> ## Content Index
> Fetch the complete content index at: https://aligrithm.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# 4.75 Stock-Bond Correlation: The Sign Flips and So Do Its Drivers
- URL: https://aligrithm.com/stock-bond-correlation-the-sign-flips-and-so-do-its-drivers/
- Published: 2026-09-01T12:39:28.000Z
- Updated: 2026-09-01T12:39:27.000Z
- Description: Stock-bond correlation flipped from +0.22 to -0.24 around 2000 across G7 markets. Real rates most stable driver. US portfolios need 80% stocks post-2000 to match 50/50 pre-2000 performance.
- Author: ali askar
- Tags: 4. Market Structure Notes

A 50/50 stock-bond portfolio that earned its target return with moderate volatility before 2000 requires an 80/20 allocation after 2000 to maintain the same profile. That 30-point shift is not a style choice or a bet on equities. It is the portfolio adjustment needed to compensate for the stock-bond correlation flipping from positive to negative, which McMillan documents across the G7 using DCC-GARCH and rolling windows from 1980 to 2023\. The flip happened around 2000 for most markets, a decade earlier for Japan, and barely at all for Italy. The correlation recently shows signs of reverting toward positive around 2022, following the inflation surge. Before you assume this is another dotcom-crash flight-to-safety story, look at the drivers. They are not constant. Real interest rates are the most stable predictor post-2000, but inflation and growth flip sign and significance across Bai-Perron structural breaks. The drivers themselves are time-varying.

## The correlation flips, measured two ways

Stock-bond correlation is usually estimated with rolling windows or a dynamic model. McMillan uses both: a three-year rolling Pearson correlation and Engle's DCC-GARCH, which estimates time-varying correlations from a multivariate GARCH model that lets the correlation evolve with its own equation. The DCC model is:

$$ \\rho\_{sb,t} = \\text{corr}(r\_{stocks,t}, r\_{bonds,t} \\mid \\text{info up to } t-1) $$

At each time t, the model computes the conditional correlation between stock and bond returns given past volatilities and correlations. The rolling window is simpler: take the last 36 months of returns and compute the Pearson correlation. Both methods agree on the broad pattern.

Worked reading: if the DCC correlation for the US in January 2000 is +0.22 and by December 2000 it is −0.05, the correlation switched from positive to negative that year. A positive correlation means stocks and bonds tend to move together, so bonds diversify but do not hedge. A negative correlation means they move opposite, so bonds hedge stocks. The switch from +0.22 to −0.05 is the switch from diversifier to hedger.

The sample is monthly total returns on stock and bond indices for Canada, France, Germany, Italy, Japan, the UK, and the US, spanning 1980 to 2023 where data permits. Table 2 in the paper splits the sample at the midpoint for each country. Pre-2000 correlations are positive for every market: Canada +0.16, France +0.22, Germany +0.18, Japan near zero at +0.02, UK +0.32, US +0.22\. Italy is the outlier at +0.30\. Post-2000, the correlation flips negative for six of the seven: Canada −0.12, France −0.03, Germany −0.20, Japan −0.24, UK −0.10, US −0.24\. Italy stays positive at +0.27\. The sign switch is not a US-only phenomenon.

Looking at the DCC time series directly, the switch occurs around 2000 for Canada, Germany, the UK, and the US. France switches around the same time but shows earlier evidence of mean reversion back toward positive in the mid-2010s. Japan switched in mid-1992 and remained negative through 2023, with only weak signs of reversion at the end. Italy had a brief negative period from 2000 to 2004 but otherwise maintained a positive correlation throughout. The rolling correlations exhibit greater volatility around the same pattern, which is expected given the smaller effective sample in each window.

By the end of 2022 and into 2023, the DCC correlations for most markets shift back toward positive or less negative, consistent with the view that the 2022 inflation surge and rate hikes moved both stocks and bonds in the same direction again.

## Drivers: real rates stable, the rest unstable

The paper regresses the time-varying correlation on a set of macro variables:

$$ \\rho\_{sb,t} = \\alpha\_0 + \\beta\_1 \\, \\text{Inflation}\_{t-1} + \\beta\_2 \\, \\text{Infl. Vol.}\_{t-1} + \\beta\_3 \\, \\text{Output}\_{t-1} + \\beta\_4 \\, \\text{Output Vol.}\_{t-1} + \\beta\_5 \\, \\text{Real IR}\_{t-1} + \\beta\_6 \\, \\text{Corr(Infl., Output)}\_{t-1} + \\varepsilon\_t $$

Inflation is the month-on-month log change in CPI, output is the log change in industrial production, inflation volatility and output volatility are estimated via GARCH or rolling standard deviation to match the correlation series, the real interest rate is the 10-year Treasury yield minus inflation, and the inflation-output correlation is itself time-varying, estimated by DCC-GARCH or a rolling window. All predictors are lagged one period.

Worked example for interpreting a coefficient: if the real interest rate coefficient is +0.055 (the US DCC regression result), a 1 percentage point increase in the real rate is associated with a +0.055 increase in the stock-bond correlation. In practice, real rates varied from near zero to 5% over the sample, so a 2-point move from 1% to 3% would lift the correlation by 0.11, which is enough to move it from neutral to moderately positive.

Table 5 reports the full-sample regression for both DCC and rolling correlations. Real interest rates are positive and significant for all markets except Canada and Germany in the DCC specification, and significant for all seven in the rolling specification. Inflation is positive and significant for Germany, Japan, the UK, and the US (DCC), and for Canada, Germany, Japan, the UK, and the US (rolling). Inflation volatility is positive and significant for France, Germany, Italy, Japan, the UK, and the US (DCC), fewer for rolling. Output growth is positive and significant in a handful of cases, including for Canada, France, Italy, and the US in the rolling specification. The typical theoretical story is that higher real rates and higher inflation both depress stock and bond prices via higher discount rates, producing a positive correlation, while higher growth should raise stock prices and lower bond prices, producing a negative correlation. The results here support the first part but contradict the second: output largely has a positive effect, suggesting that in practice higher growth raises demand for all assets rather than rotating from bonds to stocks.

The inflation-output correlation (the "nominal-real" correlation) is mostly insignificant, which contradicts the recent literature that links the stock-bond correlation flip to a flip in the inflation-output correlation from negative to positive.

## Bai-Perron breaks: the drivers change over time

The above regressions assume constant coefficients. The Bai-Perron procedure tests for multiple structural breaks and estimates separate coefficients for each regime. Table 6 reports the break dates. Most markets show breaks around 1998-1999 (dotcom period), the mid-2000s (pre-financial crisis), and the mid-2010s (post-debt crises). Japan and the US have an additional early-1990s break. Germany and Italy have breaks associated with the financial crisis itself.

Table 7 presents the regime-specific coefficients, grouped roughly by date. The variables that affect the correlation change both in sign and in significance across regimes. For the US, inflation and output volatility are significant in the 1980s, only output volatility remains significant through the 1990s, only inflation volatility is significant in the early 2000s, and neither is significant post-financial crisis. Real interest rates, in contrast, are positive and significant from the late 1990s onward for most markets, with some country variation. Germany shows a negative significant real-rate coefficient in the first half of the 2010s while most others are positive.

Inflation and inflation volatility become increasingly positive and significant from the late 1990s forward. Output growth remains mixed, sometimes positive and sometimes negative, with no stable pattern. The nominal-real correlation is also mixed.

The key takeaway is that the factors driving the stock-bond correlation are not stable. Real interest rates are the most consistent predictor post-2000, typically positive, meaning higher real rates push the correlation toward positive. Inflation also tends to push positive. Growth is unstable. The notion that the correlation is driven by a single unchanging mechanism is not supported.

## Portfolio implications: you need more stocks to stand still

The correlation flip is not an academic curiosity. It directly affects portfolio returns and risk. Figure 3 in the paper plots the cumulative return of a 50/50 stock-bond portfolio for each G7 market. For Canada, France, Germany, the UK, and the US, the cumulative path shows a visible flattening or kink around 2000, consistent with the switch to negative correlation detracting from performance. Japan shows a steeper path pre-1995 when the correlation was still positive. Italy, which maintained a mostly positive correlation, does not show the same kink.

The US case is the cleanest illustration. Figure 4 (reproduced below) plots the cumulative 50/50 portfolio for the US alongside a trend fitted to the pre-2000 period and extrapolated forward. The realized portfolio diverges sharply below the trend after 2000\. The lower panel overlays portfolios with different stock weights. From 2000 to 2008, a 65/35 stock-bond allocation is needed to track the pre-2000 trend. From 2009 onward, an 80/20 allocation is required.

![US stock-bond portfolio performance with trend and alternative weights](https://storage.ghost.io/c/27/cb/27cb0fc8-2c77-4434-af9e-d5d32a916994/content/images/2026/08/article_471-us_portfolio_weights.png)

Worked reading of the figure: the orange line is the pre-2000 trend. The blue 50/50 portfolio diverges below it after 2000\. The green 65/35 portfolio tracks the trend from 2000 to 2008\. The red 80/20 portfolio tracks the trend from 2009 to 2023\. This is not a bet on higher stock returns. It is the allocation needed to offset the drag from the negative correlation.

Table 8 quantifies the same effect from a risk perspective. The mean and standard deviation of the 50/50 portfolio over the full sample are shown, then split into pre-2000 and post-2000 periods. The table also reports the stock weight W needed in each sub-period to match the full-sample standard deviation. For the US, the full-sample 50/50 portfolio has a mean return of 0.40% per month and a standard deviation of 2.57%. Pre-2000, the same 50/50 portfolio had a mean of 0.52% and a standard deviation of 2.72%, so to bring that down to the full-sample 2.57%, you would need only 41% in stocks (W1 = 0.41). Post-2000, the 50/50 portfolio had a mean of 0.29% and a standard deviation of 2.42%, so to lift that up to 2.57%, you need 54% in stocks (W2 = 0.54). The risk shift is symmetric: lower weight pre-2000, higher weight post-2000, to maintain constant volatility.

The same pattern holds for returns. To maintain a constant return, you need to increase the stock allocation post-2000 because the negative correlation no longer provides the same diversification boost.

## Variance decomposition: the correlation contribution switches sign

Portfolio variance for a two-asset portfolio is the sum of the variance contribution from each asset plus twice the covariance. For a 50/50 stock-bond portfolio:

$$ \\sigma\_p^2 = 0.25 \\, \\sigma\_{stocks}^2 + 0.25 \\, \\sigma\_{bonds}^2 + 2 \\times 0.25 \\times \\sigma\_{stocks} \\times \\sigma\_{bonds} \\times \\rho\_{sb} $$

Dividing each term by the total portfolio variance gives the fractional contribution of each component. Table 9 reports these contributions for the full sample and the two sub-periods, using a five-year rolling window to compute variances.

For the US over the full sample, stocks contribute 88% of portfolio variance, bonds 24%, and the covariance term contributes −12%. The negative contribution from covariance is the hedging benefit: the negative correlation reduces total variance. Split by period, the story is clearer. Pre-2000, stocks contribute 55%, bonds 22%, and covariance +23%. The positive covariance contribution means the correlation is positive and adds to variance. Post-2000, stocks contribute 108%, bonds 25%, and covariance −34%. The covariance term is now large and negative, providing a substantial hedge. The stock contribution exceeds 100% because the covariance term is negative, so the total sums to 100%.

The same flip from positive to negative covariance contribution occurs for Canada, France, Germany, Japan, and the UK. Italy, which maintained a positive correlation, shows a positive covariance contribution in both periods. The bond contribution remains broadly stable across periods for all markets, while the stock contribution increases post-2000 as equity volatility rose.

The variance decomposition confirms that the correlation flip is not cosmetic. It fundamentally changes the risk profile of a stock-bond portfolio, shifting the covariance term from a drag to a hedge.

## What caused the flip, and is it stable

The paper does not settle causation. The literature points to several candidates: the dotcom crash and subsequent flight-to-safety, the prolonged low-rate environment post-financial crisis, changing inflation dynamics, and shifts in the correlation between inflation and growth. The regression results show that real rates and inflation are associated with higher correlation, while growth is mixed. The Bai-Perron breaks coincide with major crisis periods, which supports the view that regime shifts in macro risk drove the correlation changes.

The reversion toward positive correlation in 2022-2023 coincides with the inflation surge and rate hikes, where both stocks and bonds sold off together. If inflation remains elevated or volatile, the correlation may stay positive or neutral. If inflation recedes and rates stabilize, the correlation could flip back to negative. The drivers themselves are unstable, so the correlation is not anchored.

For portfolio managers, the implication is that the 60/40 or 50/50 portfolio is not a static allocation. The correlation between stocks and bonds is a regime variable, and the regime can persist for decades. Rebalancing rules that assume a constant correlation will mis-allocate. Monitoring the correlation directly, or the macro drivers that predict it, is necessary to maintain a target return or risk profile.

The paper is limited to G7 developed markets and monthly data. Higher-frequency data or emerging markets may show different patterns. The regression framework is reduced-form and does not distinguish between structural shocks (supply vs demand, monetary vs real). The variance decomposition assumes static weights and does not account for dynamic rebalancing strategies that might exploit the correlation flip. The sample ends in December 2023, before the full effects of the 2022-2023 rate cycle are observable.

The finding that the stock-bond correlation flips sign, that the drivers of the flip are themselves time-varying, and that the flip has direct portfolio consequences, is robust across methods and markets. The 80/20 allocation needed post-2000 to replicate pre-2000 50/50 performance is not a recommendation. It is a measurement of how much the correlation regime changed.

![](https://storage.ghost.io/c/27/cb/27cb0fc8-2c77-4434-af9e-d5d32a916994/content/images/2026/08/article_471-visual-0.png)

## KEY POINTS

- Stock-bond correlation flipped from positive to negative around 2000 for six of seven G7 markets (earlier for Japan, barely for Italy), measured via DCC-GARCH and 3-year rolling windows. Pre-2000 correlations ranged from +0.02 (Japan) to +0.32 (UK); post-2000 they turned negative for most, reaching −0.24 for the US and Japan.
- Correlation shows partial reversion toward positive around 2022-2023, coinciding with the inflation surge and rate hikes that moved stocks and bonds together again.
- Real interest rates are the most stable driver of the correlation post-2000, with a positive coefficient: higher real rates push the correlation toward positive. Inflation also typically positive. Output growth mixed and unstable.
- Bai-Perron structural breaks show the drivers themselves are time-varying. Breaks cluster around 1998-1999 (dotcom), mid-2000s (pre-crisis), and mid-2010s (post-debt crises). The factors that predict the correlation change in sign and significance across regimes.
- Portfolio impact is large. US investors needed to shift from 50% stocks to 80% stocks after 2000 to maintain the same return and risk profile as a pre-2000 50/50 allocation. The 30-point shift compensates for the negative correlation no longer providing diversification.
- Variance decomposition: covariance contribution to portfolio variance flipped from +23% (pre-2000) to −34% (post-2000) for the US, turning bonds from a drag into a hedge. Stock contribution increased, bond contribution stable.

## References

- [Why does the correlation between stock and bond returns vary over time? - Andersson et al. (2008)](https://scholar.google.com/scholar?q=Why+does+the+correlation+between+stock+and+bond+returns+vary+over+time+Andersson&ref=aligrithm.com)
- [Smooth transition patterns in the realized stock-bond correlation - Aslanidis & Christiansen (2012)](https://scholar.google.com/scholar?q=Smooth+transition+patterns+in+the+realized+stock-bond+correlation&ref=aligrithm.com)
- [The determinants of stock and bond return comovements - Baele et al. (2010)](https://doi.org/10.1093/rfs/hhq014?ref=aligrithm.com)
- [Estimating and testing linear models with multiple structural changes - Bai & Perron (1998)](https://doi.org/10.2307/2998540?ref=aligrithm.com)
- [Macroeconomic drivers of bond and equity risks - Campbell et al. (2020)](https://doi.org/10.1086/707766?ref=aligrithm.com)
- [Dynamic conditional correlation - Engle (2002)](https://doi.org/10.1198/073500102288618487?ref=aligrithm.com)
- [Macroeconomic factors and the correlation of stock and bond returns - Li (2002)](https://scholar.google.com/scholar?q=Macroeconomic+factors+and+the+correlation+of+stock+and+bond+returns+Li+2002&ref=aligrithm.com)
- [Stock-Bond Return Correlation: Understanding the Changing Behaviour - McMillan (2025)](https://papers.ssrn.com/sol3/papers.cfm?abstract%5Fid=5193234&ref=aligrithm.com)