Key takeaways
- Across 655 months from January 1972 to July 2026, gold's correlation with the US equity market was 0.01. In the worst 5% of those months it rose to 0.31 rather than turning negative.
- Gold still averaged 0.99% in those 33 worst months, which themselves averaged minus 9.94%, and the gold price rose in 21 of them.
- Correlation also rose in the best 5% of months, to 0.28, from 0.01 across the middle 90%. A number that climbs in booms and crashes alike is partly measuring the slice, not the metal.
- Gold fell in 11 of the 33, including minus 17.38% in October 2008 and minus 22.37% in March 1980, both months when equities were falling hard.
- A 10% gold sleeve funded from equities turned the average worst-5% month from minus 9.94% into minus 8.85%, an improvement of 1.09 percentage points.
Gold as a safe haven: the correlation goes up in a crash, and so does gold
Does gold rise when shares fall? That is the question underneath the gold safe haven claim, and the data gives an answer in two halves that look like they contradict each other.
Take the LBMA Gold Price PM, the London benchmark that ICE Benchmark Administration runs, and the monthly return on the whole US stock market from the Fama and French factor file. Line them up from January 1972 to July 2026. That is 655 months. Over all of them, the correlation between the two is 0.01. Gold and shares have essentially no month-to-month relationship.
Now take the worst 5% of those months for equities: 33 months, averaging minus 9.94%. The correlation does not go negative. It goes up, to 0.31.
And gold rose anyway. Across those same 33 months it averaged 0.99%, with a median of 2.33%, and it finished higher in 21 of them.
Both are true because correlation and direction answer different questions. Correlation asks whether the two move in step across a set of months. The average return asks which way gold actually went. In the worst months gold moved more in step with equities than usual, from a starting point of almost nothing, and still ended most of those months higher. If you have been told gold turns negatively correlated in a crisis, the record here doesn't support it. What the record supports is something weaker and more useful: gold usually goes up in the months shares fall hardest, and it does so without the correlation ever inverting.
What 655 months look like, sliced by how bad the month was
Here is the whole result in one table. Each row is a slice of the same 655 months, cut by the equity return and nothing else.
| Slice of months | Months | Average equity return | Gold correlation | Average gold return |
|---|---|---|---|---|
| All months | 655 | 0.99% | 0.01 | 0.85% |
| Middle 90% | 589 | 1.10% | 0.01 | 0.87% |
| Worst 10% | 66 | minus 7.78% | 0.21 | 1.32% |
| Worst 5% | 33 | minus 9.94% | 0.31 | 0.99% |
| Best 5% | 33 | 10.01% | 0.28 | 0.47% |
Read the correlation column downwards and the shape is clear in the chart above. It sits at 0.01 through the ordinary middle of the distribution, climbs to 0.21 in the worst decile and 0.31 in the worst twentieth. Gold's beta to equities tells the same story more sharply: 0.02 across all 655 months, 0.70 inside the worst 5%. In the tail, gold behaves like a partial equity.
Read the average-gold-return column instead and you get the opposite impression. Gold's best slice is the worst decile for equities, at 1.32%. Its weakest slice is the best 5% for equities, where gold averaged 0.47% and the median month was minus 0.54%. Gold has historically been dull when shares were flying and lively when they were falling apart.
The gold equity correlation rises in booms as well, which is the tell
Look again at the bottom row of that table. In the best 5% of equity months, correlation is 0.28, only a little below the 0.31 in the worst 5%, and both sit far above the 0.01 in the middle 90%. Whatever is happening, it isn't crisis behaviour, because it happens just as hard when equities are up 10% in a month.
This is a known trap with a name. Boyer, Gibson and Loretan set it out in a 1997 Federal Reserve International Finance Discussion Paper on what practitioners call correlation breakdown, the apparent collapse of a hedge exactly when it is needed. Their finding is that the collapse can be manufactured out of nothing. In their words, "correlation breakdowns" can easily be generated by data whose distribution is stationary and, in particular, whose correlation coefficient is constant. The point is narrower than it sounds. A correlation measured inside a subsample that was selected on one of the two variables is not the same statistic as the one measured on the whole sample, and comparing them is not a like-for-like comparison.
Their Table 2 puts numbers on it. For two variables whose true correlation is 0.20, the correlation measured inside a two-sided 5% tail event reads 0.434, while the same data outside the tail reads 0.175. Nothing about the variables changed. Only the partition did. Kristin Forbes and Roberto Rigobon reached the same conclusion for equity markets in NBER Working Paper 7267, circulated in 1999, writing that the unadjusted coefficient "is conditional on market movements over the time period under consideration, so that during a period of turmoil when stock market volatility increases, standard estimates of cross-market correlations will be biased upward."
How much of the jump from 0.01 to 0.31 is gold and how much is the partition, this data can't separate. What it can say is that the same jump happens on the other side, in months when equities rose 10%, and a crisis explanation has to account for that. The broader problem, that a correlation estimated in calm conditions is not the one you get later, is correlation instability, and it applies to every diversifier rather than to this one. What survives the objection is the return column, because averaging gold's return inside a set of months carries no such selection effect. Gold really did average 0.99% in the worst 33.
Eleven of the 33 went the wrong way, and four of them badly
The average hides the months a real holder would remember, because the record of gold in a market crash contains some outright failures. Gold fell in 11 of the 33 worst equity months, and in four it fell hard.
- October 2008: US equities minus 17.12%, gold minus 17.38%. The diversifier fell slightly further than the thing it was diversifying.
- March 1980: equities minus 11.69%, gold minus 22.37%, the sample's worst gold month inside an equity tail.
- September 2011: equities minus 7.58%, gold minus 10.67%.
- August 1998: equities minus 15.62%, gold minus 5.35%.
October 2008 is worth staying with, because it is the clearest failure in the set. It was also bracketed by two of gold's better months in the same drawdown: September 2008 saw gold rise 6.18% while equities fell 9.20%, and November 2008 saw it rise 11.46% while equities fell 7.71%. On a monthly grid, gold failed once in the middle of a crisis it otherwise handled. The months either side went the other way, which is worth holding onto before treating any single one of these figures as the answer.
For a sterling holder, the currency did some of the work
Everything above is in dollars. Gold trades in dollars, so a UK holder's return combines the dollar gold price with the exchange rate, and the exchange rate has its own behaviour in a crash. The LBMA publishes the same benchmark in sterling, so the conversion is already done.
Run the identical test on the LBMA sterling benchmark and the numbers improve. Across all 655 months the correlation with US equities is minus 0.05 instead of 0.01. In the worst 5% of equity months, sterling gold averaged 2.23% against the dollar figure of 0.99%, and it rose in 24 of the 33 rather than 21. The conditional correlation is higher, at 0.40, which is the same partition artifact showing up again on a different series.
The two failures look different too. October 2008 cost a sterling holder 7.85% rather than 17.38%. March 2020 was minus 0.06% in dollars and 3.34% in sterling. That gap is not a property of gold. The same ounce fell 17.38% in dollars and 7.85% in sterling, which can only mean the pound fell against the dollar, and a falling pound flatters a dollar-priced asset held in sterling. In an episode where the pound strengthened, the arithmetic would run the other way.
The academic record puts the safe haven window at about 15 trading days
The strongest objection to the monthly result is that a month is the wrong unit. Dirk Baur and Brian Lucey tested it directly in Is Gold a Hedge or a Safe Haven?, published in The Financial Review in 2010, using daily data for the US, UK and German markets from November 1995 to November 2005. They separate three ideas that get conflated: a hedge is uncorrelated or negatively correlated on average, a diversifier is positively but imperfectly correlated on average, and a safe haven is uncorrelated specifically in a crash.
They find gold is both a hedge and a safe haven for stocks, which supports the case here. Then they measure how long it lasts, and their conclusion is blunt: "We find that gold works as a safe haven asset only for around 15 days." The gold price jumps on the day of an extreme negative equity shock and gives the gain back over the following fortnight, because equities usually rebound and gold, being a hedge, moves the other way when they do.
That is a serious problem for any reading of the table above as a portfolio result. It says the 0.99% average is a snapshot of a window that closes. Somebody who saw the crash, did nothing, and looked again three months later may have had none of it. The honest version of the finding is that gold has historically been up while equities were down, in the same month, and that this says less about the following quarter than it appears to.
The investable version over the last decade looks weaker still
Bullion is not what most people own. Test the same idea on two exchange-traded proxies, SPDR Gold Shares (GLD) against the State Street SPDR S&P 500 ETF Trust (SPY), using 502 weekly returns between 13 September 2016 and 9 September 2026.
Three things change. The all-period correlation is 0.20, well above the metal's 0.01 across 655 months, though it drops to 0.04 across the middle 90% of weeks. In the worst 5% of weeks, when SPY averaged minus 5.75%, GLD averaged minus 0.99% and rose in only 10 of the 25. And the conditional correlation was higher in the best weeks, at 0.56, than in the worst, at 0.38, which is the both-tails signature again and this time with the boom side winning.
A weekly window over ten years is a different measurement from a monthly window over fifty, and one of them is a much smaller sample. But it is the measurement closest to what a person holding a gold ETF through the last decade experienced, and it is the weaker of the two. The same pattern shows up in crypto diversification, where the correlation with equities also rises in falling markets, with the difference that crypto's tail returns went negative rather than staying positive.
What the arithmetic is worth at a 5% or 10% weight
None of this matters at the whole-portfolio level unless the sleeve is big enough to move the total. Take the average worst-5% month, minus 9.94%, and fund a gold position out of the equity holding.
At 10% gold, the average worst month becomes minus 8.85%, an improvement of 1.09 percentage points. At 5% it is 0.55 points. In October 1987, the worst equity month in the sample at minus 22.59%, a 90/10 blend returned minus 20.13% because gold rose 2.02% that month. Two and a half points off a twenty-two point loss is real, and it is also not rescue.
The other side of the ledger runs for the remaining 622 months. Over the full sample gold compounded at 8.64% a year with annualised volatility of 19.8%, against 11.23% and 15.7% for the US equity market. That comparison is generous to gold in one direction and harsh in another: the equity figure is a total return including dividends, while gold pays no income at all, so its price change is the whole of what it earns. A gold sleeve bought a percentage point of tail cushion by holding an asset that returned less and moved more the rest of the time. Whether that trade is worth making is a question about the size of the sleeve, which is a separate argument covered in how much gold in a portfolio, and about how much of a bad month a particular holder can absorb before selling.
What this evidence cannot tell you
Start with whose story this contradicts. The World Gold Council's own correlations page says gold "decouples and becomes inversely correlated during periods of stress". Measured on the LBMA benchmark it doesn't invert, it rises. The Council says of itself that it was "formed in 1987 by some of the world's most forward-thinking mining companies", so it has an interest in the stronger claim, but the disagreement here is about which statistic to read rather than about the underlying data, and its description of the return behaviour is not far from what the table shows.
The bigger limitation is the sample. One gold series, one equity market, and one stretch of history. Gold went from 43.63 dollars an ounce at the end of 1971 to 4,026.60 at the end of July 2026, so the whole record sits inside one enormous bull market for the asset being tested. A backtest is not a forecast, and 33 tail months is a small sample in which every month carries a thirty-third of the average.
Monthly data also cannot see inside a month. A holder who sold in mid-October 2008 had a different experience from the month-end figure, and Baur and Lucey's fifteen-day result says that difference is exactly where the effect lives. Nothing here models dealing costs, ETC or trust fees, storage, spreads or tax, all of which come out of the sleeve and none of which come out of the index. And the conditional correlations are, by construction, a property of the slice as much as of the assets, which is the whole point of the section above.
What would change the conclusion
If the tail correlation stopped rising in the best 5% too, the rise in the worst 5% would stop looking like a partition artifact and start looking like behaviour. Right now the two move together, at 0.31 and 0.28, and that symmetry is the reason to treat the crisis number sceptically.
If gold's tail return turned negative, the whole case goes, because the return column is the only part of this that survives the statistical objection. The weekly ETF era has already produced a negative tail average, at minus 0.99%, on a much shorter sample. Another decade like it and the fifty-year figure stops being the relevant one.
If the crash is a gold liquidation, the average stops describing your month. October 2008 and March 1980 were both months when gold fell alongside equities, and there is nothing in this data that rules out the next one being the same.
The number to watch is not the correlation, which rises in every tail you cut and tells you very little for that reason. It is whether gold enters the crash already extended, as it was in early 1980, and how large the sleeve is relative to the loss it is being asked to offset. LedgerTouch tracks that weight continuously; a spreadsheet checked once a year tracks it too, and for a position this size that may be often enough.