Stock–bond diversification across changing correlation environments

Research significance

Multi-asset allocators and enterprise risk teams

The finding matters wherever historical correlations determine risk budgets, hedges, capital allocation, retirement glide paths, or the expected protection of a balanced portfolio.

Decision context

Whether historical diversification deserves to be treated as insurance

Do not size protection from one estimated covariance matrix alone. Ask what happens if the macro driver changes and historically distinct sleeves begin sharing the same loss factor.

The first view is historical. A rolling 126-trading-day correlation between SPY and TLT tracked the negative stock–bond relationship of 2015–2019. The same series records its reversal during the 2022–2023 interval. Those values describe this sample. They do not identify the next environment.

The second view is counterfactual. The stress held individual asset volatilities fixed. It then blended the estimated correlation matrix toward perfect positive correlation, repairing the matrix to stay positive semidefinite. At the observed structure the strategic portfolio ran 11.6% annualized volatility. At the perfect-correlation limit it reached 15.8%, the weighted average volatility of its components. About four percentage points of diversification disappeared.

The two views complement each other. History shows that a load-bearing relationship moved. The stress shows the portfolio consequence. Neither claims when or why the endpoint will occur.

• Show portfolio risk across a range of correlation states instead of one point estimate.
• Identify which pairwise relationship carries most of the diversification benefit.
• Pair covariance-based risk with named macro scenarios that explain why correlations might change.
• Monitor rolling relationships without treating the latest window as a forecast of the next regime.
• Describe “defensive” assets conditionally: protection against one shock may share exposure to another.

Inspect correlation stress → Inspect the research record → Reproduce from the executed notebook ↗

Lab measurement SPY–TLT correlation averaged −0.35 during 2015–2019. A later rolling window reached +0.37 during 2022–2023. Separately, an all-asset correlation stress raised annualized portfolio volatility from 11.6% to 15.8%.

External empirical evidence Federal Reserve and BIS evidence documents elevated inflation, expected tightening, rising Treasury yields, and falling equity prices. Inflation news became more important to stock–bond co-movement after mid-2021.

Interpretive synthesis Nominal duration is protection against particular shocks, not universal portfolio insurance. Inflation can constrain monetary easing and expose both sleeves to a common discount-rate channel.

Causal boundary: the lab measures a sign change and a portfolio stress. It does not attribute the change to one macroeconomic cause or forecast how long the positive relationship will persist.

Examine the diversification regime →

Portfolio constructionPortfolio optimization under estimation error and out-of-sample evaluationEvery construction method embeds assumptions about the environment it will face. Model calibrationHistorical VaR under coverage and independence testsA slowly adapting historical forecast can fail when the return process changes.

Bonds do not protect a portfolio in every kind of downturn.

Interpretation

Bonds often help when growth weakens and inflation is contained. They may fall with stocks when inflation keeps interest rates high. The same holds when investors demand more compensation to hold long-term debt.

Decision context

Whether a balanced allocation deserves to be treated as self-hedging, and how much confidence to place in one historical relationship among assets.

Intended analytical use

Multi-asset allocation, risk governance, retirement research, and model validation.

Principal limitation

The lab observes one liquid-ETF history. The synthesis connects that record with external evidence. It does not identify a single cause or forecast the next regime.

Lab measurement The lab records a clear change in one studied portfolio relationship. A rolling 126-trading-day correlation between SPY and TLT averaged negative during 2015–2019. It turned positive during 2022–2023 and stayed positive to the end of the frozen sample. Separately, an all-asset static stress moved the full correlation matrix toward one. That stress materially raised the studied portfolio's volatility.

Lab measurement

Values produced by the lab's frozen data, declared transformations, tests, and stress design.

External empirical evidence

Measured relationships reported by identified research outside the lab.

Institutional record

Dated policy actions and conditions documented by central banks and international institutions.

Interpretive synthesis

A bounded connection between the measured result, the external record, and an investor decision.

Synthesis ledger

Classification Empirical synthesis of existing investigations; not a seventh executed study.

Measured evidence Frozen SPY–TLT rolling correlation, correlation stress, portfolio evaluation, and historical VaR backtest.

External empirical evidence BIS, peer-reviewed, and Federal Reserve staff research remain distinct from lab estimates.

Institutional record Dated ECB, Federal Reserve, and IMF records establish chronology and stated actions without proving causality.

Data vintage Committed adjusted-close snapshot through 6 July 2026; evidence commit ad24c4999583.

Forecast status None. No present-state classification, regime probability, or return forecast.

Identification limit The observed relationships do not identify inflation, policy, fiscal, liquidity, or geopolitical causes.

Causal boundary Correlation establishes co-movement, not cause. The lab did not identify monetary, fiscal, inflation, growth, liquidity, or geopolitical shocks inside the SPY–TLT series.

External empirical evidence The negative stock–bond relationship familiar during much of the 2000s was historically unusual rather than timeless. BIS research reports that the correlation between United States equities and government bonds switched sign in mid-2021. Prolonged positive correlations had last appeared in the 1980s and early 1990s.¹ ECB evidence documents generally positive correlations from the late 1960s through the late 1990s. Correlations were predominantly negative during the 2000s.²

External empirical evidence One documented mechanism begins with the news that dominates expected policy. When inflation is low and stable, disappointing growth can reduce expected earnings while increasing the likelihood of monetary easing. Stocks fall, yields decline, and bond prices can rise. When high or volatile inflation constrains that response, inflation news can lower nominal bond prices, lift expected rates, and weigh on equities at the same time.

External empirical evidence The June 2022 Federal Reserve Monetary Policy Report records rising Treasury yields amid sustained inflation pressure and expected tightening. Broad equity prices declined sharply. The report also records rapid increases in the federal-funds target range and the start of balance-sheet reduction.³ A later Federal Reserve Board staff working paper studies announcement windows. In those windows, higher-than-expected inflation came with lower stock prices, higher nominal risk-free yields, and higher equity risk premia.⁴

Scenario map. These are mechanisms to test, not forecasts or guaranteed asset responses.

Institutional record The Federal Reserve's historical account of the Great Inflation describes interacting monetary, fiscal, energy-price, and institutional forces rather than one isolated cause.⁶ The IMF separately documents how greater sovereign-bond supply and central-bank balance-sheet runoff can place more duration with price-sensitive investors. That shift can pressure term premia. The evidence supports the fiscal and market-capacity scenario above. It does not identify the cause of the observed stock–bond reversal.⁷

Interpretive synthesis A bond hedge is partly an institutional outcome. Inflation credibility, the expected central-bank reaction, fiscal financing, debt supply, and market capacity can influence whether sovereign duration behaves as insurance. The rolling correlation alone identifies none of those channels. Together they form a disciplined scenario set for testing the allocation.

Place this mechanism inside the broader regime framework →

Lab measurement Two separate tests establish two narrow facts. The pairwise SPY–TLT relationship changed sign. In a distinct all-asset stress, moving the full correlation matrix toward one raised annualized portfolio volatility from 11.6% to 15.8%. Neither result establishes when another change will occur.

External empirical evidence BIS evidence links the recent positive correlation to an environment where inflation news mattered more for expected monetary policy. Growth-news effects weakened over the same span. Structural research also connects long-run changes in stock–bond co-movement with changes in the inflation–output relationship and time-varying risk premia.⁵

Interpretive synthesis “Bonds are defensive” is incomplete. Nominal government bonds may hedge disinflationary growth shocks. They can share losses with equities when inflation, expected policy rates, or term premia dominate. A portfolio review should name the shock each defensive sleeve is expected to absorb.

Decision rule Show risk across several correlation states. Pair the covariance matrix with named macroeconomic scenarios, duration exposure, and liquidity assumptions. State which institution is expected to absorb the shock.

Inspect the complete diversification finding →

Lab measurement Maximum-Sharpe optimization fell from about 1.12 in the 2015–2021 fit period to 0.44 in the sealed 2022–2026 evaluation. Risk parity and hierarchical risk parity delivered a more stable realized risk shape but did not establish a reliable return advantage.

External empirical evidence The fit and evaluation windows straddle a documented change in stock–bond correlation and monetary conditions. That chronology matters because every estimated mean and covariance summarizes the sample that produced it.

Interpretive synthesis Precise weights assume that fitted returns and covariances remain useful for the next decision. Here, relationships measured in one observed environment were applied in another. The optimizer itself did not estimate or forecast a policy regime.

Causal boundary The experiment did not identify a regime change as the cause of the performance ranking. It cannot attribute the Sharpe decline to stock–bond correlation alone. Estimation error, concentration, asset selection, and the realized equity path also moved the ranking.

Inspect the portfolio-estimation finding · Analyze allocation sensitivity · Inspect the research record

Lab measurement The rolling historical 95% VaR model produced 135 breaches against 107 expected across 2,140 forecasts. Coverage, independence, and conditional-coverage tests all rejected calibration. The exceptions were both too frequent and clustered. The 500-observation reference window changes only as new returns enter it.

External empirical evidence Dated Federal Reserve records and cross-asset research document a sample period in which inflation, policy rates, Treasury yields, equity valuations, and stock–bond correlation changed together.

Interpretive synthesis A breach cluster should trigger a review of the assumed return process, covariance structure, volatility state, and missing scenarios. Increasing the latest risk number without diagnosing the forecast process leaves the failure mechanism unresolved.

Causal boundary The backtests identify frequency error and dependence among exceptions. They do not reveal which return process changed. They also do not identify inflation, policy, correlation reversal, or any other macro factor as the cause. Other windows, weighting schemes, and conditional models require their own validation.

Inspect the VaR finding · Inspect the VaR backtest · Inspect the research record

The dossier does not estimate a structural macroeconomic model, identify shocks, or decompose returns into inflation, growth, policy, cash-flow, liquidity, and term-premium components. The external literature supplies credible mechanisms, not a unique explanation for the lab's observations.

Monitoring map. Every proxy is incomplete and should be interpreted with competing explanations.

• Correlation depends on return frequency, bond duration, total-return construction, and window length.
• Inflation levels, surprises, uncertainty, expectations, and inflation-risk compensation are different variables.
• Breakeven inflation also contains risk and liquidity premia; it is not a pure expectation.
• Historical analogies should use data available at the time rather than revised series alone.
• Portfolio risk depends on weights and volatilities as well as correlation.
• Political and institutional narratives require explicit alternatives and cannot substitute for identification.

A new empirical investigation would need a declared regime classifier and longer duration-consistent total-return histories. It would also need several defensible windows and real-time macroeconomic vintages. Overlapping estimates would need an uncertainty treatment, and the falsification criterion would be fixed before comparison. Until then, the defensible output is a scenario framework, not a probability assigned to the next regime.

1. Lombardi, M. J., and V. Sushko (2023), “The Correlation of Equity and Bond Returns,” BIS Quarterly Review, December. BIS research box
2. Mosk, B., L. Pangallo, and S. M. Zema (2022), “Cross-asset correlations in a more inflationary environment and challenges for diversification strategies,” Financial Stability Review, November. ECB analysis
3. Board of Governors of the Federal Reserve System (2022), Monetary Policy Report, June, part 1. Federal Reserve report
4. Knox, B., and Y. Timmer (2025), “Stagflationary Stock Returns,” Finance and Economics Discussion Series 2025-056. Federal Reserve research page
5. Campbell, J. Y., C. Pflueger, and L. M. Viceira (2020), “Macroeconomic Drivers of Bond and Equity Risks,” Journal of Political Economy 128(8), 3148–3185. doi:10.1086/707766
6. Federal Reserve History (2013), “The Great Inflation.” Historical overview
7. International Monetary Fund (2025), Global Financial Stability Report, April, chapter 1 on sovereign-bond supply and market functioning. IMF report

Lab evidence and reproduction

• Empirical finding: stock–bond diversification across changing correlation environments
• Empirical finding: portfolio optimization under estimation error
• Empirical finding: historical VaR under coverage and independence tests
• Research record: portfolio construction
• Research record: market risk and correlation stress
• Executed portfolio notebook at the evidence commit
• Executed market-risk notebook at the evidence commit

Historical frameworkMarket regimes as transmission systemsConstruct a hypothetical scenario and compare dated historical lenses. DiversificationStock–bond diversification across changing correlation environmentsInspect the measured relationship, stress surface, application, and limitations. AllocationPortfolio optimization under estimation error and out-of-sample evaluationInspect what survived when optimized allocations entered the sealed evaluation window. CalibrationHistorical VaR under coverage and independence testsSeparate forecast frequency error from dependence among exceptions. MethodsEvidence classification and bounded inferenceReview falsification, uncertainty, provenance, reproduction, and bounded interpretation.