Research · Dataset
Is gold a safe haven? The answer changes with the sampling frequency
Gold is described as a safe haven often enough that it has become a background assumption rather than a claim. The problem is that every published version of the claim tends to use monthly data, and monthly data is a filter. This page takes the same underlying price series and samples it three ways — every trading day, each month-end, and each month's average — so that the only thing changing is the frequency. The daily answer is that gold does not hedge equities on any day-to-day basis. The monthly answer is that it appears to. Both are computed here from the same fixings, and the gap between them is the actual finding.
Data through 2026-11-10 · updated 2026-11-15
Source and method
The headline numbers
Gold and equities are essentially unrelated day to day. That is not the same as gold being a safe haven, and the difference is worth stating plainly before going further: no correlation means gold does not fall when equities fall, and no hedge means gold does not rise when equities fall. A tail event study separates the two. On the worst 1% of equity days, gold's average same-day return was +0.09% and it rose on 50.0% of them — a coin flip, not protection.
One more distinction, because it is easy to conflate: an equity drawdown is not the same thing as financial-system stress. Gold's behaviour in the weeks when credit markets were genuinely freezing is tested separately against the Chicago Fed and St. Louis Fed financial-conditions indexes in our financial-stress study, which finds the same flat result across the full sample and a measurably worse outcome in the most extreme tail.
Step one: the same series, sampled three ways
Everything below rests on one control. The three paths are built from the same daily fixings — the daily path is the fixings themselves, the month-end path takes the last fixing of each calendar month, and the month-average path takes the mean of each month's fixings. The equity leg is aggregated to match. So any difference between the rows below is a sampling artefact, not a difference in the underlying market.
| Path | Periods | Gold–equity correlation | Gold annualised vol | Equity annualised vol | Worst-1% sample |
|---|---|---|---|---|---|
| daily | 13,722 | +0.0001 | 19.90% | 20.39% | 138 |
| month-end | 669 | +0.0144 | 19.73% | 20.41% | 7 |
| month-average | 669 | +0.0033 | 17.03% | 17.47% | 7 |
Read the last column carefully. "The worst 1%" sounds identical on all three rows and is not: on the daily path it is 138 events, on the monthly paths it is 7. The monthly rows do not have more evidence about a crash; they have twenty times less of it. Volatility annualises consistently on each row (252 trading days, 12 months), which is why the three gold volatility figures land within a few points of each other even though the raw frequency differs by a factor of 21.
Step two: what gold did after the worst equity days
The event study asks a narrow question: on days when equities fell hard, what did gold do on that day and over the next few weeks? Each forward return is compared with the unconditional gold return over the same horizon, because gold drifted upward across this sample and a 20-day forward return measured against a 1-day baseline would manufacture an "excess" out of that drift alone.
| Equity tail | Events | Gold that day | Gold up? | Horizon | Gold forward | Unconditional | Excess | 95% interval |
|---|---|---|---|---|---|---|---|---|
| worst 1% | 138 | +0.094% | 50.0% | 1 d | -0.118% | +0.042% | -0.160 | -0.409 to +0.060 |
| 5 d | +0.062% | +0.208% | -0.145 | -0.874 to +0.558 | ||||
| 20 d | +0.587% | +0.844% | -0.257 | -2.056 to +1.383 | ||||
| worst 2% | 275 | +0.082% | 52.7% | 1 d | +0.010% | +0.042% | -0.032 | -0.176 to +0.107 |
| 5 d | +0.236% | +0.208% | +0.028 | -0.481 to +0.511 | ||||
| 20 d | +0.961% | +0.844% | +0.117 | -1.055 to +1.283 | ||||
| worst 5% | 687 | +0.083% | 54.7% | 1 d | +0.035% | +0.042% | -0.006 | -0.106 to +0.093 |
| 5 d | +0.155% | +0.208% | -0.053 | -0.357 to +0.255 | ||||
| 20 d | +1.068% | +0.844% | +0.224 | -0.553 to +1.066 | ||||
| worst 10% | 1,373 | +0.069% | 53.2% | 1 d | +0.017% | +0.042% | -0.025 | -0.098 to +0.051 |
| 5 d | +0.068% | +0.208% | -0.140 | -0.361 to +0.086 | ||||
| 20 d | +0.855% | +0.844% | +0.011 | -0.570 to +0.613 |
Every interval on the daily path crosses zero. At the tightest threshold — the worst 1% of equity days — gold's 20-day excess return was -0.257 percentage points with a 95% interval of -2.056 to +1.383, and the 1-day excess was -0.160 points with an interval of -0.409 to +0.060. At the loosest threshold (worst 5%, 687 days) the 20-day excess was +0.224 points, interval -0.553 to +1.066. With 13,722 daily observations a real effect would be visible; there is nothing here to see.
Step three: the monthly version, and why it looks different
The same event study on the monthly paths produces a much friendlier picture. On the month-average path, after the worst 1% of equity months, gold's 20-month forward return was +15.95% against an unconditional baseline of +20.76%, an excess of -4.82 percentage points. On the month-end path the corresponding figures are +36.90%, +20.66% and +16.25 points.
| Path | Tail | Events | Gold forward (20 periods) | Unconditional | Excess | 95% interval | Verdict |
|---|---|---|---|---|---|---|---|
| daily | worst 1% | 138 | +0.587% | +0.844% | -0.257 | -2.056 to +1.383 | crosses zero |
| month-end | worst 1% | 7 | +36.90% | +20.66% | +16.25 | -23.82 to +66.77 | crosses zero |
| month-average | worst 1% | 7 | +15.95% | +20.76% | -4.82 | -22.58 to +11.21 | crosses zero |
None of these is a significant effect either. But read the two monthly rows against each other, because they are built from the same gold fixings and differ only in how a month is summarised. Month-end gives an excess of +16.25 points; month-average gives -4.82 points. Same data, same percentile, opposite sign. That is not a subtle difference in magnitude — it is the monthly framing failing to pin down a direction at all, with intervals wide enough (-23.8 to +66.8 and -22.6 to +11.2) to swallow either one. The daily row, built from 138 observations of the same event, gives a tight interval around a small negative number, which is a genuinely different kind of answer.
Note also what the monthly rows do show: the 20-period gold return after a worst-1% equity month is close to the unconditional 20-period gold return (+15.95% against +20.76% on the month-average path). What looks like hedging in the monthly framing is mostly gold's own drift — the same drift that made gold a good asset over 56 years and that has nothing to do with equities falling.
Step four: is the relationship stable, or does it come and go?
A relationship that appears in some decades and reverses in others is a sample artefact, not a property of the metal. The decade-by-decade daily correlations below range from -0.0967 to +0.1387 and change sign. Gold's average return after the worst 5% of equity days in each decade was positive in 4 decades and negative in 2.
| Decade | Days | Daily correlation | Gold after worst-5% days | Events | Gold up |
|---|---|---|---|---|---|
| 1970s | 2,206 | -0.0967 | +0.581% | 111 | 59.5% |
| 1980s | 2,466 | +0.0887 | -0.219% | 124 | 46.0% |
| 1990s | 2,467 | -0.0640 | +0.058% | 124 | 47.6% |
| 2000s | 2,451 | -0.0561 | +0.153% | 123 | 54.5% |
| 2010s | 2,452 | -0.0069 | +0.154% | 123 | 61.8% |
| 2020s | 1,680 | +0.1387 | -0.277% | 85 | 40.0% |
There is also no lead-lag structure to exploit. Across lags from minus five to plus five trading days the largest correlation in absolute terms is 0.0217, and the regression slope has a Newey-West HAC t-statistic of 0.02 (p = 0.99), which is as close to "no relationship" as a statistical test gets. Gold's own one-day autocorrelation is -0.0182.
Step five: a severity marker instead of a return threshold
The tail thresholds above are defined by the equity leg's own distribution. A different way to isolate stress is to use the VIX directly: on days when implied volatility was above a given level, what did gold do? This is a weaker test — the VIX is high on many days that are not crashes — but it separates "the market was frightened" from "the market fell".
| VIX level | Days | Equity that day | Gold that day | Gold up | Gold 20d forward | Excess |
|---|---|---|---|---|---|---|
| ≥ 30 | 723 | -0.354% | +0.049% | 50.6% | +0.084% | -0.760 |
| ≥ 40 | 203 | -0.708% | +0.129% | 48.8% | +1.976% | +1.132 |
| ≥ 50 | 75 | -0.689% | -0.048% | 46.7% | +4.296% | +3.452 |
The middle row is the interesting one. On the 203 days with a VIX above 40, equities averaged -0.708% and gold averaged +0.129%; over the next 20 days gold's excess return was +1.13 points, with a 95% interval of -1.30 to +3.67. That is the largest apparent excess anywhere on this page and it still crosses zero. At VIX above 50 (75 days) gold's same-day return was actually -0.048% — in the most violent equity days of the last three and a half decades, gold sometimes sold too. Note that the VIX series only starts on 1990-01-02, so this test covers about 37 years rather than 56.
Where gold does help: the long stretch of a bear market
The daily event study looks for insurance against a single bad day. The other case for a safe haven is slower: over a multi-year equity bear market, does gold behave differently enough to blunt the damage? Here it does. The table lists the ten deepest drawdown episodes of a daily-rebalanced equity-only portfolio. An episode runs from the running peak to the running trough and is only closed when the index makes a new high, which is why the 2000–2002 bear and the 1973–1974 bear appear as two separate entries rather than one long one.
| Peak | Trough | Days | Equity | Gold | Gold max DD | 50/50 blend | Improvement |
|---|---|---|---|---|---|---|---|
| 2000-03-10 | 2002-10-09 | 630 | -77.93% | +10.03% | -12.20% | -47.15% | +30.78 pp |
| 1973-01-11 | 1974-10-03 | 430 | -59.90% | +139.35% | -29.13% | +1.11% | +61.01 pp |
| 2021-11-19 | 2022-12-28 | 265 | -36.40% | -3.10% | -20.12% | -20.23% | +16.17 pp |
| 1987-08-26 | 1987-10-28 | 43 | -35.89% | +4.35% | -3.47% | -17.68% | +18.21 pp |
| 1989-10-09 | 1990-10-16 | 252 | -33.00% | +0.51% | -18.38% | -17.18% | +15.82 pp |
| 1983-06-24 | 1984-07-25 | 268 | -31.50% | -19.30% | -22.18% | -25.24% | +6.26 pp |
| 2020-02-19 | 2020-03-23 | 23 | -30.12% | -4.91% | -12.44% | -17.96% | +12.15 pp |
| 1998-07-20 | 1998-10-08 | 56 | -29.55% | +1.83% | -7.29% | -14.94% | +14.61 pp |
| 1981-05-29 | 1982-08-13 | 299 | -28.79% | -29.47% | -38.59% | -28.40% | +0.39 pp |
| 1980-02-08 | 1980-03-27 | 33 | -24.91% | -29.88% | -32.23% | -26.86% | -1.95 pp |
Gold rose during 5 of the 10 deepest episodes. In the deepest one, 2000-03-10 to 2002-10-09, equities fell 77.9% while gold rose +10.0% and the 50/50 blend fell 47.2% instead — +30.8 percentage points of damage avoided. But the table also shows the catch: in the 1981–1982 and 1980 episodes gold fell as hard as equities, with its own drawdowns of -38.6% and -32.2%. Gold is not a defensive asset in every crisis.
Over the full 55.8-year sample the equity-only portfolio's maximum drawdown was 77.9% and the daily-rebalanced 50/50 blend's was 47.2%. The blend compounded at 10.65% a year against 10.41% for equities alone, so the drawdown reduction cost 0.24 percentage points of annual return here. That is a reasonable trade on these numbers, and it is worth being precise about why it works: not because gold rose on the bad days, but because gold's own drawdowns did not perfectly coincide with equity's. Gold's worst drawdown over the sample was 70.3% and its compound annual growth rate was 8.71% against 10.41% for the index.
What this does not say
- A daily LBMA PM fix is a benchmark auction price, not a tradable bid/ask. Intra-day excursions inside a day are invisible, so this test can only say that gold did not move on the fixing; it cannot say gold did not trade up during the day.
- NASDAQCOM is a price index and excludes dividends, while gold here is a spot price with no carry. Both legs are therefore total-return-less, which keeps them comparable but understates the equity leg's own total return.
- No spread, slippage, financing, or margin cost is modelled anywhere on this page.
- The equity leg is a single index. A safe-haven claim about a diversified portfolio would need portfolio-level data, not one index.
- The VIX leg only exists from 1990, so the severity analysis covers about 36 years, not 55.
- This is not a live account and there is no broker statement behind these numbers.
- The site's published monthly hedge page uses the World Bank Pink Sheet monthly average for gold rather than the LBMA fix, so its gold leg is a different series; the monthly numbers on this page are computed from the LBMA fix and are not expected to match it exactly.
Nothing on this page is a forecast, a recommendation, or a claim about any account. The numbers describe what two price series did over a past window, and past windows do not obligate the future.
Frequently asked questions
Does gold actually protect you in an equity crash?
Not on any single-day basis, and the honest answer is that it does not hedge day to day at all. Over 13,722 aligned trading days from 1971-02-08 to 2026-11-10, the daily correlation between gold and the Nasdaq Composite was 0.0001 and the regression beta of gold on equities was 0.0001, with an R² of 0. On the worst 1% of equity days (138 of them), gold's average same-day return was 0.094% and it rose on 50% of them. That is a coin flip, not a hedge.
Then why does the monthly version of this test look so much better?
Because monthly sampling hides the daily behaviour, and it cannot pin down a direction at all. On the month-end path the same analysis finds gold averaging -2.29% on the worst 1% of equity months, and the 20-month forward return after such a month was 36.9%. But that "worst 1%" is only 7 months out of 669 — seven observations, where the daily path gives you 138 observations for the same nominal percentile. Worse, the two monthly paths disagree: the month-average path built from the same fixings gives a 20-month excess of -4.82 percentage points against month-end's 16.25 — opposite signs, intervals wide enough to contain either.
What is the fairest way to compare the two?
Compare against the unconditional baseline at the same horizon. Over the 669-month path, gold's unconditional 20-month forward return was 20.66%, so the "excess" after a worst-1% equity month was 16.25 percentage points. On the daily path the unconditional 20-trading-day forward return was 0.844% and the excess after a worst-1% equity day was -0.257 points. Comparing a 20-day forward return against a 1-day baseline — which is a common mistake in this kind of write-up — would manufacture an "excess" out of gold's own drift.
Is any of the daily tail result statistically meaningful?
None of it. After the worst 1% of equity days, gold's 20-day excess return was -0.257 percentage points with a moving-block bootstrap 95% interval of -2.056 to 1.383 — the interval spans zero. The 1-day and 5-day horizons behave the same way. At the loosest threshold (worst 5% of days, 687 events) the 20-day excess was 0.224 points with an interval of -0.553 to 1.066, again crossing zero. With 13,722 daily observations you can detect small effects; there simply are none here to detect.
So does gold have ANY relationship to equities?
None that survives scrutiny, and no leading one either. The largest correlation across lags from minus five to plus five trading days is 0.0217, which is indistinguishable from zero. The unconditional beta has a Newey-West HAC t-statistic of 0.02 (p = 0.99), so the slope is not distinguishable from zero. Gold's own one-day autocorrelation is -0.0182. The decade-by-decade correlations range from -0.0967 to 0.1387, and the sign flips: 4 of 6 decades show a positive average gold return after the worst 5% of equity days, and 2 show a negative one.
Where DOES gold help?
Over the long stretch of a genuine bear market, not on the day of the crash. In the deepest equity drawdown in the sample, 2000-03-10 to 2002-10-09, equities fell 77.9% while gold rose 10%, and a 50/50 portfolio fell 47.2% instead of 77.9%. Across the ten deepest equity drawdowns, gold rose during 5 of them. The full-sample maximum drawdown of the equity-only portfolio was -77.9% against -47.2% for a daily-rebalanced 50/50 blend. Gold's contribution is path shape, not daily insurance.
Did the 50/50 portfolio cost me return?
Very little, over this sample. A daily-rebalanced 50/50 portfolio compounded at 10.65% a year against 10.41% for equities alone, over 55.8 years — a gap of 0.24 percentage points a year in exchange for cutting the worst drawdown from 77.9% to 47.2%. But note what this is not: the blend did not lose much because gold was uncorrelated on bad days; it lost little because gold had its own very large drawdowns, -70.3% at its worst, in the same periods.
What are the limits of this study?
They are substantial and they bound the conclusion. A daily LBMA PM fix is a benchmark auction price, not a tradable bid/ask, so intra-day moves are invisible: this test can say gold did not move on the fixing, but not that gold did not trade up during the day. The Nasdaq Composite is a price index with no dividends while gold here carries no financing cost, so neither leg is a total return. No spread, slippage, margin or financing cost is modelled anywhere. The equity leg is a single index, and a safe-haven claim about a diversified portfolio would need portfolio-level data. The VIX leg only starts on 1990-01-02, so the severity analysis covers about 37 years rather than 56. And there is no live account behind any of this.
How is this different from the existing monthly hedge page?
The site's earlier hedge analysis uses the World Bank Pink Sheet monthly average for gold and a monthly equity series; this page uses the LBMA PM daily fixing and rebuilds the monthly paths from that same daily source, so the only thing that varies between the daily, month-end and month-average paths here is the sampling frequency. That is what makes it a controlled comparison rather than a second opinion. It also means the monthly numbers on this page will not match the other page exactly, because the gold series is different.
Sources
-
LBMA Gold Price PM benchmark fixings, via
scripts/daily_gold_lbma_pm.csv(LBMA PM gold price benchmark (daily fixing), continued by World Gold Council FSAPI from 2026-02-26) -
Nasdaq Composite daily close, FRED series
NASDAQCOM(FRED NASDAQCOM — Nasdaq Composite daily close (price index, no dividends)) -
CBOE Volatility Index daily close, FRED series
VIXCLS(FRED VIXCLS — CBOE Volatility Index daily close (1990 onward)) -
All figures recomputed by
scripts/gold_daily_safe_haven_analysis.pyand rechecked with numpy byscripts/verify_gold_daily_safe_haven.py. Accessed 2026-11-15.