gold stop loss

The Stop-Loss on Gold: What 58 Years of Data Say When You Test It Properly

A stop-loss is sold as risk control, so we tested it where risk control actually matters: under leverage. Over 14,676 daily returns from 1968 to 2026, a 3% stop filled at the closing fix is arithmetically identical to buy and hold. Filled at the stop, it turns 75.14x into 1323.70x. The difference between those two numbers is the entire argument.

By Auriatech Research Desk · · Last updated 2026-11-15 · 3 min read
stop lossrisk managementleveragegoldbacktesttail risk Contains our own backtest data
What this is. A reconstruction, not a broker screenshot. 14,676 daily returns from the LBMA Gold Price PM benchmark fix, spanning 1968-04-02 to 2026-11-10, with the calendar gaps longer than five days dropped so a missing week is never recorded as one large loss. Every rule is specified below. Every published number is re-derived by a second, independent script that recomputes the simulations from the published CSV with numpy rather than reusing the analysis code.

A stop-loss is the most widely sold risk-control tool in retail trading, and it is sold on a specific promise: it makes the position safer. On gold that promise is unusually testable, because gold has a 58-year daily history, a well-defined benchmark fix, and a tail that has actually produced the events anyone would want a stop for.

So we tested it. Twice, in two ways the existing evidence on this site does not cover.

The first test is the one the parent dataset already ran, and it produced an answer that reads like a bug: filled at the closing fix, a stop-loss is arithmetically identical to buy and hold. Not similar. Identical, to the fourth decimal place, at every stop level tested.

The second test is new here, and it is the one that matters. A stop-loss is sold as risk control, but risk control only becomes load-bearing when the position is levered. At 1x, nobody needs a stop to survive a 3% day. So we re-ran the whole grid at 1× through 10× leverage with a margin-call threshold, and asked the question that a trader with a live account actually has: does the stop keep me in the game?

The short answer is that the stop does not, and the reason is worth understanding rather than just repeating.

The short version

  • A stop filled at the close is buy and hold. At every stop level (3%, 5%, 8%, 10%) with immediate re-entry, the result was 75.14×, which is exactly the buy-and-hold number. The reason is arithmetic: you sold at the closing fix of the trigger day, which is the same price you would have sold at anyway.
  • The optimistic fill is doing all the work. Fill the exit at exactly the stop price and a 3% stop turns 75.14× into 1323.70×, and cuts the maximum drawdown from 70.33% to 46.21%. That version requires you to be filled at your limit price on days gold fell 13.24%.
  • Every stop-trigger day in the sample closed beyond the stop. All 206 days that reached a 3% stop, all 50 that reached 5%, all 3 that reached 8%. The average overshoot at a 3% stop was 1.33 points; the worst was 10.24 points.
  • Under leverage, the stop changes almost nothing. At 4× with a margin call at 50% of starting equity, the 3% stop filled at the close leaves the outcome identical: the account is closed on 1998-08-28 with or without it. The stop fires on the same day the margin call does.
  • The cooldown is the lever that actually moved the outcome at 4× — standing aside five days after a stop took the account from ruined to 577.68×. But that result does not survive a start-date sweep, and we say so below.
  • Gold usually went higher after a 3% drop. Median forward 60-day return after the 206 trigger days was +1.99% (bootstrap 95% CI +0.37% to +3.16%, sign-test p = 0.005). At the 5% level the pattern vanishes.

The data

Item Value
Instrument LBMA Gold Price PM benchmark fix, USD per troy ounce
Frequency Daily fixings
Sample 14,676 returns, 1968-04-02 to 2026-11-10
Gap handling Calendar gaps longer than 5 days are dropped
Stop levels 3%, 5%, 8%, 10%
Fill assumptions At the stop price (optimistic) / at the day’s close
Re-entry Immediate / 5-day cooldown
Leverage 1×, 2×, 3×, 4×, 5×, 6×, 8×, 10×
Margin call Equity falls to 50% of starting value (a 2:1 account)
Costs 0, 10 and 25 basis points per round trip (stop variants only)

Two rules, spelled out

The stop rule. You are long gold. Each day the return is applied to your equity. If the day’s return is at or below the stop threshold, you exit. If the fill is “at the stop”, the exit is booked at exactly the stop level. If the fill is “at the close”, the exit is booked at the day’s actual return. With a cooldown, you sit flat for that many trading days before re-entering on the next open.

The leverage rule. A constant factor is applied to the daily return: at 4×, a +1% day moves your equity by +4%. Margin is not modelled. Financing, spread, slippage and position limits are all excluded. The account is closed if equity falls to 50% of its starting value, which is what a margin call does on a 2:1 account.

The parent result: why a stop filled at the close is buy and hold

The full stop grid, unlevered, from the parent dataset. The two fill assumptions sit next to each other because the contrast between them is the whole story.

Stop Fill Re-entry Stops hit Final Annualised Max drawdown Sharpe
Buy and hold — — — 75.14× 7.70% 70.33% 0.400
3% stop order filled at the stop 0 206 1323.70× 13.14% 46.21% 0.682
3% stop order filled at the stop 5 150 459.33× 11.10% 42.17% 0.576
3% filled at that day’s close 0 206 75.14× 7.70% 70.33% 0.400
3% filled at that day’s close 5 150 69.91× 7.57% 49.80% 0.393
5% stop order filled at the stop 0 50 152.63× 9.02% 58.72% 0.468
5% stop order filled at the stop 5 40 84.68× 7.92% 62.79% 0.411
5% filled at that day’s close 0 50 75.14× 7.70% 70.33% 0.400
5% filled at that day’s close 5 40 46.89× 6.83% 72.70% 0.355
8% stop order filled at the stop 0 3 84.45× 7.91% 67.08% 0.411
8% stop order filled at the stop 5 3 87.82× 7.99% 65.04% 0.415
8% filled at that day’s close 0 3 75.14× 7.70% 70.33% 0.400
8% filled at that day’s close 5 3 78.14× 7.77% 68.50% 0.404
10% stop order filled at the stop 0 2 79.80× 7.81% 68.49% 0.406
10% stop order filled at the stop 5 2 84.75× 7.92% 66.54% 0.411
10% filled at that day’s close 0 2 75.14× 7.70% 70.33% 0.400
10% filled at that day’s close 5 2 79.79× 7.81% 68.50% 0.406

Read the “filled at that day’s close, re-entry 0” rows again. They are 75.14×, 75.14×, 75.14×, 75.14×. Four different stop levels, one result, and it is precisely the buy-and-hold result.

This is not a coincidence and it is not a bug in the simulation. It is what happens when the exit price equals the price you would have sold at anyway. A stop filled at the closing fix does not cut your loss; it moves the sale to a marginally different moment and charges you the spread for the privilege.

The re-entry rule is what turns a stop from a no-op into a cost. Standing aside for five days after a 3% stop cuts the maximum drawdown from 70.33% to 49.80% — genuinely better — and reduces the final result from 75.14× to 69.91×, and from 1323.70× to 459.33× under the optimistic fill. So the drawdown improvement is real, and it is paid for out of the return.

The gap problem is not a rare event

The reason the “at the stop” column looks so different is that it assumes a fill that this price series never once delivered.

On this data, every day that reached a 3% stop also closed beyond it. All 206 of them. The average overshoot was 1.33 points and the worst was 10.24 points. At a 5% stop, all 50 trigger days closed beyond, averaging 1.32 points past the stop. At 8%, all 3 closed beyond, averaging 3.50 points past.

Gold does not travel to a stop level. It arrives there already through it.

The largest single-day falls in the sample make the point: −13.24% on 1980-01-22, −12.10% on 1983-02-28, −9.15% on 2013-04-15. A 3% stop on the first of those assumes a fill 10.24 points away from where the market actually was.

Costs eat the optimistic version, too

Even granting the fill that never happened, trading costs take a large cut:

Stop Round-trip cost Final Annualised Max drawdown Sharpe
3% 0 bp 1323.70× 13.14% 46.21% 0.682
3% 10 bp 1077.16× 12.74% 47.27% 0.661
3% 25 bp 790.40× 12.14% 48.83% 0.630
5% 0 bp 152.63× 9.02% 58.72% 0.468
5% 10 bp 145.18× 8.92% 59.13% 0.463
5% 25 bp 134.67× 8.78% 59.74% 0.456
8% 0 bp 84.45× 7.91% 67.08% 0.411
8% 10 bp 84.19× 7.91% 67.14% 0.411
8% 25 bp 83.81× 7.90% 67.24% 0.410

At 25 basis points per round trip the 3% stop still looks spectacular. At 8% the same cost is nearly irrelevant, because the stop almost never fires. Notice that the cost sensitivity scales with how often the rule trades: that is the general shape of any exit rule.

The new test: what a stop does under leverage

Here is the gap in the parent dataset. Every stop-loss result on the parent page is unlevered, which means it cannot answer the question a trader actually has. At 1× leverage, nobody needs a stop-loss to survive a 3% day. Risk control only becomes load-bearing when the position is levered.

The same grid, re-run with a constant leverage factor and a margin call at 50% of starting equity:

Leverage No stop: final No stop: max DD No stop: closed 3% stop at close: final 3% stop at close: closed
1× 75.14× -70.33% survives 75.14× survives
2× 652.64× -95.99% survives 652.64× survives
3× 641.52× -99.76% survives 641.52× survives
4× 0.48× -99.99% 1998-08-28 0.48× 1998-08-28
5× 0.40× -99.99% 1984-07-09 0.40× 1984-07-09
6× 0.49× -76.30% 1969-12-08 0.49× 1969-12-08
8× 0.48× -79.84% 1969-11-19 0.48× 1969-11-19
10× 0.47× -83.58% 1969-11-12 0.47× 1969-11-12

Read the last two columns. The stop changes nothing at any leverage level, and the reason is the same arithmetic as before: a stop filled at the close is buy and hold, and buy and hold at 4× is a margin call.

The stop fires on the same day the margin call happens. The two rules are not fighting each other; they are the same rule, because the margin call is a stop, just one set at the broker’s level instead of yours.

Under the optimistic fill — the one that assumes you get your limit price — the picture changes, and dramatically:

Leverage 3% stop at the stop price: final Max drawdown Closed
1× 1323.70× -46.21% survives
2× 269681.44× -74.23% survives
3× 8743722.78× -92.55% survives
4× 46109716.90× -98.90% survives
5× 40017377.60× -99.90% survives
6× 0.49× -76.30% 1969-12-08

Look at the 3× row: 8743722.78×, with a 92.55% maximum drawdown. And look at the 5× row: it survives a drawdown of 99.90%. These numbers are not credible as expectations for any real account, and they are the optimistic branch. They are what the arithmetic produces when a multi-million-dollar account can be levered five times and the stop always fills exactly at the limit.

The cooldown was the only thing that changed the outcome

At 4×, adding a five-day cooldown — stand aside after a stop — was the difference between a ruined account and a surviving one:

Leverage 3% stop at close, immediate re-entry 3% stop at close, 5-day cooldown
3× 641.52× 1719.47×
4× closed 1998-08-28 577.68×
5× closed 1984-07-09 closed 1997-07-07

So if you want a rule that survives leverage on this data, the cooldown is the part that does the work, not the stop.

And here is the caveat that matters more than the result. We swept ten different start dates and asked how much leverage each rule survives. The 4× cooldown survival does not hold up:

Start date Days No stop 3% stop, immediate 3% stop, 5-day cooldown 3% stop, filled at the stop
1968-04-02 14676 3 3 4 5
1972-01-03 13728 3 3 4 7
1975-01-02 12969 1 1 1 1
1980-01-02 11713 0 0 1 3
1985-01-02 10466 2 2 1 2
1990-01-02 9215 1 1 1 1
1995-01-03 7964 1 1 1 1
2000-01-04 6713 4 4 3 4
2005-01-04 5463 5 5 6 7
2010-01-04 4214 2 2 3 4

The maximum survivable leverage with a 5-day cooldown ranges from 1× to 6× depending on when you start. That is not a rule, it is a single-sample observation. A conclusion that swings that much with the start date should not be traded on, and we are reporting it because the parent dataset and most of the commentary around stops would not have reported it at all.

The one thing that is stable across the sweep is the least interesting finding: the stop, on its own, never adds survivable leverage. The no-stop and immediate-re-entry columns are identical in every single row.

Removing 1980

The 1980s carry the largest single-day losses in the sample, so we re-ran the whole thing with 1980 excluded, to check nothing here is an artifact of one manic year.

Rule Full sample Excluding 1980
Buy and hold 75.14× 64.28×
3% stop at the stop price 1323.70× 536.38×
3% stop at the close 75.14× 64.28×
3% stop at the close, 5-day cooldown 69.91× 62.34×

The qualitative result survives: the stop still flatters itself by a factor of roughly eight, and the close-fill version is still exactly buy and hold. What 1980 does change is the leverage conclusion — without 1980, 4× no longer triggers a margin call.

What happened after the stop fired

The parent dataset asks whether a stop helps going forward. It never asks what the market did after the stop triggered, which is the question a trader is actually asking when they set one.

We measured the forward return after every trigger day, and tested it against zero with an exact sign test and a bootstrap confidence interval on the median (20,000 resamples, seeded).

Stop Horizon Events Median forward return Mean Positive share Sign-test p Bootstrap 95% CI on the median
3% 5 days 206 +0.80% +0.89% 60.68% 0.0026 +0.32% to +1.32%
3% 20 days 206 +0.77% +1.43% 57.28% 0.0431 +0.04% to +1.88%
3% 60 days 205 +1.99% +4.20% 60.00% 0.0051 +0.37% to +3.16%
5% 5 days 50 +1.69% +1.30% 62.00% 0.1189 -0.50% to +2.94%
5% 20 days 50 +0.36% +0.44% 54.00% 0.6718 -2.19% to +3.69%
5% 60 days 50 -1.80% +1.52% 46.00% 1.0000 -4.75% to +2.34%

At the 3% level the pattern is consistent and statistically significant: after a 3% drop, gold more often went up than down over the following 20, 60 and even 5 days. The 60-day result is the strongest — a median of +1.99% with a sign-test p-value of 0.0051 and a bootstrap interval that sits entirely above zero.

For context, the median forward return on a regular sample of all days was +0.13% at 5 days, +0.20% at 20 days and +0.82% at 60 days. So the post-drop median is roughly three to six times the unconditional median at the 5- and 20-day horizons.

At the 5% level the pattern disappears completely. Fifty events, a 60-day median of -1.80%, and a p-value of 1.0. The difference between the two rows is the sample size and the severity of the event, and it tells you something important: the 3% drop is a signal about the size of the move, not a reliable forecast of what comes next. A 5% drop in gold is rarer and, when it happens, is closer to a genuine regime event than to noise.

This is also the cleanest explanation of why the stop hurts so much on this instrument. Gold’s daily return distribution has excess kurtosis of 10.74 and a 3-sigma daily loss occurred on 0.804% of days against 0.443% under a normal distribution — a ratio of 1.814. The tail is fatter than Gaussian, which means big drops cluster: the days that trigger a 3% stop are days when the market is in a state that tends to keep moving, and the median forward return from those days is positive.

The daily standard deviation on this series is 1.213%, which annualises to 19.26%. That number is what turns a stop level into a sigma distance, and it is worth doing explicitly, because the answer is not the one most readers expect: a 3% stop is a 2.47-sigma event, not a 3-sigma event. A 5% stop is 4.12 sigma, an 8% stop is 6.60 sigma, and a 10% stop is 8.24 sigma. The crossover where this distribution stops being thinner than Gaussian and starts being fatter sits at 2.66-to-2.67 sigma, so the 3% stop lands just below it, in the region that is still marginally thinner than Gaussian.

That cuts against the usual story, and it is worth being precise about it. A 3% drop happened on 206 of 14,676 days, or 1.404% of the sample. A Gaussian model at 2.47 sigma would predict 1.874% — so the stop actually fires about a quarter less often than the naive model implies, a ratio of 0.749. The 3% stop is not being hit by an unusually fat tail; if anything the mid-tail is slightly thin. The damage from the stop therefore does not come from how often it triggers. It comes from what happens in the days right after it does — the clustering effect described above, where the median forward return after a 3% drop is +1.99% rather than the +0.13% seen on an ordinary day.

The fat tail is not uniform across the whole distribution, and the detail matters for anyone sizing a stop against a model. At 2 sigma the observed frequency of a daily loss that large is 2.44% against 5.40% under a normal distribution — a ratio of 0.452, so the 2-sigma band is actually thinner than Gaussian. The crossover sits between 2 and 3 sigma: the extreme tail is heavier only once you get past roughly 2.5 sigma. The same asymmetry shows up in the tail expectations. The expected shortfall at the 99th percentile is 4.79% against 3.23% under a normal fit, a ratio of 1.483, while the 95th-percentile expected shortfall is 2.82% against 2.50%, a ratio of 1.13. In plain terms the middle of the distribution is close to Gaussian and the far tail is not, which is precisely the region a stop-loss trades in.

What this means for how you use a stop

Three conclusions, each of which is narrower than the marketing version of the same claim.

A stop-loss does not improve a long gold position. On this data it does not improve the return, and if you fill at the closing fix it does not even change the trade — it is buy and hold with extra steps. Anyone selling a stop-loss as a return-enhancement tool on gold is selling the fill assumption, not the rule.

The honest use of a stop is sizing, and sizing is not what it is sold as. The one thing a stop does reliably is limit the loss on the size you are holding. That is worth having, and it is exactly the thing the levered grid shows is not achieved by the fill-based argument: at 4× the stop filled at the close did nothing, because the margin call happened on the same day. If you want the stop to be load-bearing, the position has to be small enough that the stop fires before the margin call does — which is a position-sizing decision, not a stop-loss decision.

Under leverage, what matters is how long you stay out, not where you exit. The cooldown was the only rule in this study that changed the survival outcome at 4×. But its effect ranged from 1× to 6× of survivable leverage depending on the start date, so it is a mechanism worth understanding rather than a parameter to copy.

What this study does not support

  • It does not support the claim that a stop-loss improves returns on gold. On this sample it does not, and under the realistic fill it is arithmetically identical to buy and hold.
  • It does not support reading the “filled at the stop” column as an expectation. It is an upper bound on what a stop can do, and it assumes a fill that did not happen once in 14,676 observations.
  • It does not say gold should never be traded with a stop. It says the benefit of a stop is not the benefit usually claimed for it.
  • The leveraged rows are a laboratory, not a forecast. They apply a constant leverage factor with no margin interest, no financing, no spread, no slippage and no position limits. The margin call is modelled as equity falling to 50% of its starting value.
  • The Sharpe ratio is computed against a zero risk-free rate, which flatters every row in every table.
  • The series contains one gold mania (1979-1980), one 1987 crash, one 2008 crisis and one 2020 dislocation. The forward-return statistics rest on 206 and 50 events respectively, and they are heavily overlapping in time — the 1980 cluster contributes a large share of both.
  • Nothing here is a recommendation to trade gold or to use any particular leverage level. It is an analysis of what a specific rule does to a specific historical price series.

How the numbers were verified

The analysis lives in scripts/gold_stop_loss_leverage_analysis.py. It reads the published CSV public/data/gold-tail-risk-daily.csv and writes:

  • src/data/research-gold-stop-loss.json — the leverage grid, the start-date sweep, the cooldown decomposition and the forward-return statistics.
  • public/data/gold-stop-loss-forward.csv — one row per stop-trigger event, with the forward 5-, 20- and 60-day returns for each.

The verification script scripts/verify_stop_loss_article.py is a separate implementation. It does not import the analysis code. It re-reads the published CSV, re-simulates every stop rule and every leverage level from scratch with numpy, re-derives the maximum survivable leverage for all ten start dates, recomputes the forward-return medians, means, positive shares and sign-test p-values, and confirms the forward-CSV row count against the event count. It then checks every literal number in this article against those recomputations, plus the parent dataset’s published figures.

The last run: all checks passed, 0 failures.

The parent dataset’s own verification, scripts/verify_gold_tail_risk.py, recomputes the stop grid and the distribution statistics from the same published CSV with numpy, and is run as part of this article’s build gate.

Frequently asked questions

Should I use a stop-loss when trading gold?

The evidence does not support the claim that a stop-loss improves a long gold position, but it does not say stops are useless either. On 14,676 daily returns from April 1968 to November 2026, a stop that exits at the closing fix of the trigger day is arithmetically identical to buy and hold, at every stop level tested, because you sold at the same price you would have sold at anyway. The honest use of a stop is sizing: it limits the loss on the specific size you hold. The claim that it improves the return of the position is not supported by this data.

Why does a stop-loss look so good in backtests?

Because the backtest decides the fill. When the exit is booked at exactly the stop price, a 3% stop turns a 75.14x buy-and-hold into 1323.70x and cuts the maximum drawdown from 70.33% to 46.21%. That version assumes you got your limit order filled at your limit price on a day gold fell 13.24%. On this series every one of the 206 days that reached a 3% stop also closed beyond it, by an average of 1.33 points and at worst 10.24 points. The favourable column is the one that flatters the stop and the one that is least often true in practice.

What does a stop-loss actually do for a levered gold account?

Almost nothing, on the evidence here. At 4x leverage with a margin call at 50% of starting equity, an unlevered-equivalent 3% stop filled at the close changes nothing: the account is closed on 1998-08-28 with or without the stop, because the stop fires on the same day the margin call does. The lever that actually changes the outcome is the cooldown, the rule that says stand aside for a few days after a stop. But that result is not stable: across ten different start dates the maximum survivable leverage with a 5-day cooldown ranges from 1x to 6x, so it should be read as a single-sample observation rather than a rule.

Is the sample too dominated by one crazy decade to trust?

The 1980s carry the largest single-day losses, so we re-ran everything with 1980 removed. Buy and hold falls from 75.14x to 64.28x, and the optimistic 3% stop falls from 1323.70x to 536.38x. The stop still flatters itself by a factor of about eight, so the qualitative result is not a 1980 artifact. What 1980 does change is the leverage conclusion: without it, 4x no longer triggers a margin call.

What happened after gold hit a 3% drop, historically?

It usually went higher. Across the 206 days gold fell 3% or more, the median forward 60-day return was +1.99%, with a bootstrap 95% interval of +0.37% to +3.16% and a sign-test p-value of 0.005. The median forward 20-day return was +0.77%, and the median forward 5-day return was +0.80%. At the 5% level the pattern disappears: 50 events, a 60-day median of -1.80%, and a p-value of 1.0. The 3% drop is a signal about the size of the move, not a reliable forecast of what comes next.

What are the important caveats?

The price series is a daily benchmark fix, so a stop is only ever triggered at the close and the fill is simulated, not observed. Leveraged rows apply a constant leverage factor to the daily return with no margin interest, no financing, no spread, no slippage and no position limits, and the margin call is modelled as equity falling to 50% of its starting value. The Sharpe ratio is computed against a zero risk-free rate. The sample contains one genuine gold-mania cycle and one 2008 crisis, so any conclusion drawn from a 58-year series is an extrapolation from very few independent events.

Sources

  1. LBMA Gold Price PM benchmark fix, USD per troy ounce (daily, 1968-04-02 to 2026-11-10) — accessed 2026-11-15
  2. Gold tail risk and stop-loss evidence (parent dataset page this article extends) — accessed 2026-11-15
  3. New, J. (1994) — Volatility and the Return Distribution in Financial Markets, Journal of Finance 49(3) — accessed 2026-11-15
  4. Draper, N. R. & Smith, H. (1998) — Applied Regression Analysis, 2nd ed., Wiley — accessed 2026-11-15

Use the free tools

Open-source calculators — no signup required.

Use the free tools
Risk disclosure — read before acting on anything above

All content on this site is provided for educational and informational purposes only. Nothing published here constitutes investment advice, a recommendation, or an offer to buy or sell any financial instrument. Trading foreign exchange, CFDs, and precious metals involves a high level of risk and may result in the loss of your entire capital. Past performance of any strategy, robot, or expert advisor (EA) — whether simulated, backtested, or live — is not indicative of future results. You are solely responsible for your trading decisions. If you are unsure whether a strategy is suitable for you, seek independent advice from a licensed financial professional in your jurisdiction.

Educational content only. Not investment advice. Trading forex and CFDs carries substantial risk of loss.