volatility targeting gold
Volatility Targeting on Gold: Does Sizing to Risk Actually Cut the Drawdown, or Just Hide the Tail?
We applied inverse-volatility sizing to 802 months of World Bank gold prices (1960-2026) and checked whether it really lowers drawdown or just hides tail risk. It cut max drawdown from -69.58% to -63.28% and volatility from 15.24% to 11.81%, but CAGR fell from 6.18% to 5.44% and the rule levered up into 45.4% of the falls it preceded.
There is a rule that feels like it should work. Estimate how volatile gold has been, divide a fixed risk budget by that volatility, and hold whatever size falls out. When gold is calm you lever up; when gold is violent you step back. On paper it is the obvious way to make a position less dangerous without throwing away the upside.
On a 66-year gold price series it does something less flattering than that. It genuinely lowers volatility. It genuinely lowers the maximum drawdown. And it also gives up roughly half the compound return, and it responds to past volatility in a way that makes it worse exactly when you would want it to be better. The rule is not a fraud and it is not a free lunch either. It is a risk-control tool that changes the shape of the return, and the shape it changes is mostly the wrong way for a long-only gold position.
This article works through what the rule actually did, month by month, over 802 months of real prices — and where the reduction in drawdown came from.
The short version
- Volatility targeting works as advertised and underdelivers on the promise. Max drawdown improved from -69.58% to -63.28% (rolling) and -54.92% (EWMA). Realised volatility fell from 15.24% to 11.81% and 10.42%.
- It is not free. CAGR fell from 6.18% (static 1x) to 5.44% (vol-target, 3x cap) and 5.08% (EWMA).
- It is mechanically procyclical. The rolling estimator cut exposure in 367 months and 48.8% of those cuts were followed by a rise. It levered up in 328 months and 45.4% of those were followed by a fall.
- The drawdown reduction is mostly de-risking, not avoidance. The rule never forecasts anything. It reduces your exposure to the loss you were already going to take, and it does so by being small, not by being right.
- The tail is not hidden, only shallower. The worst rolling 12-month price return was -36.39% (1980-07 to 1981-07) and no sizing rule touches a price statistic.
- It beats leverage on risk-adjusted terms and loses badly on return. Vol-target at a 3x cap had the best Sharpe in the ladder (0.461), while static 3x compounded at 12.12% with a -99.41% drawdown.
Every number here is reproduced by
scripts/gold_vol_target_analysis.py and independently re-derived by
scripts/verify_gold_vol_target.py, which recomputes the whole study in
NumPy along a deliberately different code path and checks 103 assertions
against the published CSVs.
The data
The price series is the World Bank’s Commodity Price Data (The Pink Sheet), monthly gold in nominal US dollars per troy ounce, column 70 of the Monthly Prices sheet, data updated 3 November 2026.
| Coverage | Value |
|---|---|
| Monthly observations | 802 |
| First month | 1960-01 |
| Last month | 2026-10 |
| Realised volatility, full sample | 15.24% annualised |
| Monthly volatility, full sample | 4.40% |
This is the same price basis used by the rest of this desk — the grid versus trend-following study and the multi-account capital study run on the identical series, so numbers can be compared across articles without a basis adjustment.
The two estimators
A volatility estimate has to come from somewhere, and the choice of “somewhere” turns out to matter more than most of the published discussion of this rule admits. Two are tested here:
| Estimator | Definition | Character |
|---|---|---|
roll12 |
Trailing 12-month sample standard deviation of monthly log returns, annualised by ×√12 | Simple, transparent, and slow to react |
ewma94 |
Exponentially weighted moving average with λ = 0.94, annualised | Faster to react, smoother in the tail |
Both are computed with the standard sample standard deviation (ddof = 1) and both are annualised with √12.
The sizing rule
f_t = min(cap, target / sigma_hat_t)
with target = 10% annualised and cap ∈ {1×, 2×, 3×}. The monthly return
is then f_t × r_t.
The critical implementation detail is where the volatility estimate comes
from. sigma_hat_t is computed from returns realised strictly before
the month being sized — sigma[i] excludes rets[i]. That means the
position for month t is set using information available at the end of
month t-1. There is no look-ahead anywhere in the study, and the
verification script asserts this property explicitly.
This is the part most retail write-ups of volatility targeting get wrong. If you compute the volatility of the same month you are trading and then size with it, you are trading on the move you already know happened. The result looks spectacular and is entirely fictional.
Three eras
Gold’s realised volatility is not stationary, so a single full-sample number is misleading. The study splits the history into three eras.
| Era | Months | Realised vol | Buy & hold | Max drawdown | Avg roll12 sigma |
|---|---|---|---|---|---|
| 1960-01 – 1979-12 | 240 | 15.37% | +1200.0% | -40.22% | 10.07% (66.1% of full sample) |
| 1980-01 – 1999-12 | 240 | 17.52% | -58.1% | -62.07% | 14.10% (92.5% of full sample) |
| 2000-01 – 2026-10 | 322 | 12.90% | +1345.4% | -39.28% | 12.12% (79.6% of full sample) |
The 1980s and 1990s were, on this measure, close to twice as volatile as the full-sample average suggests. Any rule that scales with volatility will be operating at very different settings in each era, and the first era is dominated by a stretch where the price was administratively fixed and the volatility estimate is mechanically degenerate. That last point deserves its own section, because it is where the honest version of this study starts.
The 84 months where the rule has no answer
From 1961-02 through 1968-01, the trailing 12-month window sits entirely
inside the era when the official gold price was fixed at $35 per troy ounce.
The standard deviation of monthly log returns across that window is exactly
zero, and target / 0 is not a number.
| Peg fact | Value |
|---|---|
| Months with zero-leverage sizing | 84 |
| First month | 1961-02 |
| Last month | 1968-01 |
| Price during the era | $35/oz (administered) |
This implementation sets exposure to 0 in those months rather than dividing by zero, and the data confirms the price was an official administered rate rather than a market price throughout. The London gold market was closed from 15 March 1968 to 1 April 1968, and the fixing changed from a sterling quote to a twice-daily dollar quote on 1 April 1968 as the Gold Pool collapsed.
The honest conclusion is that a live system could not have traded this series at all during those 84 months. It was not a tradeable price. Any backtest that includes it without flagging it is quietly pretending you could have been long gold at $35 on demand.
There is a second, more uncomfortable consequence, and it is worth staring at because it is the single most dangerous behaviour in the whole rule. The month after the peg breaks, the trailing volatility estimate is computed from a window that contains only one or two non-zero returns. The rolling estimator’s volatility for 1968-02 was 2.82%, which produces a leverage of 3.00× — the cap — and 1968-03 was also at the cap. In other words, the rule’s most aggressive possible position was set in the month immediately after the single largest structural break in the history of the price, because a volatility estimate that has barely seen any volatility yet concludes that there is no volatility.
That is the failure mode in miniature, and it is not an artefact of the 1968 episode. It is what inverse-volatility sizing does when the distribution changes rather than merely widens.
Result 1: the volatility reduction is real
The headline promise of volatility targeting is that it holds the position’s volatility near the target. It does that. Over the 789 simulated months from 1961-02 to 2026-10:
| Strategy | CAGR | Total return | Realised vol | Max drawdown | Sharpe | Avg leverage | Months at cap | Months at zero |
|---|---|---|---|---|---|---|---|---|
| Static 1x (buy & hold) | 6.18% | +5388.0% | 15.24% | -69.58% | 0.406 | 1.000 | — | — |
| Static 2x | 10.32% | +70033.1% | 30.47% | -94.36% | 0.339 | 2.000 | — | — |
| Static 3x | 12.12% | +206845.0% | 45.71% | -99.41% | 0.265 | 3.000 | — | — |
| Vol-target 10%, cap 1x | 5.10% | +2540.3% | 10.57% | -52.58% | 0.483 | 0.714 | 31.1% | 10.6% |
| Vol-target 10%, cap 2x | 5.29% | +2872.8% | 11.69% | -64.28% | 0.453 | 0.810 | 2.2% | 10.6% |
| Vol-target 10%, cap 3x | 5.44% | +3157.2% | 11.81% | -63.28% | 0.461 | 0.819 | 0.3% | 10.6% |
| Vol-target 10% EWMA, full sample | 5.08% | +2496.0% | 10.42% | -54.92% | 0.487 | 0.723 | 0.4% | 10.6% |
Read the volatility column first. The target was 10%, and the realised volatilities were 10.57%, 11.69% and 11.81% for the three caps and 10.42% for the EWMA. That is a 4% to 18% overshoot on a 10% target, which is close to what the academic literature reports for slow volatility estimators — a target is not a guarantee, and the slower the estimator, the further the realised volatility drifts from it.
Now the drawdown column. The full-sample buy-and-hold maximum drawdown was -69.58%, peaking in 1980-01 and bottoming in 2001-04, with recovery taking 82 months. Volatility targeting cut that to -63.28% (rolling) and -54.92% (EWMA). With a 1x cap it was -52.58%, the shallowest number in the entire ladder.
So: yes. The rule reduces volatility and it reduces the maximum drawdown. That part is not marketing, it is arithmetic.
And here is the part that is arithmetic too, and that the marketing does not mention: the CAGR fell from 6.18% to 5.44%. Over 65.75 years of compounding, giving up 0.74 percentage points of annual return is not a rounding error. Static 2x compounded at 10.32% and static 3x at 12.12%. The vol-target rule with the same 3x ceiling returned less than half of static 2x.
The rule bought a shallower hole with less of the journey.
Result 2: where the drawdown reduction actually came from
Here is the question the article exists to answer, so let me answer it directly instead of describing the mechanism in the abstract.
The reduction is entirely a de-risking effect. The rule never avoided a drawdown; it just made sure it was small when it arrived.
Consider the post-1980 peak collapse, January 1980 through December 1981:
| Strategy | Total return | CAGR | Max drawdown | Worst 12m | Avg leverage | Min lev | Max lev |
|---|---|---|---|---|---|---|---|
| Static 1x | -43.54% | -25.79% | -43.54% | -38.47% | — | — | — |
Vol-target roll12 |
-15.41% | -8.36% | -15.48% | -13.63% | 0.304 | 0.191 | 0.490 |
Vol-target ewma94 |
-13.57% | -7.33% | -13.57% | -12.39% | 0.280 | 0.236 | 0.386 |
The rolling estimator’s leverage path through that collapse:
| Month | Gold price | roll12 sigma |
Leverage (roll12) |
Equity (vol-target) | Equity (static 1x) |
|---|---|---|---|---|---|
| 1979-12 | $455 | 18.88% | 0.530 | 4.9596 | 10.2616 |
| 1980-01 | $675 | 20.42% | 0.490 | 5.9175 | 14.3089 |
| 1980-02 | $665 | 38.83% | 0.258 | 5.8947 | 14.0953 |
| 1980-03 | $554 | 40.26% | 0.248 | 5.6273 | 11.5212 |
| 1980-06 | $601 | 49.35% | 0.203 | 5.7143 | 12.3297 |
| 1980-12 | $538 | 50.78% | 0.197 | 5.5857 | 10.8260 |
| 1981-01 | $557 | 52.48% | 0.191 | 5.6227 | 11.2018 |
| 1981-06 | $461 | 28.50% | 0.351 | 5.2833 | 9.2001 |
| 1982-01 | $384 | 19.21% | 0.521 | 4.8348 | 7.5495 |
By June 1980 the rule was running at 0.203× leverage, and by December 1980 at 0.197×. It was never out of the market, but it was carrying roughly one fifth of the exposure. The result: the account fell 15.41% where buy-and-hold fell 43.54%.
Nothing about the rule predicted the 1980 peak. It reacted to volatility that had already happened. The drawdown was smaller because by the time the loss was happening the rule had already shrunk the position — which is exactly what it was designed to do, and exactly why it did not protect the return.
Compare with the two-decade bear market, 1980-01 through 1999-12:
| Strategy | Total return | CAGR | Max drawdown | Worst 12m | Avg leverage | Min lev | Max lev |
|---|---|---|---|---|---|---|---|
| Static 1x | -66.72% | -5.37% | -69.45% | -38.47% | — | — | — |
Vol-target roll12 |
-55.88% | -4.03% | -63.28% | -34.09% | 1.015 | 0.191 | 2.859 |
Vol-target ewma94 |
-47.00% | -3.14% | -54.92% | -31.42% | 0.839 | 0.236 | 1.789 |
Note the average leverage in the 1980s–1990s window: 1.015× for the rolling estimator. The rule was, on average, running slightly above 1× through the worst two decades in the history of gold — and it still lost 55.88%. In 1996-01 it was at 2.859×, the highest leverage of the entire era, because the trailing volatility had fallen to 3.50%. Gold was in a quiet consolidation, the rule concluded that gold was safe, and it bought the maximum the cap allowed.
That is the shape of the whole thing. The rule is not wrong. It is doing exactly what it says. But it is doing it with a one-to-twelve-month memory of how violent the market has recently been, and gold’s volatility is strongly mean-reverting on that horizon.
Result 3: the procyclicality bill
A rule that responds to realised volatility is going to cut size after volatility rises and add size after it falls. The question is what that costs. The study measures it directly: for every month where the leverage changed, it records the return in the month the new leverage applied to.
| Estimator | Months deleveraged | Mean return after deleveraging | Deleveraged into a rise | Months levered up | Mean return after levering up | Levered into a fall |
|---|---|---|---|---|---|---|
roll12 |
367 | +0.777% | 48.8% | 328 | +0.546% | 45.4% |
ewma94 |
218 | +1.339% | 53.2% | 484 | +0.364% | 47.1% |
Three things fall out of that table.
First, the rule cuts size roughly half the time it should not. With the rolling estimator, in 48.8% of the 367 months where leverage was cut, the month that followed was up. With the EWMA estimator it was 53.2% of 218 months. In other words the faster estimator misses more often, and it misses by a wider margin.
Second, the rule adds size ahead of falls at close to a coin-flip rate.
roll12 levered up in 328 months and 45.4% of them were followed by
a fall. ewma94 levered up in 484 months with 47.1% followed by a
fall. Notice that ewma94 levered up nearly 50% more often than roll12
while producing a lower average return after doing so (+0.364% against
+0.546%). Reacting faster made the procyclicality worse, not better.
Third, the deleveraging months had a higher average return than the
levering-up months. +0.777% against +0.546% for roll12. A rule that
sounds intuitive — step back when it’s risky, step in when it’s calm —
produces a book that systematically underweights the months following
calm and overweights the months following turbulence. That is the opposite
of what an intuitive reading of the rule suggests.
This is the mechanical explanation for the 0.74 percentage point CAGR give-up, and it is worth being precise about the logic: the rule is not “wrong” in the sense of being buggy. It is trading a real return stream for a smoother equity curve, and it pays for that smoothing with a systematic tilt that costs money in this particular sample.
Result 4: the full stress-window table
Seven windows, three strategies each. This is the table that answers “does it actually protect me when it matters”.
| Window | Description | Months | Static 1x total | roll12 total |
ewma94 total |
Static 1x MDD | roll12 MDD |
ewma94 MDD |
|---|---|---|---|---|---|---|---|---|
| 1980-01 – 1981-12 | Post-1980 peak collapse | 24 | -43.54% | -15.41% | -13.57% | -43.54% | -15.48% | -13.57% |
| 1980-01 – 1999-12 | Two-decade bear market | 240 | -66.72% | -55.88% | -47.00% | -69.45% | -63.28% | -54.92% |
| 2000-01 – 2011-09 | Dot-com era with gold bottoming | 141 | +456.24% | +446.73% | +295.22% | -22.40% | -13.76% | -13.14% |
| 2011-09 – 2015-12 | Post-peak gold bear market | 52 | -41.12% | -39.19% | -33.74% | -41.12% | -39.19% | -33.74% |
| 2018-01 – 2020-12 | COVID crash and recovery | 36 | +37.52% | +42.98% | +39.81% | -10.40% | -12.05% | -10.46% |
| 2020-01 – 2022-10 | 2020 peak to 2022 low | 34 | +4.87% | +4.75% | +5.51% | -16.37% | -16.97% | -14.86% |
| 2024-01 – 2026-10 | Recent volatility expansion | 34 | +94.00% | +107.15% | +79.55% | -19.32% | -12.19% | -11.87% |
The pattern is consistent and it is the whole argument of this article.
In every window that hurt, the rule hurt less. In every window that
helped, it helped less — with one exception. The 2024–2026 window is the
exception: roll12 returned +107.15% against static 1x at +94.00%,
because gold’s realised volatility fell for long enough that the rule
re-levered into the move.
But notice what happened to the drawdown in that same window: static 1x
took -19.32%, roll12 took -12.19%, ewma94 took -11.87%. And
notice the 2018–2020 window, where roll12 had a worse drawdown than
static 1x (-12.05% against -10.40%) while returning more. The
2020-01 to 2022-10 window is the same story: roll12 at -16.97% versus
static at -16.37%.
So the improvement is not uniform. The rule reliably helps in the windows where volatility is genuinely elevated and sustained. It is not reliable in the windows where volatility spikes and then mean-reverts quickly, which is precisely when the estimator is slowest to react and most likely to be carrying the wrong size.
Result 5: what the tail actually is
The question in the title deserves a direct answer, so here it is with the numbers attached.
The worst rolling 12-month price return in the entire 802-month sample was -36.39%, between 1980-07 and 1981-07. That is a property of the price series. No position sizing rule of any kind changes it. What sizing changes is the multiplier you apply to it.
Against that, the equity worst 12-month figures from the stress table are what the rule actually delivers:
| Window | Static 1x worst 12m | roll12 worst 12m |
ewma94 worst 12m |
|---|---|---|---|
| 1980-01 – 1981-12 | -38.47% | -13.63% | -12.39% |
| 1980-01 – 1999-12 | -38.47% | -34.09% | -31.42% |
| 2000-01 – 2011-09 | -12.98% | -7.77% | -8.97% |
| 2011-09 – 2015-12 | -28.18% | -26.47% | -21.90% |
| 2018-01 – 2020-12 | -3.94% | -4.16% | -3.53% |
| 2020-01 – 2022-10 | -9.72% | -9.14% | -8.60% |
| 2024-01 – 2026-10 | -0.60% | +4.01% | +0.86% |
The tail is not hidden. It is scaled. And scaling it down costs you the upside, because the same multiplier works on both sides of the ledger.
This is the honest version of “does volatility targeting reduce tail risk”: it reduces the severity of the tail you experience without reducing the probability of the tail in the underlying asset, and it pays for that severity reduction in the returns you forego. That is a defensible trade for an investor who cannot stomach a -40% year. It is a bad trade for an investor who was going to hold gold through one anyway, and who can actually tolerate the drawdown because the position is small enough that a -40% year does not threaten the account.
Result 6: where we are right now
The most recent volatility regime is the one most likely to be relevant to a live decision, so it gets its own numbers rather than being buried in an appendix.
Over the last 34 months:
| Statistic | Value |
|---|---|
Average realised volatility (roll12) |
12.06% |
| Minimum | 7.33% |
| Maximum | 21.91% |
At a 10% target and an average estimate of 12.06%, the implied leverage is
0.83×. At the maximum observed estimate of 21.91%, it would be
0.46×. Both are below the 3× cap, which is why the cap binds in only
0.3% of months for roll12 — the target is what binds, not the cap.
The monthly detail of the recent volatility expansion:
| Month | roll12 sigma |
Leverage (roll12) |
ewma94 sigma |
Leverage (ewma94) |
|---|---|---|---|---|
| 2025-01 | 10.47% | 0.96 | 11.24% | 0.89 |
| 2025-03 | 10.53% | 0.95 | 12.11% | 0.83 |
| 2025-05 | 9.76% | 1.02 | 13.31% | 0.75 |
| 2025-07 | 8.75% | 1.14 | 12.77% | 0.78 |
| 2026-01 | 11.98% | 0.83 | 15.55% | 0.64 |
| 2026-04 | 15.15% | 0.66 | 17.04% | 0.59 |
| 2026-06 | 17.18% | 0.58 | 16.37% | 0.61 |
| 2026-09 | 21.91% | 0.46 | 17.89% | 0.56 |
And the equity path through the same stretch:
| Month | Gold price | Equity (vol-target roll12) |
Equity (static 1x) |
|---|---|---|---|
| 2024-12 | $2648 | 20.9085 | 36.5455 |
| 2025-07 | $3340 | 26.1550 | 45.8199 |
| 2026-01 | $4753 | 35.4918 | 64.2875 |
| 2026-06 | $4228 | 33.0741 | 56.8247 |
| 2026-09 | $4319 | 33.3740 | 57.8184 |
| 2026-10 | $4105 | 32.5721 | 54.8802 |
The gap between the two equity curves widened through the 2025–2026 rally and then held through the correction, because the rule was running below 1× for most of it. Over the 2024-01 to 2026-10 window the rule returned +107.15% against +94.00% for buy-and-hold, but it also produced a -12.19% drawdown against -19.32%.
Right now, at 21.91% trailing volatility, the rolling rule would size a position at 0.46×. Anyone reading this as “the rule told me gold is risky” is misreading it: the rule is telling you that gold has been risky over the last twelve months, and it is making that statement by reducing your size. Whether you want that is a question about your drawdown tolerance, not about the forecast.
What this study is not
Every backtest on a single price series is a study of that series, and the sins of this one are worth listing.
It is price-only. No financing cost, no storage cost, no futures roll, no dividend, no spread, no commission. Leveraged gold is precisely the trade where carry bites hardest, so every leveraged result above is optimistic, and the error grows with leverage. Static 3x is the most flattered number in the ladder.
The Sharpe ratio uses a zero risk-free rate. Every strategy in the table is measured against 0%, which flatters the leveraged runs and the long-sample runs alike. It does not change the ranking much, but the absolute Sharpe numbers should not be read as investable.
One asset is not a portfolio. Volatility targeting is a risk-budgeting rule and the interesting case for it is a multi-asset book where you can scale down one sleeve and up another. On a single asset, “scaling down” can only mean “less gold”, and the opportunity cost of that gold has nowhere to go. This is the same structural problem that runs through the risk parity study on this desk, where equal-risk-contribution weighting across seven assets also failed to beat the simplest alternative. Volatility targeting on a single asset is risk parity on a single asset, and it inherits the same limitation.
The estimator choice was made in advance and not varied. Two estimators were tested. A proper study would sweep the lookback window, the target volatility, the cap, the rebalance frequency and the estimator family, and it would report the distribution of outcomes rather than two points on it. We report two points on it honestly rather than implying the pair is canonical.
The 84 peg months are a policy artefact, not a market observation. They are flagged and set to zero exposure. A different but equally defensible choice — dropping the era entirely, or treating sigma = 0 as an insufficient-data signal — would change the first era’s numbers.
Monthly rebalancing friction is not modelled. The rule implies rebalancing monthly, which in a real account means commission and spread on a partial position adjustment. For a retail account trading one instrument on a 100-contract-per-lot contract size, this is not zero.
If you are going to use it anyway
Given all of the above, the honest guidance is not “don’t”. It is “use it for the thing it is good at, and stop expecting it to do the thing it is not good at.”
Use it if you cannot survive a -40% year. That is the rule’s genuine domain. If a -40% drawdown would force you to close the position, then a rule that halves the drawdown is worth a lot of money even at the cost of 0.74 percentage points of CAGR. The position sizing framework on this desk makes the same argument from the other direction: the sizing rule that matters is the one that keeps the account alive long enough for the thesis to work.
Do not use it if your position is already small enough that the drawdown does not threaten you. In that case the rule is pure drag on the return. A position sized so that a -40% year costs you 8% of the account needs no volatility overlay; it needs no overlay at all, and the overlay’s procyclical tilt is a pure cost.
Do not confuse it with a forecast. The rule is a function of past volatility. It has no view on direction. When it cut to 0.191× in January 1981 it was not predicting the bottom; it was reacting to a fall that had already happened. It is a risk-budgeting rule dressed up in the language of timing.
Set the target against your actual drawdown tolerance, not against a number you read somewhere. A 10% target with a 3× cap produced an average leverage of 0.819× over this sample, which means the rule was, on average, reducing exposure relative to unlevered gold. That is a materially different position from what “10% volatility target” sounds like to most readers. And the minimum non-zero volatility estimate in the whole series was 2.82%, which produced a 3.00× position — the cap — for reasons that had nothing to do with risk management.
Understand the leverage it will actually carry. With roll12, leverage
exceeded 2× in only 17 of 789 months, and reached the 3× cap in only
2 months (1968-02 and 1968-03, both artefacts of the peg break). The
rule is a de-risking rule on this asset, not a leverage rule. If you want
leverage, take it deliberately and size it for the -99% scenario, as
the multi-account capital study
does with separated sleeves, rather than picking it up accidentally through
a volatility estimator.
Do not put a grid ladder under it. A grid ladder assumes a fixed notional size that grows as price moves against you — the opposite of what inverse-volatility sizing does. Combining them means the two rules fight over the position size every month, and the loser is whichever one you did not intend. The grid mechanics article sets out how a ladder’s depth decides whether it survives a trend; adding a volatility overlay on top changes the ladder’s effective depth without changing its arithmetic, and that is how accounts get funded to zero.
How the numbers were produced
The full pipeline is reproducible from the repository:
python scripts/gold_vol_target_analysis.py # writes the JSON and both CSVs
python scripts/verify_gold_vol_target.py # independent NumPy re-derivation
node scripts/audit-dist.mjs # built-site SEO / JSON-LD audit
The analysis script reads the World Bank workbook directly, computes the two volatility estimators, runs the sizing simulation and the static-leverage benchmarks, decomposes the result into eras and stress windows, and writes:
src/data/research-vol-target.json— the full research payload behind the dataset page.public/data/gold-vol-target-monthly.csv— 789 monthly rows: price, log return, both volatility estimates, both leverage paths, and all three equity curves, every numeric cell at 12 decimal places.public/data/gold-vol-target-stress.csv— 7 stress windows × 3 strategies, same 12-decimal precision.
The verification script is a separate implementation. It re-reads the source workbook, re-computes the estimators and the simulation with NumPy vectors rather than Python loops, and then checks the published CSVs cell-by-cell against that recomputation — including a check that every numeric cell really is in 12-decimal fixed-point format, so the published files cannot silently drift to a different precision than the page quotes. It also re-derives every stress-window statistic from the published monthly CSV, so the stress table is provably a function of the equity curve that was published, not a separate calculation.
The last run: 103 checks, 0 failures.
Related reading
- Why Your Grid Backtest Beats Your Live Account: Data Frequency, Not Strategy - the price-path granularity effect, measured on four grid configurations
- Gold Intra-Month Drawdown, 1968-2026 - what a monthly close hides about the drawdown that happened inside the month
- Risk Parity on Gold: Where the Leverage Argument Breaks — the multi-asset version of this question, and the finding that a plain 60/40 book beat the whole risk-parity construction
- Gold Position Sizing: How Much to Risk per Trade — the sizing framework this rule is an alternative to
- Grid vs Trend-Following on Gold: 66 Years of Evidence — what happened to a grid ladder and a 12-month trend rule on the same 802 monthly closes
- Gold Grid Trading Mechanics — how ladder depth decides which grid survives a trend
- Multi-Account Gold Trading — six sleeves over 66 years, and why accounts holding the same asset lose money together
- Reading an Equity Curve: Drawdown, Recovery, and What They Hide — how to read the curves this article produces, including the parts they smooth over
Frequently asked questions
Does volatility targeting actually reduce drawdown on gold?
Yes, and the reduction is real but modest. Over the 789 simulated months the rolling 12-month estimator produced a -63.28% maximum drawdown against -69.58% for static 1x, and the EWMA estimator -54.92%. Realised volatility fell from 15.24% to 11.81% and 10.42% respectively. The rule delivered on its stated promise — it just delivered less than most people expect, and it charged half the compound return for it.
Why did the CAGR drop if the risk was reduced?
Because the rule is mechanically procyclical. With the rolling estimator the position was cut in 367 months, and 48.8% of the time that cut was followed by a rise. It levered up in 328 months, and 45.4% of those were followed by a fall. Static 1x compounded at 6.18% a year against 5.44% for vol-target at a 3x cap. You gave up 0.74 percentage points of annual return to remove 6.30 percentage points of drawdown.
Does it make the tail risk better or just smaller?
Smaller, not better. The worst rolling 12-month price return in the sample was -36.39% between 1980-07 and 1981-07, and that is a price statistic no sizing rule touches. What the rule changes is how much of that move you are exposed to: in 1980-06 the rolling estimator had cut leverage to 0.203x, so the account lost 15.41% over the 1980-1981 collapse against -43.54% for static 1x. The hole is shallower because you were barely in the market, not because the rule anticipated anything.
What did the rule actually do during the 1980-1999 bear market?
It lost less money, and it still lost money. Over those 240 months the rolling estimator averaged 1.015x leverage, ranged from 0.191x to 2.859x, and returned -55.88% with a -63.28% drawdown that never recovered inside the window. Static 1x returned -66.72% with a -69.45% drawdown. The EWMA version returned -47.00% with a -54.92% drawdown. The rule lost roughly a quarter of the damage and took the full two decades to do it.
Is volatility targeting better than just using leverage?
No. Static 2x compounded at 10.32% a year with a -94.36% maximum drawdown; static 3x compounded at 12.12% with a -99.41% drawdown. Vol-target with a 3x cap compounded at 5.44% with a -63.28% drawdown. On raw return, leverage wins decisively. On the risk-adjusted measure, the Sharpe ratios run the other way: 0.406 for static 1x, 0.339 for 2x, 0.265 for 3x, against 0.461 for vol-target at a 3x cap. Volatility targeting bought the best Sharpe in the ladder without touching the 2x and 3x ruin scenarios.
What are the important caveats?
The series is price-only: no financing, storage, roll or dividend cost is included, so every leveraged result is optimistic relative to a real position, and carry bites hardest exactly where the leverage is. 84 months from 1961-02 to 1968-01 sit entirely inside the administratively fixed $35/oz price era, where the trailing standard deviation is exactly zero and inverse-volatility sizing has no defined answer; this implementation sets exposure to 0 rather than dividing by zero. The Sharpe ratio is computed against a zero risk-free rate, which flatters every leveraged run. And one asset over one price path is not a portfolio.
Sources
- World Bank — Commodity Price Data (The Pink Sheet), monthly, updated November 3, 2026 — accessed 2026-11-11
- World Bank — CMO Historical Data Monthly (direct download used for this analysis) — accessed 2026-11-11
- LBMA — March 1968 and the London Gold Fixing — accessed 2026-11-11
- World Gold Council — Historic Gold Agreements — accessed 2026-11-11
- IMF — The International Monetary Fund 1966-1971, chapter 20 — accessed 2026-11-11
- Moreira, A. & Muir, T. (2017) — Volatility-Managed Portfolios, Journal of Finance 72(4), 1611-1644 — accessed 2026-11-11
Use the free tools
Open-source calculators — no signup required.
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.