This article is about the question that has preoccupied allocators since 2022: what happens to a portfolio when the stock-bond correlation changes sign? For two assets with weights w, volatilities σ and correlation ρ:

The bonds' risk share at ρ = −0.3 is remarkable: 0.4%. The 40% of capital in bonds then acts almost like insurance that lowers equity risk without adding any of its own. If the correlation turns, their share rises to 14.8%. The chart shows the whole pattern across ρ, including for a risk parity portfolio with equal risk contributions.

A century of shifts: positive, negative, positive again

The negative correlation that many regard as normal is historically the exception. Ilmanen (2003) shows for the 20th century that the correlation between equities and government bonds was mostly positive, and negative only in the early 1930s, the late 1950s and around the turn of the millennium. Campbell et al. (2017) find a slightly positive average covariance with equities for US Treasuries from 1953 to 2014, unusually high in the early 1980s and negative in the early 21st century, especially in the downturns of 2001 and 2007 to 2009.

For the euro area and the US, the European Central Bank finds the same pattern: mostly positive from the late 1960s to the late 1990s, negative in the 2000s and recently higher in both economies (Mosk et al. 2022). Brixton et al. (2023) therefore call the correlation the “golden parameter” of traditional portfolios and warn that greater uncertainty about inflation could bring back the positive correlation of the 1970s to 1990s. In 2022, equities and government bonds fell at the same time for long stretches, following the pattern of those earlier decades.

Drivers: growth shocks versus inflation shocks

Why does the sign change? The explanation lies in the type of shock. A share is the present value of future earnings, a government bond the present value of fixed payments. Both depend on the discount rate, but equities also depend on earnings. Ilmanen (2003) puts it this way: growth and volatility shocks drive equities and bonds apart, while inflation shocks move them in the same direction via the common discount rate.

• Growth shock: the economy disappoints. Earnings expectations fall, and equities fall; the central bank cuts rates or investors flee to safety, and bonds rise. Result: negative correlation.
• Inflation shock: inflation surprises on the upside. Nominal rates rise, and bonds fall; higher discount rates and cost pressure also weigh on equities. Result: positive correlation.
• Monetary policy: if the central bank responds mainly to inflation, it amplifies the inflation channel; if it mainly fights recessions, it supports the growth channel. The ECB also sees the recent rise as a consequence of high inflation and rate hikes; the correlation was lowest when inflation was close to 2% (Mosk et al. 2022).

Which channel dominates depends on the relative volatility of the two shocks. A simple model with two independent shocks makes this visible; the loadings are illustrative figures, not estimates:

Research confirms this logic, but with caveats. Campbell et al. (2020) explain why the equity beta of US Treasuries was positive from 1979 to 2001 and negative from 2001 to 2011: the co-movement of inflation and the output gap had reversed. However, this only explains the data if risk premia change as well. Campbell et al. (2017) link the covariance to the term premium: the more closely inflation and the real interest rate move with the real economy, the higher the premium; its decline since the early 1980s lowered the modelled term premium of ten-year zero-coupon bonds by about two percentage points. Baele et al. (2010) urge modesty: in their model, macroeconomic fundamentals explain little of the correlation; liquidity measures and other factors play a greater role.

Implications for 60/40, risk parity and hedging

The correlation determines not only volatility but above all what bonds deliver in an equity crash. With jointly normally distributed returns, the expected bond return given an equity shock is linear in ρ:

Three consequences follow. First, 60/40 remains essentially an equity portfolio; the correlation decides whether the bonds are a buffer or ballast (see 60/40 portfolio). Second, risk parity depends on it particularly strongly, because the bond risk is levered and equally weighted; for two assets, the weights do not depend on ρ, but the leverage needed for a volatility target does. Third, hedging with government bonds only works reliably in growth crises. Against an inflation shock, short maturities, inflation-linked bonds, commodities or strategies such as trend following are more likely to help; other approaches are described in tail hedging and gold in the portfolio.

According to the ECB, a higher correlation can initially lead to portfolio risk being underestimated, because historical risk models carry the old correlation forward; losses are then larger or more frequent than expected (Mosk et al. 2022). Brixton et al. (2023) see two ways out, both of which come at a cost: first, taking on more risk or reallocating and giving up expected return in the process; second, adding alternatives that replace the missing diversification.

Rolling correlation: windows and estimation error

Practitioners track the correlation with a rolling window, e.g. 36 monthly returns. But such an estimator is imprecise. The usual approximation for the standard error, (1 − ρ²) / √n, assumes a symmetric interval. In small samples, however, the distribution of the estimator is skewed. The Fisher z-transformation is better: it makes the distribution of the estimator approximately normal and gives an interval that does not extend beyond −1 or +1.

The interval assumes independent returns with constant correlation. In practice, it is likely to be even wider. On top of that come three pitfalls. Overlapping windows: two consecutive 36-month values share 35 months; the rolling curve therefore looks smoother and more persistent than it is. Frequency: daily correlations are shaped by different exchange closing times and short-term flight-to-safety moves and can differ markedly from monthly correlations, which tend to reflect the macroeconomic channels. Weighting: an exponentially weighted estimator reacts faster to a regime change but is noisier; more on this under covariance estimation. If you want to model regimes explicitly, you will find approaches under regime models.

For advanced readers: derivation and biases

General shock model. With r_A = a_g g + a_π π and r_B = b_g g + b_π π and uncorrelated shocks g and π, Cov(r_A, r_B) = a_g b_g σ_g² + a_π b_π σ_π². Because a_g > 0 and b_g < 0, the growth term is negative; because a_π < 0 and b_π < 0, the inflation term is positive. The sign flips at σ_π² / σ_g² = −a_g b_g / (a_π b_π). Adding real interest rate shocks as a third source, which hit both assets via the discount rate, makes the positive part larger. That is why the correlation can also rise without an inflation surge if real interest rates fluctuate strongly.

Volatility bias. Volatility rises in crises, and a correlation measured only in such phases is mathematically biased. Forbes and Rigobon (2002) correct for this with ρ* = ρ / √(1 + δ (1 − ρ²)), where δ measures the relative increase in variance. A crisis correlation of 0.5 with doubled variance (δ = 1) corresponds to only 0.38 after adjustment. For the stock-bond correlation, this means that a short, very negative reading during a panic is not automatically a new regime. Conversely, Longin and Solnik (2001) show for stock markets that correlation really does rise in bear markets (see asymmetric correlation).

Nominal or real. Inflation-linked bonds carry only real interest rate risk. Their correlation with equities therefore reacts less to inflation shocks than that of nominal bonds, but fully to real rate shocks. If you want to manage inflation risk deliberately, you separate the two channels via breakeven inflation. And finally, currency: for euro investors, what counts is the correlation in euros; an unhedged US Treasury contains the dollar, which moves differently in crises from the bond itself.

What the research shows

• Long run: Ilmanen (2003) documents the mostly positive correlation in the 20th century and explains it with inflation shocks; with low inflation, he expects low correlations, and particularly negative ones in deflationary recessions and flight-to-safety phases.
• Inflation and term premium: Campbell et al. (2017) model the change via the covariance of inflation and the real interest rate with the real economy; a high covariance means a high term premium and a concave yield curve.
• Macro dynamics: Campbell et al. (2020) attribute the change in the bond beta around 2001 to the altered co-movement of inflation and the output gap, combined with time-varying risk premia.
• Limits of the macro explanation: Baele et al. (2010) find that liquidity measures explain more than fundamentals and that models do a particularly poor job of capturing covariances.
• Current situation: Mosk et al. (2022) show a non-linear relationship with inflation for the euro area and the US, lowest at around 2%; Brixton et al. (2023) analyse the consequences of a persistently positive correlation for portfolios.

Limitations and common mistakes

• Extrapolating the average of the last 20 years: it comes almost entirely from a regime of falling interest rates and stable inflation.
• Reading the sign from short windows: with 36 months, a correlation of −0.3 is statistically hard to distinguish from zero.
• Mixing daily and monthly values: optimisation with monthly data and monitoring with daily data give different numbers for the same relationship.
• Confusing correlation with hedging: a negative average correlation says little about whether bonds will help in exactly the type of crisis that hits you.
• Forgetting duration: long-term bonds hedge growth crises more strongly, but also lose more in inflation shocks (see duration).
• No forecast: no model reliably predicts the next regime change. This article is not investment advice.

How Portfolio Lab calculates it

On the Assets page, the Correlation tab in Portfolio Lab shows the matrix of period returns in the base currency over the common period and a rolling correlation for a selectable pair. The window defaults to 36 months or 252 trading days and can be set from 6 to 2,520 periods; the correlation is calculated using Pearson's method from simple returns. The page does not show a confidence interval. Use the formula above for that. For long series, the Asset classes “Equities world (developed markets)” (@AKTIEN_WELT), “German Bunds 10y” (@BUND_10J), a synthetic ten-year German government bond built from interest rate series, and “US Treasuries 10y” (@US_TREASURY_10J) are suitable; asset classes always use monthly values.

How a change in correlation affects a portfolio is shown on the Risk page: it breaks down volatility by position, with the sample covariance, exponentially weighted (λ = 0.97 per month) or Ledoit-Wolf, and with windows covering the whole period or the last 36, 60 or 120 months. In the Backtest, you can test 60/40 and risk parity in periods with both signs, such as the 1970s, 2008 and 2022. More on the calculation in the asset analysis methodology.

This article explains a method; it is not investment or tax advice. Worked examples are simplified and rounded. All terms in the glossary.