Key takeaways
- Sorting 72 years of US annual returns by the direction of the 10-year Treasury yield barely moves the stock bond correlation: +0.05 in rising-rate years, +0.12 in falling ones, -0.18 when the yield sat still. All three confidence intervals straddle zero.
- What the rate regime does move is the payment. The 10-year Treasury returned -2.89% a year in rising-rate years, +4.37% in flat ones and +17.34% in falling ones, a spread of just over 20 percentage points.
- The volatility cut held almost constant anyway. A 60/40 mix reduced the annual standard deviation of an all-equity portfolio by 38.1% in rising-rate years, 37.4% in falling and 41.0% in flat.
- Sort the same 72 years by inflation instead and the record separates: +0.29 when December CPI ran at 3% or more, -0.25 below it. From 1998 to 2020 the correlation was -0.54; from 2021 to 2025 it was +0.77.
- Bonds paid in 12 of the 15 years US equities fell, averaging +4.76%. In rising-rate years they still cut volatility, and still lost 8.10% a year of purchasing power doing it.
The stock bond correlation barely moves with the direction of rates
You want to know whether bonds still diversify a portfolio when interest rates are rising. It's a fair question, and 2022 is the reason you're asking.
Here's the answer from the record. We took Damodaran's annual return series for the S&P 500 and for the US T. Bond (10-year), which runs from 1928 to 2025, and matched each year to the change in the Federal Reserve's 10-year constant maturity yield over that same year. That gives 72 usable years, 1954 to 2025, because the Fed's H.15 yield series starts in April 1953.
Split them three ways. A year counts as rising if December's average yield sat at least 0.50 percentage points above the previous December's, falling if it sat at least 0.50 points below, and flat in between. That produces 21 rising years, 18 falling years and 33 flat ones.
The correlation between annual stock returns and annual Treasury returns comes out at +0.05 in rising-rate years, +0.12 in falling ones and -0.18 in flat ones. Those look different. They aren't. The confidence interval on the rising-rate figure runs from -0.39 to +0.47, and on the flat-rate figure from -0.50 to +0.17. They overlap almost completely. Sorted by the direction of rates, the record has no finding in it.
What a rate regime changes is not the correlation, it's the size of the coupon you collect
The interesting split is one notch over. Correlation describes the shape of the relationship. It says nothing about the level of either return, and the level is where rate regimes do their work.
In the 21 rising-rate years the 10-year Treasury returned -2.89% a year on average. In the 33 flat years it returned +4.37%. In the 18 falling years it returned +17.34%. That's a spread of just over 20 percentage points between the best rate regime for bonds and the worst, against a correlation spread of 0.30 that the confidence intervals cannot distinguish from noise.
Equities barely noticed. They averaged 13.17% in rising-rate years, 12.91% in falling ones and 12.33% in flat ones. The rate regime sorted the bond sleeve and left the equity sleeve alone, which is what you'd expect from an asset whose price is mostly a claim on nominal earnings rather than a fixed nominal coupon. If you want the mechanism at instrument level, the arithmetic of bond duration is the same arithmetic, applied to one holding instead of a sleeve.
The volatility cut held between 37% and 41% in every rate regime
Diversification isn't the correlation number. It's what the correlation buys you, and the standard way to measure that is the volatility of the blend against the volatility of the thing you started with.
Run a 60/40 of those same two series inside each regime. In rising-rate years, equities alone had an annual standard deviation of 18.17% and the 60/40 had 11.26%, a cut of 38.1%. In falling-rate years, 19.57% became 12.24%, a cut of 37.4%. In flat years, 14.73% became 8.69%, a cut of 41.0%.
Three regimes, three different macroeconomic worlds, and bond diversification delivered the same 37.4% to 41.0% reduction in every one of them. The worst single year for the 60/40 was -17.96% in a rising-rate regime, -13.89% in a falling one and -7.12% in a flat one, against a worst equity year of -36.55% in 2008.
That is the honest headline finding. The diversification benefit of bonds is close to regime-independent. The price you pay for it is not.
Correlation with equities by rate regime: Treasuries, credit and gold
The same exercise across three more sleeves shows where the ballast actually sits. Each cell is the correlation of that asset's annual return with the S&P 500's, inside the regime named at the top.
| Annual correlation with S&P 500 | Rising rates (21 yrs) | Falling rates (18 yrs) | Flat rates (33 yrs) |
|---|---|---|---|
| 10-year US Treasury | +0.05 | +0.12 | -0.18 |
| Baa corporate bond | +0.42 | +0.67 | +0.30 |
| Gold | -0.24 | -0.13 | -0.07 |
Baa corporate bonds are the row worth staring at. They correlate with equities at +0.42, +0.67 and +0.30 across the three rate regimes, far above the Treasury row in every one. A corporate bond carries the same borrower risk the equity does, so the credit sleeve is partly an equity sleeve wearing a coupon. Gold ran mildly negative in all three, which is a different argument with a different sample problem, and the wider version of it sits in the evidence on diversification in a crisis.
Sort by inflation instead and the record finally separates
If rate direction is the wrong axis, something else has to be the right one. Sort the same 72 years by December-to-December CPI-U from the Bureau of Labor Statistics and the sample splits cleanly. In the 33 years inflation ran at 3% or more, the stock bond correlation was +0.29. In the 39 years below 3%, it was -0.25.
Chop it by era and the gap is wider still. From 1998 to 2020 the correlation across 23 annual observations was -0.54, with a confidence interval of -0.78 to -0.17 that excludes zero. From 2021 to 2025 it was +0.77. Across 1954 to 1997 it was +0.19.
That's not a novel result, and it shouldn't be. The Bank for International Settlements documented the same break with daily data in its December 2023 Quarterly Review, noting that "the correlation between US equity and government bond returns switched sign in mid-2021" and that "one has to go back to the 1980s and the early 1990s to find a prolonged period with positive correlations". Its explanation is about which shock dominates. When inflation is low, markets price growth news, and growth news pushes equities and bonds in opposite directions. When inflation is high, markets price the inflation outlook, and an inflation surprise pushes both the same way.
Vanguard put a number on the threshold before the event. Its September 2021 paper on the stock bond correlation concluded that "10-year trailing inflation would have to be around 3% on average over the next five years to have a significant impact on correlation regimes", and that reaching it would need annual core inflation "maintained at the minimum rate of 5.7% over the same period". Ten-year trailing CPI-U inflation stood at 2.14% a year in December 2021. By December 2025 it was 3.20%. The condition the paper named as unlikely is the condition that arrived.
1973 to 1981: bonds did their job and still lost 39% in real terms
Nominal returns flatter bonds in exactly the regime where they hurt most. The stretch from 1973 to 1981 is the cleanest case on the record.
In 1974 equities fell 25.90% and the 10-year Treasury returned +1.99%. The hedge worked. Inflation that year was 12.34%, so the bondholder's +1.99% was a real loss of around a tenth of their money. Compound the whole stretch and the 10-year Treasury lost 39.0% of its purchasing power between 1973 and 1981, while doing precisely what the textbook said it would do in the equity drawdowns.
Across all 21 rising-rate years, the average real return on the 10-year Treasury was -8.10% a year, against +7.36% for equities. In flat years the bond's real return was +1.69%; in falling years, +14.27%. So the regime-independence of the volatility cut comes with a caveat the volatility number cannot show you. You keep the shock absorber. You pay for it in purchasing power, and in a rising-rate regime the bill is roughly 8 points a year.
Bonds paid in 12 of the 15 equity-down years, and 2022 was one of the three that didn't
The test most people actually care about is narrower than a correlation. Did the bond go up when the stock market went down?
Across 1954 to 2025 the S&P 500 fell in 15 calendar years. The 10-year Treasury finished positive in 12 of them, averaging +4.76% across all 15. The three failures were 1969, when equities lost 8.24% and the Treasury lost 5.01%; 2018, when the Treasury finished at -0.02% against an equity loss of 4.23%; and 2022, when equities fell 18.04% and the Treasury fell 17.83%.
Two of those three sat inside a rising-rate regime, and 2022 is the year that made the question urgent. It's also a single observation. One year in which a hedge fails is evidence about that year, and a correlation estimated over five years is mostly evidence about how few years you have. The broader problem with reading a correlation off a short window is the subject of correlation instability, and it applies to this piece as much as to any other.
The UK version: gilts fell 21.9% over five years while UK equities rose 69.8%
The series above is American, and a UK reader holds gilts. The FTSE Actuaries UK Conventional Gilts All Stocks Index returned -21.9% in total over the five years to 31 August 2026, or -4.8% a year, with a maximum drawdown over that window of 32.8%. The FTSE All-Share returned +69.8% over the same five years, or +11.2% a year.
That is the rising-rate regime in one pair of figures, and it's harsher than the US annual averages suggest. It's also not the same test. A cumulative five-year window is not a correlation, and the gilt factsheet publishes no calendar-year returns, so the two indices can't be lined up year by year here. What the pair does show is the size of the gap. UK equities returned +0.3% in calendar 2022, the year US equities and Treasuries fell together, and the gilt index spent that five-year window setting a 32.8% drawdown.
The strongest objection: 18 observations cannot prove a negative
The case against everything above is straightforward, and it's a good one. Annual data gives 18 to 33 observations per rate regime. At that size the standard error on a correlation is roughly 0.2, which means this study could fail to detect a real effect of meaningful size and report it as noise. Absence of evidence here is genuinely weak evidence of absence.
The regime definition is also a choice, and the choice matters. Move the threshold from 0.50 percentage points to 1.00 and the rising-rate bucket shrinks to 11 years and its correlation jumps from +0.05 to +0.37. That is a large change from a small edit, and it's the honest way to read the robustness: the rising-rate correlation is not reliably zero, it is unreliably estimated.
The 2021 to 2025 figure of +0.77 deserves the same scepticism, and more of it. Five annual observations produce a confidence interval so wide it reaches from -0.34 to +0.98. That number carries weight only because the BIS found the same sign break in daily data, and because the inflation threshold Vanguard named in advance was crossed on the way. On its own it would be an anecdote with a decimal point.
Finally, the data cannot speak to timing inside a year. Annual returns hide the path. A bond that ended 2022 down 17.83% spent parts of that year as a worse hedge than the annual number shows, and parts as a better one.
What would change this conclusion
Three things would move it. The first is a sustained fall in trailing inflation back under the level Vanguard named: if 10-year trailing CPI-U drifts back below 3.20% toward its 2021 level, the post-2021 positive correlation loses its stated cause, and the 1998 to 2020 pattern becomes the live one again. The second is a long stretch of rising yields with low inflation. That combination is thin in this sample, and it's the one case where the rate regime and the inflation regime give different answers, so it would settle which axis is doing the work.
The third is arithmetic rather than macroeconomics. A diversifier that cuts volatility by 38.1% while losing 8.10% a year in real terms is a different proposition from one that cuts volatility by 38.1% while earning its coupon. Starting yields set that, not correlations. The 10-year Treasury yield averaged 4.14% in December 2025, which is a materially different starting point from the yields that ran through the negative-correlation era of 1998 to 2020. If you track a blended portfolio, the drift between those two sleeves is visible in LedgerTouch long before the correlation of their annual returns is.