Diversification in a Crisis: What Actually Worked

12 min read

Key takeaways

  • Across the six times developed-world equity fell 20% or more since July 1990, a 60/40 split between world equity and 10-year Treasuries lost between 9.4% and 26.5% from peak to trough, against 21.4% to 57.0% for equity on its own. In the November 2021 to October 2022 decline the bond sleeve absorbed only 3.8 of the 26.7 points equity lost.
  • Treasuries lost money in two of the six windows, and both were oil shocks. They returned -0.8% through the 1990 Gulf War fall and -17.1% through 2022. Only 2022 was also a protection failure: in 1990 the 40% Treasury sleeve still absorbed 8.55 of the 22.16 points equity lost.
  • Gold ended the 2007-09 window up 17.0% but fell 29.5% from its own high inside that window. In the 2020 crash its daily correlation with US equity was +0.454, against -0.004 across the whole 1990-2026 sample.
  • US REITs lost 68.9% in the 2007-09 window against 54.2% for US equity, and their daily correlation with US equity ran 0.945 through the 2020 crash against 0.737 over 2005-2026.
  • The convergence story is overstated. On weeks ending Wednesday, correlation between US and developed ex-US equity was 0.723 over the full sample and 0.785, 0.793 and 0.829 inside the three longest crisis windows -- higher, but a long way from 1. Move the week's closing day and the rise ranges from -0.02 to +0.15.

Treasuries and Treasury bills carried the protection; most of the rest didn't

You hold several asset classes. In the crash that actually mattered to you, did they help?

Here's the record across every 20% fall in developed-world equity since 1990. The answer is narrow. Government bonds and Treasury bills carried nearly all of the protection, and one of those two lost money twice. Gold's reputation depends on where the end date goes. Listed property behaved like equity with a different label. International equity co-moved more in a crisis than in calm markets, though by less than the folklore claims.

The windows come from a running maximum, not from memory

Picking crisis dates by recollection is how you end up with a result that agrees with you. These windows come from a rule stated in advance.

The reference series is the Fama/French Developed market portfolio: every listed stock in the developed markets of that universe, value-weighted, dividends included, in US dollars, daily from 2 July 1990. A running maximum is carried forward from the first day of the sample. A window opens on the running maximum preceding any fall of 20% or more from it, and closes on the lowest close before the index regains that maximum. Nothing else qualifies a window, and no judgement about what counted as a crisis enters anywhere. That produced six.

Peak-to-trough total return in US dollars, held from the peak with no rebalancing. "60/40 bonds" is 60% developed-world equity and 40% 10-year US Treasuries; "60/40 bills" swaps the Treasury sleeve for 3-month Treasury bills.
WindowTrading daysWorld equity60/40 bonds60/40 bills
17 Jul 1990 - 28 Sep 199052-22.16%-13.61%-12.69%
20 Jul 1998 - 8 Oct 199857-21.39%-9.39%-12.41%
24 Mar 2000 - 9 Oct 2002615-47.97%-13.00%-24.84%
31 Oct 2007 - 9 Mar 2009328-57.03%-26.48%-33.42%
12 Feb 2020 - 23 Mar 202028-33.79%-16.89%-20.23%
8 Nov 2021 - 12 Oct 2022221-26.69%-22.87%-15.56%

Three conventions matter. Every figure is a total return, dividends and coupons reinvested, so none is comparable with a headline price index. Every figure is nominal: no inflation adjustment appears anywhere, which flatters the 2022 window most, because that is when inflation was highest. And everything is gross of tax, trading cost and fees.

Two other pieces here put a 60/40 through the same 2007-09 crash at 22.4% and 31.4% rather than 26.48%; both use US rather than world equity, and one is monthly and real. The full reconciliation is there.

The Treasury series is built rather than published: no free daily total-return index for 10-year US Treasuries covers 1990 to 2026. This one comes from the Federal Reserve's daily 10-year constant-maturity yield -- each day a par bond yielding yesterday's rate is repriced at today's rate, with one day of coupon accrued. Compounded to calendar years and set against Aswath Damodaran's published annual 10-year Treasury returns, the two correlate at 0.993 across the 36 years from 1990 to 2025, a mean absolute difference of 0.70 percentage points a year. For 2008 the constructed series gives +20.5% against Damodaran's +20.1%; for 2022, -16.4% against -17.8%. That gap is the error bar on every Treasury figure below.

Government bonds lost money in two of six windows, and both were oil shocks

The 10-year Treasury has the cleanest record and the loudest exceptions. Peak to trough, it returned -0.8% in 1990, +8.6% in 1998, +39.5% through 2000-02, +19.3% through 2007-09, +8.5% through the 2020 crash, and -17.1% through 2022.

The chart above puts that in portfolio terms: percentage points of each equity fall absorbed by a 40% Treasury sleeve. The range runs from 34.97 points in 2000-02 to 3.82 in 2022 -- the same allocation, doing almost none of its job in the most recent episode.

Losing money and failing to protect aren't the same thing, though. The sleeve lost 0.8% in 1990 and still absorbed 8.55 of the 22.16 points equity gave up -- more than double the 2022 figure, and 0.9 points behind cash. Only 2022 is a protection failure on that measure. And -0.8% sits inside the constructed series' own error bar, so read the sign as indicative.

The correlations say the same from the other side. Across the full sample of 8,716 daily observations from July 1990 to May 2026, the correlation between Treasury and US equity daily returns is -0.203. Inside the four windows where Treasuries worked it was more negative still, between -0.36 and -0.49. Inside the two where they didn't, it flipped: +0.590 in 1990 and +0.068 in 2022.

The two losses have the same shape: the oil price rose hard enough to move inflation expectations, and a nominal government bond has no defence against that. West Texas Intermediate spot crude gained 116.8% over the 1990 window; in 2022 it was up as much as 50.9% before ending 7.2% higher. In both, the yield discounting a Treasury's fixed coupons rose while equity multiples fell, so the two legs lost together.

The 2022 damage is bigger than the window shows. Measured against its own running maximum, the constructed Treasury total-return index peaked on 4 August 2020, fell 27.1% to 19 October 2023, and as at 29 May 2026 was still 15.2% below that peak, 69.8 months later. On an inflation-adjusted measure the hole is deeper and longer still.

Gold's 2008 record is an endpoint, not a hedge

Gold ended four of the six windows higher: +13.2% in 1990, +1.8% in 1998, +12.1% through 2000-02 and +17.0% through 2007-09. It lost in the other two: -2.5% through the 2020 crash and -8.3% through 2022.

Those endpoint returns don't show what happened in between. Inside the 2007-09 window, gold rose 28.1% and then fell to 9.8% below where it began -- a 29.5% drawdown from its own high, taken while equity was in free fall. Anyone selling gold to buy equity during that window — the point of holding a hedge — was as likely to be selling it 30% down as 28% up. Inside the 2020 crash it fell 12.4% high to low over 28 trading days, and inside 2022 it fell 19.9%.

The correlation record is worse than the return record. Over the full sample gold's daily correlation with US equity is -0.004, near enough to zero to look like the textbook diversifier. In the 1990, 1998, 2000-02 and 2007-09 windows it stayed at or below zero. Then, in the fastest crash of the six, it went to +0.454: gold fell with stocks on the days stocks fell hardest, and recovered afterwards. That window runs 28 trading days, so the estimate is loose: its 95% interval is +0.10 to +0.71. It excludes zero, but only just, and every within-window figure here carries a band like it. Even so, that's a different product from a hedge that arrives on time.

Listed property is equity wearing a property label

The MSCI US REIT index, a gross total return in dollars from February 2005, is the clearest case of an asset class that looks like diversification on a fact sheet and isn't. Over the 5,175 trading days to May 2026 its daily correlation with US equity is 0.737 -- already high. Inside the 2007-09 window it rose to 0.815 and the index lost 68.9% against 54.2% for US equity: it didn't merely fail to protect, it amplified. Inside the 2020 crash the correlation reached 0.945 and the loss was 42.8%. Inside 2022 it lost 26.1% against equity's 25.5%, with a deeper interim fall of 30.5%.

Emerging-market equity, on the same panel, is a milder version: -62.5% in 2007-09, -31.5% in 2020 and -29.6% in 2022, with a full-sample correlation of 0.452 that rose to 0.516 and 0.622 in the first two. Corporate credit went the same way. The spread between Moody's Baa corporate yield and the 10-year Treasury started the 2007-09 window at 1.99 percentage points and ended at 5.40, touching 6.16 on 4 December 2008; in 2020 it went from 2.06 to 4.31 in 28 trading days.

International equity co-moved more, but nowhere near perfectly

The claim that correlations go to 1 in a crisis is the most repeated line here, and the data only partly supports it.

Daily correlation is a poor instrument here, because the markets trade in different hours: London and Tokyo close before most of a US move happens. Across the full sample the US-to-developed-ex-US daily figure is 0.505 and the weekly figure 0.723 — that gap is the clock rather than the economics. Japan is the extreme case: its daily correlation with US equity over 36 years is 0.049, which reads like perfect diversification and means almost nothing.

Weekly returns have to close on some weekday, and the choice moves the answer. On weeks ending Wednesday, the convention used here, that correlation was 0.785 in 2000-02 (133 weeks), 0.793 in 2007-09 (71) and 0.829 in 2021-22 (49), against 0.723 for the full 1,874-week sample. Close the week on any of the other four weekdays and the windows land between 0.726 and 0.879, the full sample between 0.723 and 0.748. Each window interval overlaps the full-sample one anyway: 2000-02 runs [0.71, 0.84] against [0.70, 0.74]. Co-movement did rise, in fourteen of those fifteen anchor-window combinations, and it's nowhere near 1 -- but how big the rise looks is partly the calendar. The separate question of how much a portfolio holds abroad turns on more than this.

The strongest objection: crisis correlations are biased upward by construction

There's a serious statistical case that the whole exercise above overstates its finding, and Kristin Forbes and Roberto Rigobon set it out in NBER working paper 7267. Selecting a subsample by picking the period when one market was volatile mechanically raises its measured correlation with any other market, even when the true relationship hasn't changed. In their words: "The measure of cross-market correlations central to this standard analysis, however, is biased." Their equation 8 corrects for it, dividing the measured correlation by a factor that grows with the relative rise in the volatile market's variance.

The correction changes the picture materially. US equity's daily variance inside the 2020 window was 15.6 times its full-sample variance. The measured US-to-developed-ex-US correlation of 0.710 in that window falls to 0.248 once adjusted -- below the full-sample 0.505. In the 2007-09 window, variance was 4.4 times normal and the measured 0.553 falls to 0.302. In 1998 it falls from 0.460 to 0.333, and in 2000-02 from 0.420 to 0.330.

One window survives the correction, and it's the slow one. In 2021-22, US equity variance was only 1.8 times its full-sample level, so the adjustment is small: the measured 0.663 becomes 0.551, still above the 0.505 baseline. The slow decline is where international diversification genuinely got weaker; the fast crashes mostly produced a volatility artifact dressed as contagion. That is a sharper distinction than "correlations rise in a crisis".

Two caveats belong with the correction. Forbes and Rigobon are explicit that it rests on the shock starting in the market whose variance rose, and on no unobserved common factor: "These assumptions are critical to obtain the results reported in this paper." And the adjustment says nothing about money: an asset can have a low adjusted correlation and still lose 30% alongside everything else, as emerging-market equity did.

Treasury bills never lost, in any window, by any amount

The dullest line in the table is the only one that didn't lose. Three-month Treasury bills returned +1.5%, +1.1%, +9.9%, +2.0%, +0.1% and +1.1% across the six windows, with a maximum drawdown of exactly zero in each.

Against Treasuries the record is mixed rather than inferior. A 40% bill sleeve beat a 40% Treasury sleeve in the two oil-shock windows -- by 0.9 points in 1990 and 7.3 points in 2022 -- and lost to it in the other four, by 3.0 points in 1998, 11.8 in 2000-02, 6.9 in 2007-09 and 3.4 in 2020. Bills give up the duration rally that made Treasuries worth 39.5% through 2000-02, and the duration risk that cost 17.1% in 2022. Which looks better depends on whether the next 20% equity fall arrives with falling or rising inflation. A holdings tracker such as LedgerTouch shows what the mix is; it can't show which shock is coming.

What this data cannot settle

Six windows is six observations, four drawn from a period of near-uninterrupted US disinflation. The bond record is a sample from an era, not a law: this series begins in 1990, nine years after the 10-year yield peaked at 15.84% on 30 September 1981, and one starting in 1968 would contain different failures.

The panel is also incomplete by construction. It carries developed-world, US and developed ex-US equity, 10-year Treasuries, Treasury bills, gold and crude oil from 1990, and adds US REITs and emerging-market equity only from 2005. It contains no hedge funds, private credit, unlisted property or managed-futures strategies, because no daily published total-return series for those was available -- so this piece can't say whether those diversified. The oil figure is a spot price, not an investable commodity index: a futures fund also earns a roll and collateral return that spot ignores.

Days on which any one of these markets was closed were dropped, costing 653 of 9,370 trading days. Every source was truncated at 29 May 2026, the last date in the Fama/French files, so no figure mixes vintages. And correlation summarises average co-movement, not payoff -- which is why the return and correlation columns disagree as often as they do.

What would change the conclusion

A crisis that arrives with falling inflation would restore the bond result. The Treasury record splits cleanly on that one variable. Four windows with disinflation or deflation gave returns between +8.5% and +39.5%; two with an oil-driven inflation shock gave -0.8% and -17.1%. Nothing in the data says which comes next, and 1990 shows the inflationary version isn't unique to 2022.

A longer window would rehabilitate gold, and a longer one still would damn it. Gold's failures in these episodes are timing failures: it fell with equity and recovered afterwards. Against its own history it looks worse -- the London afternoon price reached $850.00 on 21 January 1980, bottomed at $252.80 on 20 July 1999, 70.3% lower, and did not close above the 1980 figure again until 3 January 2008, 28 years later. Those are nominal dollars. After inflation, clearing that same 1980 top took at least 45 years, not 28. The evidence here is about what gold did in the weeks that hurt, not over a decade.

Correcting the correlations differently would change how much of the convergence survives. The Forbes-Rigobon adjustment assumes the shock starts in US equity. Applying it with developed ex-US equity as the source market gives different numbers. So would a genuinely global shock: the paper's own footnote 7 says that if endogeneity or unobservable aggregate shocks exist, "the adjustment to the correlation coefficient is slightly different than that presented here". The same footnote then reports that such shocks "have little impact on the results reported in this paper", and that alternative procedures addressing them "reinforce the results reported below". So treating the raw figures -- 0.710 in 2020, 0.663 in 2022 -- as the better estimate departs from the authors' own reading rather than applying it. The caveat bounds the size of the correction, not its direction.

A different reference index would move the window dates. These six come from developed-world equity in dollars. A sterling or euro investor, or one measuring against an index that includes emerging markets, would get peaks and troughs days or weeks apart -- and in the 1990 and 1998 cases, both close to the 20% threshold, might get five windows rather than six. The 2007-09, 2020 and 2022 episodes sit far enough past it that no reasonable index choice removes them.

The number worth carrying out of this is not a correlation. It is the 3.8 points Treasuries absorbed of equity's 26.7-point fall in 2022, against the 30.6 points they absorbed of the 57.0-point fall in 2007-09. Same two assets, same allocation, an eightfold difference in what the second one did. Diversification is worth a different amount in each crisis, set by the kind of shock rather than by the allocation. What that implies for acting during the fall itself is a separate question again.

Sources

  1. Kenneth R. French, Data Library — Fama/French Developed 3 Factors [Daily] (Developed_3_Factors_Daily_CSV.zip). Downloaded 27 July 2026; the file states it was created from the 202605 Bloomberg database and its last observation is 29 May 2026. The developed-world equity total return used throughout is the market factor plus the risk-free rate (Mkt-RF + RF), in US dollars with dividends included. Supplies the running-maximum crisis-window rule and its six windows (17 Jul 1990-28 Sep 1990, -22.16%; 20 Jul 1998-8 Oct 1998, -21.39%; 24 Mar 2000-9 Oct 2002, -47.97%; 31 Oct 2007-9 Mar 2009, -57.03%; 12 Feb 2020-23 Mar 2020, -33.79%; 8 Nov 2021-12 Oct 2022, -26.69%), the 60/40 peak-to-trough figures, and the 8.55/12.00/34.97/30.55/16.90/3.82 percentage-point absorption figures. (mba.tuck.dartmouth.edu)
  2. Kenneth R. French, Data Library — Fama/French 3 Factors [Daily] for the US (F-F_Research_Data_Factors_daily_CSV.zip). Downloaded 27 July 2026; created from the 202605 CRSP database, last observation 29 May 2026. US equity total return is Mkt-RF + RF. Supplies the US peak-to-trough figures (-54.19% in 2007-09, -33.89% in 2020, -25.50% in 2021-22), the full-sample daily correlations with Treasuries (-0.203), gold (-0.004), T-bills (0.004) and oil (0.095) over 8,716 daily observations from 3 July 1990 to 29 May 2026, and every within-window correlation and Forbes-Rigobon variance ratio. (mba.tuck.dartmouth.edu)
  3. Kenneth R. French, Data Library — Fama/French Developed ex US 3 Factors [Daily] (Developed_ex_US_3_Factors_Daily_CSV.zip). Downloaded 27 July 2026; created from the 202605 Bloomberg database, last observation 29 May 2026. Supplies the US-to-developed-ex-US daily correlation of 0.505 and the weekly correlation of 0.723 across 1,874 weeks, the within-window daily correlations (0.500, 0.460, 0.420, 0.553, 0.710, 0.663) and the developed ex-US peak-to-trough returns. Weekly returns are taken on weeks ending Wednesday, giving within-window figures of 0.785 (133 weeks), 0.793 (71) and 0.829 (49). Repeating the exercise on the other four weekday anchors gives full-sample figures of 0.748 (Monday), 0.725 (Tuesday), 0.731 (Thursday) and 0.732 (Friday), and window figures spanning 0.726 to 0.879 — negative against its own full sample in one of the fifteen anchor-window combinations, the Monday reading for 2000-02. (mba.tuck.dartmouth.edu)
  4. Kenneth R. French, Data Library — Fama/French Japan 3 Factors [Daily] (Japan_3_Factors_Daily_CSV.zip). Downloaded 27 July 2026; created from the 202605 Bloomberg database, last observation 29 May 2026. Source of the 0.049 full-sample daily correlation between Japanese and US equity total returns in US dollars, used to show that a near-zero daily cross-border correlation is a trading-hours artifact. (mba.tuck.dartmouth.edu)
  5. Federal Reserve Bank of St. Louis, FRED series DGS10 (Market Yield on U.S. Treasury Securities at 10-Year Constant Maturity, daily, percent) — CSV downloaded 27 July 2026, series running 2 January 1962 to 23 July 2026 and truncated at 29 May 2026 for this analysis. This is the input to the constructed 10-year Treasury total-return index described in the article: each day a par bond yielding the previous day's rate is repriced at the current rate and one day of coupon accrued. It supplies the Treasury peak-to-trough returns (-0.78%, +8.59%, +39.45%, +19.34%, +8.46%, -17.14%), the -27.08% nominal drawdown from 4 August 2020 to 19 October 2023, and the 15.22% shortfall still outstanding at 29 May 2026. (fred.stlouisfed.org)
  6. Federal Reserve Bank of St. Louis, FRED series DTB3 (3-Month Treasury Bill Secondary Market Rate, Discount Basis, daily, percent) — CSV downloaded 27 July 2026, series running 4 January 1954 to 23 July 2026 and truncated at 29 May 2026. Accrued daily on a 360-day basis to give the Treasury bill total return used for the +1.52%, +1.06%, +9.85%, +1.99%, +0.10% and +1.13% window returns, the zero maximum drawdown in all six windows, and the 60/40-in-bills comparisons. (fred.stlouisfed.org)
  7. Federal Reserve Bank of St. Louis, FRED series DCOILWTICO (Crude Oil Prices: West Texas Intermediate, Cushing, Oklahoma, daily, dollars per barrel) — CSV downloaded 27 July 2026, series running 2 January 1986 to 20 July 2026 and truncated at 29 May 2026. Source of the 116.8% gain across the 1990 window and the 50.9% intra-window high and 7.2% close in the 2021-22 window. A spot price, not an investable commodity index. (fred.stlouisfed.org)
  8. Federal Reserve Bank of St. Louis, FRED series BAA10Y (Moody's Seasoned Baa Corporate Bond Yield Relative to Yield on 10-Year Treasury Constant Maturity, daily, percentage points) — CSV downloaded 27 July 2026, series running 2 January 1986 to 23 July 2026. Source of the spread widening from 1.99 to 5.40 points across the 2007-09 window with a peak of 6.16 on 4 December 2008, and from 2.06 to 4.31 across the 2020 window. (fred.stlouisfed.org)
  9. London Bullion Market Association, LBMA Gold Price PM auction history (JSON, US dollars, British pounds and euros per troy ounce) — a rolling file that serves the latest available history; downloaded 27 July 2026 with a last observation of 24 July 2026, truncated at 29 May 2026 for this analysis. Source of gold's window returns (+13.22%, +1.83%, +12.11%, +17.00%, -2.45%, -8.32%), the 29.54%, 12.44% and 19.85% intra-window drawdowns, the +0.454 correlation with US equity in the 2020 window, and the 21 January 1980 price of $850.00, the 20 July 1999 low of $252.80 (-70.3%) and the first close above $850.00 on 3 January 2008. (prices.lbma.org.uk)
  10. MSCI, End of Day Index Data — daily gross total-return levels in US dollars for MSCI index code 128456, the MSCI US REIT Index. Retrieved 27 July 2026. The endpoint returns a JSON object carrying only four fields — msci_index_code 128456, index_variant_type GRTR, ISO_currency_symbol USD, and an array of dated closing levels. It does not carry the index's name or a printed 'Index Level: Gross' line; those appear in MSCI's separate CSV export, which is a different endpoint. With the start date set to 1 January 1999 the response begins on 7 February 2005, the first level MSCI publishes for this index, and runs to 24 July 2026; truncated at 29 May 2026 here. Source of the -68.87% return in the 2007-09 window, -42.81% in 2020 and -26.13% in 2021-22 with a 30.45% intra-window drawdown, and the daily correlations with US equity of 0.737 (full 5,175-day sample), 0.815, 0.945 and 0.772. (app2.msci.com)
  11. MSCI, End of Day Index Data — daily gross total-return levels in US dollars for MSCI index code 891800, the MSCI Emerging Markets Index. Retrieved 27 July 2026. As with the REIT endpoint, the JSON carries only the index code, the variant type GRTR, the currency USD and an array of dated closing levels; the series name and index-level labelling come from MSCI's separate CSV export. With the start date set to 1 January 1999 the response begins on 29 December 2000, the first level MSCI publishes for this index, and runs to 24 July 2026; truncated at 29 May 2026 here. Source of the -62.50%, -31.51% and -29.59% window returns and the 0.452 full-sample, 0.516 and 0.622 within-window daily correlations with US equity. (app2.msci.com)
  12. Aswath Damodaran, Historical Returns on Stocks, Bonds and Bills: United States (histretSP.xls), NYU Stern. Downloaded 27 July 2026; the workbook's document properties record it as last saved 14 May 2026 and the 'Returns by year' sheet runs to 2025. Used only to validate the constructed 10-year Treasury total-return series against an independently published one: correlation 0.993 across the 36 calendar years 1990-2025, mean absolute difference 0.70 percentage points a year, +20.5% against Damodaran's +20.1% for 2008 and -16.4% against -17.8% for 2022. (pages.stern.nyu.edu)
  13. Kristin J. Forbes and Roberto Rigobon, 'No Contagion, Only Interdependence: Measuring Stock Market Co-movements', NBER Working Paper 7267 (July 1999). Source of the quoted sentence 'The measure of cross-market correlations central to this standard analysis, however, is biased.' (abstract), of the correction in the paper's equation 8 that divides the measured correlation by the square root of one plus the relative rise in the volatile market's variance times one minus the squared measured correlation, and of footnote 7 in full, which states that the identifying assumptions 'are critical to obtain the results reported in this paper', that if endogeneity or aggregate unobservable shocks exist 'the adjustment to the correlation coefficient is slightly different than that presented here', and then that 'Preliminary results from Rigobon [1998] suggest that any endogeneity and/or unobservable aggregate shocks have little impact on the results reported in this paper. Alternative test procedures which address these problems reinforce the results reported below.' Retrieved 27 July 2026. (nber.org)

Research Disclosure

This content is for informational purposes only and does not constitute financial advice. Always do your own research or consult a qualified financial advisor before making investment decisions.

Published · Last reviewed . Data can revise after publication, so validate critical figures at source before making allocation changes.