Research · Dataset
Gold tail risk and what a stop-loss actually does, 1968–2026
Almost every gold trading conversation eventually arrives at one instruction: put a stop-loss on it. This page takes that instruction seriously and tests it on 14676 daily fixing-to-fixing returns, alongside the tail-risk numbers — Value at Risk, Expected Shortfall, excess kurtosis — that decide how big a position someone can actually carry. Two of the results are not what the folklore expects.
Data through 2026-11-10 · updated 2026-11-13
Source and method
VaR and Expected Shortfall are historical: VaR is the loss threshold exceeded on the worst 5% or 1% of days, and Expected Shortfall is the average loss conditional on being in that tail. They describe the past and imply nothing about the next crisis.
The headline numbers
Gold's daily standard deviation over this sample is 1.213%, which annualises to 19.26%. The distribution is not normal: excess kurtosis is 10.74, so the far tail is materially heavier than a bell curve implies.
The exceedance test makes that concrete. Moves beyond two standard deviations occurred on 2.44% of days against a normal benchmark of 5.40% — rarer than normal. Moves beyond three standard deviations occurred on 0.80% of days against 0.44% — 1.81× as often as normal predicts. Gold is placid most of the time and violent occasionally.
What the worst days looked like
| Window | Worst cumulative move | From | To |
|---|---|---|---|
| 1 day | -13.24% | 1980-01-22 | 1980-01-22 |
| 2 days | -18.24% | 1980-01-22 | 1980-01-23 |
| 3 days | -17.36% | 1980-03-13 | 1980-03-17 |
| 5 days | -26.59% | 1980-01-22 | 1980-01-28 |
| 10 days | -23.72% | 1980-03-05 | 1980-03-18 |
| 20 days | -26.83% | 1980-02-19 | 1980-03-17 |
The worst single day was -13.24% on 1980-01-22. But a single day is not how gold breaks accounts — the -26.59% five-day stretch from 1980-01-22 to 1980-01-28 is what a leveraged position actually experiences, because a stop that triggers on day one still leaves you flat into days two through five.
Ten worst single days on record: 1980-01-22 (-13.24%), 1983-02-28 (-12.10%), 2013-04-15 (-9.15%), 2026-01-30 (-7.83%), 2008-10-13 (-7.66%), 1980-03-17 (-7.46%), 1980-03-26 (-7.26%), 1980-01-04 (-7.26%), 1973-11-14 (-7.12%), 1980-02-20 (-7.09%).
How long does it take to get back?
The question after a bad day is always "how long until I am back to even". The median answer and the average answer are not the same number, and the gap between them is the whole story.
After a −5% day the median recovery took 28.5 trading days. The mean was 614 days, the 90th percentile was 1,136 days, and the longest was 6,985 days — 28 years. A stop-loss is often justified by "it recovers fast". It usually does recover fast; the ones that do not take a decade.
Recovery is measured as the first later fixing that regains the level in force immediately before the drop. Two of the 50 −5% events never recovered within the sample.
Is the tail stable across time?
A risk number is only useful if it does not change regime every few years. Rolling the one-year window forward gives a range rather than a single figure.
The 99% Expected Shortfall over a rolling year has ranged from 1.12% to 9.32% across 14425 windows — a factor of 8.3. A risk model calibrated on a calm year will be badly wrong in 1980, and a position sized off a calm year will not survive a 1980.
| Decade | Days | Annualised vol | Excess kurtosis | ES 99% | Normal ES 99% | Days ≤ −3% |
|---|---|---|---|---|---|---|
| 1960s | 441 | 10.12% | 3.62 | 2.38% | 1.70% | 0.00% |
| 1970s | 2522 | 22.56% | 6.34 | 5.04% | 3.79% | 2.30% |
| 1980s | 2498 | 26.20% | 10.48 | 6.49% | 4.40% | 2.88% |
| 1990s | 2502 | 11.96% | 10.50 | 2.94% | 2.01% | 0.24% |
| 2000s | 2499 | 18.67% | 4.85 | 4.36% | 3.13% | 1.56% |
| 2010s | 2498 | 15.60% | 6.06 | 3.67% | 2.62% | 0.60% |
| 2020s | 1716 | 18.03% | 4.04 | 4.39% | 3.03% | 0.93% |
The 1980s are the reference decade for tail risk: 26.20% annualised volatility and a 99% Expected Shortfall of 6.49%, against a normal-model figure of 4.40%. Every decade on this sample carries positive excess kurtosis.
Does a stop-loss help?
This is the test the page exists for. A stop-loss on gold is a position-management rule: exit when the loss reaches a threshold. Whether it helps depends almost entirely on what happens between the threshold and your exit — so both fill assumptions are simulated.
| Stop | Fill | Re-entry | Final | Annualised | Max drawdown | Stops hit |
|---|---|---|---|---|---|---|
| — | — | — | 75.14× | 7.70% | 70.33% | — |
| 3% | at the stop | immediate | 1,323.70× | 13.14% | 46.21% | 206 |
| 3% | at the stop | 5 days | 459.33× | 11.10% | 42.17% | 150 |
| 3% | at the close | immediate | 75.14× | 7.70% | 70.33% | 206 |
| 3% | at the close | 5 days | 69.91× | 7.57% | 49.80% | 150 |
| 5% | at the stop | immediate | 152.63× | 9.02% | 58.72% | 50 |
| 5% | at the stop | 5 days | 84.68× | 7.92% | 62.79% | 40 |
| 8% | at the stop | immediate | 84.45× | 7.91% | 67.08% | 3 |
| 8% | at the stop | 5 days | 87.82× | 7.99% | 65.04% | 3 |
| 10% | at the stop | immediate | 79.80× | 7.81% | 68.49% | 2 |
| 10% | at the stop | 5 days | 84.75× | 7.92% | 66.54% | 2 |
The full grid — stops of 3%, 5%, 8% and 10%, both fill assumptions, and both re-entry rules — is published with the dataset download. Buy and hold for reference: 75.14× over the sample, 7.70% annualised, 70.33% maximum drawdown.
Why the fill assumption decides everything
Every day that reached a stop closed beyond it.
On this series, all 206 days that reached a 3% stop also closed beyond it — 100% of them, not one exception. The average overshoot was 1.33 points and the worst was 10.24 points. At a 5% stop the same holds on all 50 trigger days, with an average overshoot of 1.32 points.
Gold does not travel to a stop level. It arrives there already through it. That is why a stop-loss is a claim about the future that the closing data refuses to support.
The two fill assumptions give opposite-looking answers, and the difference between them is the entire result.
Filled at the stop price, the 3% stop turns 75.14× into 1,323.70× and cuts the maximum drawdown from 70.33% to 46.21%. Filled at the close, the same stop produces 75.1378× — which is not close to buy and hold. It is exactly buy and hold. The reason is arithmetic, not subtle: if you sell at the closing fix on the day the stop triggers, you sold at the same price you would have sold at anyway. You paid the spread on a round trip and gave up the recovery. A stop that fills at the close does not cut the loss; it only moves the sale.
The optimistic fill is the one that flatters the stop, and it is the one that is least often true in practice. Neither assumption includes bid-ask spread, financing, margin, custody or tax. The optimistic column should be read as an upper bound on what a stop can do, not as an expected result.
Does re-entry timing matter?
A stop-loss on its own is half a rule. The other half is when you come back in. Standing aside for a few days after a stop means missing the recovery — and the recovery is where most of the return lives.
| Rule | Final | Annualised | Max drawdown | Time below prior peak | Stops hit |
|---|---|---|---|---|---|
| Buy and hold | 75.14× | 7.70% | 70.33% | 97.3% | — |
| 5% stop, filled at stop, immediate | 152.63× | 9.02% | 58.72% | 96.9% | 50 |
| 5% stop, filled at stop, 5-day wait | 84.68× | 7.92% | 62.79% | 97.0% | 40 |
| 5% stop, filled at close, 5-day wait | 46.89× | 6.83% | 72.70% | 97.5% | 40 |
Waiting five days after a 5% stop cut the annualised return from 9.02% to 7.92% under the optimistic fill, and made the close-filled version worse than buy and hold at 6.83% with a 72.70% drawdown. The recovery arrives quickly most of the time — the median is 28.5 trading days after a −5% day — and a cooling-off period spends that window out of the market.
The stop only stops out 50 times in 14676 observations. A rule that fires on a fraction of a percent of days cannot carry much of the risk-management load on its own.
What happens to the result after costs?
| Round-trip cost | Stop | Final | Annualised | Max drawdown |
|---|---|---|---|---|
| none | 3% | 1,323.70× | 13.14% | 46.21% |
| 10 bp | 3% | 1,077.16× | 12.74% | 47.27% |
| 25 bp | 3% | 790.40× | 12.14% | 48.83% |
| none | 5% | 152.63× | 9.02% | 58.72% |
| 10 bp | 5% | 145.18× | 8.92% | 59.13% |
| 25 bp | 5% | 134.67× | 8.78% | 59.74% |
| none | 8% | 84.45× | 7.91% | 67.08% |
| 10 bp | 8% | 84.19× | 7.91% | 67.14% |
| 25 bp | 8% | 83.81× | 7.90% | 67.24% |
A stop-loss turns the position over. Every exit is a round trip, so the cost is charged per stop, not per year. Even at 25 basis points per round trip — a generous figure for a liquid market — the 3% stop still produced 12.14% annualised against 13.14% before costs. Costs are not what decides this question. The fill assumption is.
Costs here are a flat charge on each exit and re-entry. They do not model spread on entry, financing, margin, custody or tax.
Is there a better day of the week to trade gold?
The "buy gold on Friday" rule is one of the most searched gold patterns, so this page tests it properly rather than reporting the statistic that looks best.
| Weekday | Days | Mean return | Standard deviation | t against zero |
|---|---|---|---|---|
| Mon | 2753 | +0.0320% | 1.3119% | 1.28 |
| Tue | 2995 | +0.0163% | 1.1925% | 0.75 |
| Wed | 2994 | +0.0614% | 1.2116% | 2.77 |
| Thu | 3007 | +0.0149% | 1.1933% | 0.69 |
| Fri | 2927 | +0.0595% | 1.1574% | 2.78 |
The individual t-statistics are a trap.
Wednesday and Friday both have average returns that look significant against zero: 0.0614% with t = 2.77, and 0.0595% with t = 2.78. The problem is that "is this weekday's mean different from zero" is not the question. The question is "do the weekday means differ from each other", and that is what a one-way ANOVA tests.
The answer is no: F = 1.0324 with p = 0.7809 on 4 and 14671 degrees of freedom. There is no detectable day-of-week effect in gold's daily returns over this sample. The Friday statistic is ordinary noise that happened to be tested against the wrong null.
Weekend gaps are included in the sample as returns from the last fixing before the weekend to the first fixing after it — they are real exposure. Holiday gaps longer than 5 calendar days are dropped so a missing week is not recorded as one large loss.
Where the tail stands right now
Over the last 63 fixing days, gold's realised volatility annualises to 21.07% and the 99% Expected Shortfall is 4.59% — the 74.8th percentile of that measure since 1968-04-02. Tail risk is running above its long-run norm without reaching the levels seen in 1980 or 2008.
A 63-day window is short and the percentile is a rank, not a forecast.
What this dataset does and does not support
Supported by the data: gold's daily return distribution has fat tails. Excess kurtosis is 10.74, the worst single day was -13.24%, and the worst five-day stretch was -26.59%. Expected Shortfall at 99% is 4.79% against 3.23% under a normal model, and the measure moves by a factor of 8.3 between its calmest and most violent rolling years. Gold also spent 97.3% of its history below a prior peak.
Not supported: the claim that a stop-loss protects gold positions on its own. On this series every single day that reached a 3% stop closed beyond it, and a stop filled at the close produces a result arithmetically identical to buy and hold (75.1378× against 75.1378×). A stop filled at the stop price does improve the drawdown, but that is the assumption the data cannot support, and it is the one most often wrong in practice. Nor is there any day-of-week effect: the ANOVA gives p = 0.7809.
These are arithmetic results on one historical price path, with no transaction cost, spread, financing, margin, custody or tax modelled except where stated. Nothing on this page is investment advice, a recommendation, or a forecast.
Frequently asked questions
Does a stop-loss make gold safer?
Not by itself, and the answer depends entirely on how the stop is filled. If the stop price trades and you are filled there, a 3% stop turns 75x into 1324x and cuts the maximum drawdown from 70.33% to 46.21%. But every one of the 206 days that reached a 3% stop also closed beyond it — 100% of them — with an average overshoot of 1.33 points. If you are filled at the close instead, the stop changes nothing at all: the final result is 75.14x, identical to buy and hold, because you sold at the same price you would have sold at anyway.
Why does a close-filled stop-loss produce exactly the buy-and-hold result?
Because it does not cut the loss, it only moves the sale. When the stop triggers you sell at the day's closing fix, which is the same price the next morning's buy-and-hold position would have been sold at. You pay the spread on a round trip and you give up the recovery. The result is not a small loss of edge — it is arithmetically identical to not having the stop at all: 75.1378x against 75.1378x for buy and hold.
Is the gap through the stop the real risk?
Yes, and it is total on this series. All 206 days that touched a 3% stop closed beyond it, not one exception. The worst overshoot was 10.24 points beyond the stop, and the average was 1.33 points. At a 5% stop the same thing happens on all 50 trigger days, with an average overshoot of 1.32 points. Gold does not drift to a stop; it arrives there already through it.
How does gold behave in its worst days?
Gold's daily returns are not normally distributed. The worst single day on record was -13.24% on 1980-01-22, and the worst five-day stretch was -26.59% from 1980-01-22 to 1980-01-28. Excess kurtosis over the full sample is 10.74, so the tail is much fatter than a normal distribution would imply. The practical consequence is that moves beyond three standard deviations happen 1.81x as often as the normal benchmark predicts, even though moves beyond two standard deviations are less common than normal would suggest.
How long does it take to recover from a big drop?
The median is short and the tail is very long. After a -5% daily drop, gold regained its previous level a median of 28.5 trading days later across 50 such events, with 48 of them recovering at all. But the mean is 614 days, the 90th percentile is 1136 days, and the longest was 6985 days. The median and the mean are not describing the same experience.
What is gold value-at-risk and expected shortfall on this data?
At the 95% level, the historical Value at Risk on gold is 1.75% and Expected Shortfall is 2.82% — meaning that in the worst 5% of days gold lost more than 2.82% on average. At the 99% level, VaR is 3.36% and Expected Shortfall is 4.79%, against 3.23% for a normal distribution with the same volatility. Both are historical figures: they describe what happened, and they carry no promise about the next crisis.
Is there a day of the week that is better for gold?
No, not one that survives a proper test. Wednesday and Friday show the highest average returns (0.0614% and 0.0595%) and both look individually significant against zero — t-statistics of 2.77 and 2.78. But the correct test is whether the weekday means differ from each other, and that one-way ANOVA gives F = 1.0324 with p = 0.7809 on 4 and 14671 degrees of freedom. That is no effect at all. Comparing each day against zero is the wrong test, and it is how the Friday-gold myth survives.
What does this say about how I should size a gold position?
It argues for sizing on the tail rather than on the average. Gold spent 97.3% of its history below a prior peak, with the longest stretch underwater lasting 7020 trading days. A position sized so that a 2.82% day is survivable is a position that can be held through the 4.79% days that actually happen. The stop-loss results point the same way: the 5% stop with a five-day cooldown still produced a 7.92% annualised return against buy-and-hold's 7.7%, but the close-filled version of the same rule fell to 6.83% with a worse drawdown of 72.7%. What protects capital here is position size and re-entry discipline, not the stop order itself.
Where is gold tail risk right now?
On the last 63 fixing days, gold's realised volatility annualises to 21.07%, with a 95% Value at Risk of 2.64% and Expected Shortfall of 3.21%. The 99% Expected Shortfall over that window is 4.59%, which sits at the 74.8th percentile of its own history since 1968-04-02. That is elevated but not extreme — the tail risk is running above its long-run norm without reaching the levels seen in 1980 or 2008.
Sources
-
Gold: LBMA Gold Price PM benchmark fix, USD per troy ounce —
scripts/daily_gold_lbma_pm.csv(1968-04-02 to 2026-11-10). -
Analysis script:
scripts/gold_tail_risk_analysis.py. Independent recomputation with numpy:scripts/verify_gold_tail_risk.py(385 checks over the published figures). - For volatility-based position sizing on a longer monthly series, see volatility-targeting position sizing; for the option-market side of the same question, see gold implied versus realised volatility; for drawdown episodes measured inside the month, see gold intra-month drawdown; for the entry rule that follows a crash, see does buying gold dips actually work.
Value at Risk is the loss threshold exceeded on the worst 5% or 1% of observations; Expected Shortfall is the average loss conditional on exceeding that threshold. Both are historical and neither is a prediction.