Research · Sizing study

Volatility-targeting position sizing on gold, 1960–2026

The most common claim in position-sizing literature is that scaling exposure inversely to volatility produces a better risk-adjusted result. This page tests that claim on one asset — monthly gold prices — with a rule fixed in advance, no look-ahead, and every number recomputed by a published script.

Coverage: 802 months, 1960-01 to 2026-10 · Source data updated November 03, 2026 · Generated by scripts/gold_vol_target_analysis.py

The rule under test

Sizing rule
ft = min(cap, target / σ̂t)
Target volatility
10% annualised
Leverage cap
3×
Volatility estimators
roll12 (trailing 12-month stdev) · ewma94 (lambda 0.94)
Look-ahead
none: sigma[i] excludes rets[i]
Realised volatility in sample
15.24% annualised (4.40% per month)

f_t = min(cap, target / sigma_hat_t). sigma_hat_t is estimated from returns realised strictly BEFORE the month that f_t sizes, so the position for month t is set using information available at the end of month t-1. No look-ahead.

Two estimators are run so the conclusion does not depend on one particular volatility model. roll12 is the standard deviation of the trailing 12 monthly returns; ewma94 is an exponentially weighted estimate with λ = 0.94, which reacts faster to a change in regime. If a result only holds under one of them, it is a result about the estimator, not about gold.

Leverage ladder: static versus volatility-targeted

Static leverage multiplies every month by a fixed number. Volatility targeting sets exposure each month from the volatility observed before that month. Both are applied to the same monthly gold returns over the same window.

Position rule CAGR Ann. vol Max drawdown Sharpe Total Avg leverage Months at cap
static 1x +6.18% 15.24% -69.58% 0.406 +5,388.0% — 100%
static 2x +10.32% 30.47% -94.36% 0.339 +70,033.1% — 100%
static 3x +12.12% 45.71% -99.41% 0.265 +206,845.0% — 100%
vol-target 10% / cap 1x +5.10% 10.57% -52.58% 0.483 +2,540.3% 0.714× 31.1%
vol-target 10% / cap 2x +5.29% 11.69% -64.28% 0.453 +2,872.8% 0.810× 2.2%
vol-target 10% / cap 3x +5.44% 11.81% -63.28% 0.461 +3,157.2% 0.819× 0.3%

Read the drawdown column before the return column. Static 3× produced a -99.41% peak-to-trough loss — on a price series alone, with no financing cost, no spread and no margin interest. A static 2× position produced -94.36%. A volatility-targeted position capped at 2× produced -64.28% while giving up most of the extra return: the targeted version compounds at 5.29% against 10.32% for the static version.

The mechanism is visible in the leverage columns. Static leverage is fixed by definition, so it stays fully invested into every calm month and every panic. Volatility targeting spends most months below 1×: the 1×-cap rule averaged 0.714× exposure and sat at its cap in only 31.1% of months, while the 3×-cap rule averaged 0.819× and reached 3× in 0.3% of months — the leverage cap almost never binds, because the target volatility of 10% already keeps the position below 1× most of the time. It buys capital preservation by accepting a lower average exposure, and it pays for that in CAGR, not in Sharpe. The best Sharpe in the table, 0.483, belongs to the volatility-targeted rule at a 1× cap — a 10.57% annualised volatility with a -52.6% drawdown.

Does it hold across regimes?

A sizing rule that only works in one era is not a rule. The series is split into three eras — the administered-price era, the 1980 boom-bust cycle, and the modern bull market — and the same rule is applied to each.

Era Realised vol Buy & hold Static 1× CAGR Static 1× max DD Vol-target CAGR Vol-target max DD Vol-target Sharpe
1980–1999 17.52% -58.1% -5.37% -69.45% -4.03% -63.28% -0.339
2000–2026 12.90% +1,345.4% +9.55% -41.12% +9.89% -39.19% 0.846

The 1980–1999 era is the decisive test: gold fell 58.1% over those twenty years. Static 1× compounded at -5.37% per year and never recovered its 1980 peak inside the era — the drawdown ran from 1980-01 to 1999-07 and was still open when the era ended. Volatility targeting cut that drawdown from 69.45% to 63.28% but did not make money: -4.03% per year. In a flat-to-falling asset, a volatility rule changes the shape of the loss, not its sign.

The volatility regime explains the exposure levels. Average trailing-12 volatility was 14.10% in 1980–1999 (92.5% of the full-sample average) and 12.12% since 2000 (79.6%). Over 1980–1999 the rule averaged 1.015× exposure, but that average hides wide swings: it ranged from 0.19× to 2.86× and never once hit its 3× cap (0.0% of months). The full-sample average of 0.819× is lower only because 10.6% of months sat at zero exposure during the pegged-price era, where the rule is undefined. So the rule did de-lever in the bear market — briefly and deeply — but the de-leveraging was not sustained long enough to prevent the 63.28% peak- to-trough loss. Gold's volatility in that decade was high enough to trigger de-risking and not high enough to keep the position small.

Procyclicality: the rule leans into the trend

Volatility-targeting is often described as contrarian because it cuts exposure after a rally that raised volatility. The measurement below shows what actually happened: the direction of the next month's return after exposure was cut, and after it was raised.

Estimator Months de-leveraged Mean next return Of those, months into a rise Months leveraged up Mean next return Of those, months into a fall
roll12 (trailing 12-month stdev) 367 +0.78% 48.8% 328 +0.55% 45.4%
ewma94 (lambda 0.94) 218 +1.34% 53.2% 484 +0.36% 47.1%

The direction of the trade is what the estimate cannot deliver. After the rule cut exposure, the next month rose 48.8% of the time under roll12 and 53.2% under ewma94 — a coin flip. After the rule raised exposure, the next month fell 45.4% of the time under roll12 and 47.1% under ewma94, and the mean next-month return was positive (+0.55% and +0.36%). The rule reduces volatility by construction without adding a repeatable timing edge — it stays exposed into strength and de-risks into weakness, which is why the drawdown comparison above shows a reshaped loss rather than a smaller one.

Where the rule breaks down

1 · The pegged-price era makes the maths undefined

For 84 months, from 1961-02 to 1968-01, gold traded at an administratively fixed price. The standard deviation of monthly returns in that window is exactly zero, so target / σ̂ is a division by zero. This implementation sets exposure to 0 in those months rather than producing an infinite leverage number. No live system could have traded this series: the price was an official rate, and the free market was closed from 15 March 1968 to 1 April 1968.

Sources for the closure: LBMA, 'March 1968 and the London Gold Fixing' (https://www.lbma.org.uk/wonders-of-gold/items/march-1968-and-the-london-gold-fixing) · World Gold Council, 'Historic Gold Agreements' (https://www.gold.org/about-us/what-we-do/central-banks/historic-gold-agreements) · IMF, 'The International Monetary Fund 1966-1971', ch. 20 (https://www.elibrary.imf.org/display/book/9781451971477/ch020.xml)

2 · The drawdown reduction is not a drawdown elimination

The worst rolling 12-month return on the full price series is -36.39%, from 1980-07 to 1981-07. Volatility targeting does not remove that; it changes how much of it reaches the account. The deep drawdown in the targeted curve ran from 1980-01 to 1999-07 — a 63.28% loss that took 81.00 months to recover. The static 1× curve fell 69.58% over a longer stretch and took 82.00 months to get back to its prior peak. Fewer months of pain, but the same fact: the position was underwater across two decades.

3 · The capital is idle exactly when the price is worth holding

Average leverage over the full sample was 0.819× for roll12 and 0.723× for ewma94, and 10.6% of months carried zero exposure under the pegged-price treatment. Exposure below 1× is a drag on the best months, because gold's largest monthly gains arrive in the volatile stretches that trigger de-leveraging.

What the rule did during the two volatility shocks

The two windows below are the sharpest expansions in trailing 12-month volatility in the sample. Each row shows the estimated volatility and the leverage the rule would have set for that month, using only data available at the end of the previous month.

MonthEst. volatility (annualised)Target leverage set
2020-01 9.52% 1.05×
2020-02 10.18% 0.98×
2020-03 10.19% 0.98×
2020-04 9.88% 1.01×
2020-05 10.06% 0.99×
2020-06 9.72% 1.03×
2020-07 9.13% 1.10×
2020-08 10.01% 1.00×
2020-09 10.18% 0.98×
2026-01 11.98% 0.83×
2026-02 13.09% 0.76×
2026-03 12.95% 0.77×
2026-04 15.15% 0.66×
2026-05 16.06% 0.62×
2026-06 17.18% 0.58×
2026-07 20.34% 0.49×
2026-08 21.02% 0.48×
2026-09 21.91% 0.46×

In the 2020 dislocation the rule did almost nothing: estimated volatility was still under 11% when the crash was already under way, because a trailing estimator reacts to volatility only after it appears. The leverage it set in those months stayed close to 1×. By September 2026 the same estimator had measured 21.91% and cut exposure to 0.46×. Over the last 34 months the trailing volatility estimate has averaged 12.06%, ranging from 7.33% to 21.91%.

That gap — reacting late to a volatility shock — is the central weakness of trailing estimators on a monthly frequency. It is the reason the estimator choice is not academic: the ewma94 variant has a materially different leverage profile from roll12 over the same months, and the two produce different return and drawdown numbers from the same underlying price series.

How to use this — and how not to

  • These are historical results on a single price series. They are not a forecast, not a performance claim, and not a basis for sizing a live position on its own.
  • The series is price-only. No financing cost, storage cost, spread, commission, slippage or margin interest is deducted. Every leveraged result here is therefore optimistic relative to a real position, and the deeper the leverage, the more the result depends on costs that are missing.
  • Monthly averages smooth intramonth volatility. Drawdown depths measured on daily closes are deeper than the numbers on this page.
  • A sizing rule that only reduces volatility is a risk-management tool, not a source of edge. The procyclicality table shows the rule does not reliably time reversals.
  • The 1960s comparison set reflects administered prices and a closed institutional market. It is not a usable baseline for anything a retail trader could have executed.

Source: World Bank Commodity Price Data (The Pink Sheet), Monthly Prices, column 70 (Gold, $/troy oz). monthly average price, nominal USD. Data updated November 03, 2026. Generated by scripts/gold_vol_target_analysis.py; every figure above is recomputed from the workbook by that script.

The underlying price series and its drawdown structure are published in full on the gold price history page. For the arithmetic of position sizing, see Gold Position Sizing and the calculators on the tools page. Every estimator above is a trailing one, which means it always reacts to volatility after it has arrived — the same rule driven by what the options market is charging for gold volatility in advance is measured on the gold implied volatility page. For what the tail of the gold distribution actually costs — Value at Risk, Expected Shortfall and whether a stop-loss changes any of it — see gold tail risk and stop-loss evidence. For the same volatility-targeting ladder run on a diversified multi-asset book — where the leverage is spread across seven assets rather than concentrated in gold — see risk parity on leverage. The full narrative version of this dataset — every result above worked through with the procyclicality decomposition and the peg-break artefact — is in Volatility Targeting on Gold.

Frequently asked questions

What does volatility-targeting actually do to a gold position?

It sets exposure each month as f_t = min(3, 10% / sigma_hat_t), where sigma_hat_t is the volatility estimated strictly from returns realised BEFORE the month being sized — no look-ahead. Two estimators are used: roll12 (trailing 12-month stdev) and ewma94 (lambda 0.94). Over the 789 months from 1961-02, the rolling 12-month version averaged 0.819x leverage and produced 11.81% annualised volatility with a -63.28% maximum drawdown, against 15.24% and -69.58% for static 1x. Volatility targeting did reduce the volatility, and it reduced the drawdown — but it also gave up most of the return.

Does volatility targeting improve the risk-adjusted return?

Barely. The Sharpe ratio over the full sample is 0.461 for the rolling 12-month estimator and 0.487 for the EWMA, against 0.406 for static 1x, 0.339 for static 2x and 0.265 for static 3x. Meanwhile CAGR fell from 6.18% (1x) to 5.44% (vol-target) — the sizing rule delivers less than half the CAGR and roughly the same Sharpe. That is the honest summary: volatility targeting is a risk-control tool that changes the shape of the return, not a return enhancer.

What happened to volatility targeting in the 1980s and 1990s?

It was still losing money, just less of it. From 1980-01 to 1999-12 the rolling 12-month version returned -55.9% (-4.03% a year) with a -63.28% drawdown that never recovered within the window; the EWMA version returned -47% with a -54.92% drawdown. Static 1x lost -66.7% over the same period. In 2000 onwards the picture inverts: vol-target returned +1145.4% against buy-and-hold static 1x at +1048.4%, and the EWMA variant did better still at +735.4%.

Is volatility targeting procyclical — does it sell high and buy low?

It deleverages into strength, and the data shows that costs something. With the rolling estimator the position was cut in 367 months, and the average return in those months afterwards was 0.777% — but 48.8% of the time the deleveraging was followed by a RISE, meaning the rule stepped out of the way of gains 48.8% of the time. It also levered up in 328 months and was followed by a fall 45.4% of the time. This is the mechanical price of a rule that responds to realised volatility, and it is why the CAGR drops.

What is the worst twelve-month stretch in this sample?

-36.39% between 1980-07 and 1981-07. Static 1x over the full sample had a -69.58% maximum drawdown, peaking 1980-01 and bottoming 2001-04, with recovery taking 82 months. Volatility targeting with the rolling estimator cut that to -63.28% and the EWMA to -54.92% — a real improvement in the depth of the hole, achieved by giving up roughly half the compound return over the same 789 months.

Does leverage beat volatility targeting here?

On raw CAGR, yes, and badly. Static 2x returned 10.32% a year and static 3x 12.12%, against 5.44% for vol-target with a 3x cap. But the drawdowns are the reason not to take it: -94.36% for 2x and -99.41% for 3x, against -63.28% for vol-target at the same 3x cap. Volatility targeting with a 1x cap produced a -52.58% drawdown, the shallowest in the whole ladder. And the Sharpe ratios run in the opposite direction to the CAGRs: 0.406 for static 1x, 0.339 for 2x and 0.265 for 3x, against 0.483 for vol-target at a 1x cap. Leverage did not buy better risk-adjusted returns, only more volatility and a bigger hole.

What is the current volatility estimate?

Over the last 34 months the rolling realised volatility has averaged 12.06%, with a minimum of 7.33% and a maximum of 21.91%. Realised volatility over the full sample was 15.24% annualised, equivalent to 4.4% monthly. With a 10% target, the implied leverage at the average estimate is 0.83x, and at the maximum it would be 0.46x — below the 3x cap, which is why the cap rarely binds.

What are the limits of this backtest?

Several, and they matter. The series is price-only: no financing, storage, roll or dividend cost is included, so the leveraged results are optimistic relative to a real position — and leveraged gold is exactly where carry bites hardest. There are 84 months from 1961-02 to 1968-01 where the trailing window sits entirely inside the administratively fixed $35/oz price era, so the standard deviation of monthly log returns is exactly zero and inverse-volatility sizing has no defined answer; this implementation sets exposure to 0 in those months rather than dividing by zero, and a live system could not have traded that series at all. Finally, every result is a single-asset price path with no rebalancing friction modelled beyond the monthly step.

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