gold risk parity
Risk Parity on Gold: Where the Leverage Argument Breaks, and What Beats It
We ran equal-risk-contribution allocation across gold, silver, copper, oil, wheat, coffee and the Nasdaq at 1x, 2x and 3x leverage over 537 common months. Leverage cut the Sharpe ratio every step. But the bigger finding is that the sophisticated weighting barely beat equal weight, and a plain 60/40 gold/Nasdaq book beat the whole thing.
Risk parity is the most technically impressive idea in retail portfolio construction, and it has one very simple pitch: stop allocating by dollars, allocate by risk. Weight each asset so that it contributes the same amount of risk to the portfolio, and no single commodity can wreck you. The pitch usually comes with a picture of a diversified book, and the pitch almost never comes with the leverage number.
That is the gap this study walks into. Real risk-parity products run at two to four times the unlevered book. We took the standard seven-asset universe, built the book properly, sized it to a volatility target, and ran it at 1×, 2× and 3× leverage over the same 537 months.
The leverage result was the expected one, and it is worth stating plainly because it is the whole argument: leverage made the risk-parity book strictly worse. The Sharpe ratio fell at every step, and the drawdown deepened from survivable to unsurvivable.
But the more interesting result was somewhere the pitch never points. On the same data, with the same monthly rebalancing and the same risk target, the sophisticated equal-risk-contribution weighting barely beat simply splitting the money seven ways. And a plain 60/40 book of gold and the Nasdaq beat both of them.
The short version
Over 537 common months (February 1982 – October 2026):
- Leverage destroyed the Sharpe ratio. ERC at 1× scored 0.621; at 2× it was 0.458; at 3× it was 0.403. Meanwhile the maximum drawdown went from −36.87% to −72.81% to −92.96%.
- The allocation scheme barely mattered. ERC 0.621, inverse volatility 0.624, equal weight 0.617. The whole spread was 0.008 — noise, not skill.
- Diversification did not deliver the promised risk reduction. The book targeted 10% annualised volatility and realised 12.15% at 1×, 24.40% at 2× and 36.66% at 3×. The overshoot was roughly 21–23% at every level.
- A plain 60/40 gold/Nasdaq book beat the whole construction, scoring 0.843 with a smaller drawdown (−35.71%) and no leverage.
- The estimation window was worth more than the allocation scheme. Moving the covariance window from 24 to 120 months moved the 1× Sharpe ratio from 0.525 to 0.754.
That last point deserves emphasis. The scheme everyone argues about — equal risk versus inverse volatility versus equal weight — was worth 0.008 of Sharpe ratio. The one parameter nobody argues about, because it was simply chosen in advance and never varied, was worth 0.23.
The data
| Item | Value |
|---|---|
| Assets | Gold, silver, copper, crude oil (average), wheat (US HRW), coffee (Arabica), Nasdaq Composite |
| Source | World Bank Pink Sheet (Monthly Prices); FRED NASDAQCOM; FRED DFF |
| Price basis | Monthly average, USD |
| Common sample | 537 months, 1982-02 to 2026-10 |
| Estimation windows | 24, 36, 60, 120 months (re-run on the same 537 months) |
| Target volatility | 10% annualised, before leverage |
| Leverage levels | 1×, 2×, 3× |
| Rebalancing | Monthly, to weights computed from the prior window |
| Data updated | November 3, 2026 |
Why the sample starts in 1982
The window comparison needs a sample on which every window can be evaluated. A 120-month covariance window requires ten years of history before it produces its first reading, so the common sample starts in February 1982 — the first month for which even the longest window has full trailing history.
This is a real cost and it is worth being explicit about it. The parent dataset page ran on 657 months (1972–2026) with a 36-month window. Every window variant here is measured on the shorter 537-month common sample instead, because otherwise a longer window would be flattered by having more data. The consequence is that the absolute numbers on this page are not directly comparable to the parent page — on the common sample, ERC at 36 months scores 0.621, against 0.736 on the parent’s longer sample. The comparison between schemes and windows, which is what this page is about, is internally consistent.
How the book is built
Monthly returns for seven assets. At each month t, the covariance matrix is computed from the trailing window months, the allocation weights are solved, and the portfolio is sized so that its ex-ante annualised volatility equals 10%. Gross exposure is then multiplied by the leverage factor.
Financing is modelled explicitly: cash earns the effective federal funds rate, and any borrowed amount is charged the same rate. Over a sample where the fed funds rate averaged well above zero for three decades, that is not a rounding error.
Weights for the ERC book are solved by the standard risk-contribution condition — every asset contributes an equal share of portfolio variance. The production script solves it with a closed-form quadratic coordinate update; the verification script solves it with a scipy root finder on the first-order conditions. Two independent algorithms, same answer to within 1e-9 on the parent configuration.
Result 1: leverage makes it worse, every single step
This is the finding the whole exercise was set up to look for, and it is unambiguous.
| Configuration | CAGR | Realised vol | Sharpe | Max drawdown | Median gross leverage |
|---|---|---|---|---|---|
| ERC, 36-month window, 1× | +7.02% | 12.15% | 0.621 | −36.87% | 1.23 |
| ERC, 36-month window, 2× | +8.43% | 24.40% | 0.458 | −72.81% | 2.46 |
| ERC, 36-month window, 3× | +8.03% | 36.66% | 0.403 | −92.96% | 3.70 |
The mechanism is the one the parent dataset page documented in detail: a volatility-targeted book must be levered up when measured volatility falls, which is exactly when volatility is about to rise again. The 36-month window adds a second layer — a covariance estimate that is still dominated by prices from a decade ago is slow to notice a regime change at all.
What is new here is the magnitude of the damage on a common sample with the benchmarks in place. A 92.96% drawdown is not a drawdown; it is the end of the account. At 3× leverage the worst single month was February 1986, and the peak-to-trough span ran from February 1982 to November 2001 — nearly twenty years underwater.
Result 2: the allocation scheme was worth almost nothing
| Scheme (36-month window, 1×) | CAGR | Realised vol | Sharpe | Max drawdown |
|---|---|---|---|---|
| Equal risk contribution (ERC) | +7.02% | 12.15% | 0.621 | −36.87% |
| Inverse volatility | +6.75% | 11.56% | 0.624 | −34.96% |
| Equal weight | +6.26% | 10.81% | 0.617 | −36.89% |
ERC beat equal weight by 0.0038 of Sharpe ratio and lost to inverse volatility by 0.0038. Those differences are far inside the noise of a 44-year monthly series.
This is worth stating carefully, because it is not the same claim as “risk parity does not work.” Risk parity is a coherent idea and the construction here is correct — the ERC weights do deliver equal risk contributions by design. The finding is narrower and more damning: on this asset set, over this period, at this risk target, doing the sophisticated thing was not measurably better than doing the simple thing.
Equal weight actually delivered the lowest realised volatility (10.81% vs 12.15%) and a nearly identical drawdown. The extra machinery bought a higher return and slightly higher risk, and the Sharpe ratio came out the same.
Result 3: what the book is actually competing with
The parent dataset page compared three allocation schemes against each other. It never asked what a gold trader would otherwise do. So here are three benchmarks, all run on the same 537 months with the same data:
| Benchmark | CAGR | Realised vol | Sharpe | Max drawdown |
|---|---|---|---|---|
| 60/40 gold + Nasdaq, monthly rebalanced | +8.57% | 10.43% | 0.843 | −35.71% |
| ERC risk parity, 1× | +7.02% | 12.15% | 0.621 | −36.87% |
| Buy and hold gold | +5.44% | 13.28% | 0.465 | −47.86% |
| Buy and hold gold at a 10% vol target | +4.95% | 11.00% | 0.494 | −59.61% |
Two things follow.
First, the risk-parity book does beat simply owning gold — by 0.156 of Sharpe ratio, and by 1.58 percentage points of annualised return. If the alternative was holding gold, the seven-asset book earned its complexity. That comparison is fair and it is worth stating.
Second, and this is the part that undermines the pitch: a 60/40 gold and Nasdaq portfolio beat the risk-parity book outright, by 0.22 of Sharpe ratio, with a smaller drawdown, with no leverage, and with no covariance matrix. Two assets beat seven.
There is a control worth noting in the same table. Buy and hold gold, sized to the same 10% volatility target, had a median gross exposure of just 0.87 — meaning the sizing rule usually told it to hold less gold than it had capital for, because gold’s trailing volatility was above 10%. That single-asset book scored 0.494, barely better than unsized gold and well behind the multi-asset construction. Volatility targeting did not rescue a one-asset portfolio; the diversification was what did the work.
Result 4: the 60/40 result is not just the Nasdaq bubble
A Sharpe ratio of 0.843 for a book that includes the Nasdaq invites the obvious objection: that is the bubble, not a strategy. So the sample is split into three eras and every run is re-derived within each:
| Era | Months | ERC 1× | Equal weight 1× | Buy & hold gold | 60/40 gold/Nasdaq |
|---|---|---|---|---|---|
| 1982–1999 | 215 | 0.304 | 0.411 | −0.061 | 0.680 |
| 2000–2009 | 120 | 0.894 | 0.874 | 1.073 | 0.630 |
| 2010–2026 | 202 | 0.810 | 0.658 | 0.676 | 1.213 |
The 60/40 book wins two eras out of three and loses the third — and the one it loses is the decade when gold was the strongest single asset, which is exactly the era where a gold-heavy book had nothing to gain from the equity sleeve. In 2000–2009 the Nasdaq fell hard while gold went up 14.90% a year, and a book with 40% in the Nasdaq paid for that.
The honest summary is that the 60/40 advantage is era-dependent, not universal. It is a better book in this sample, on this data, with a Sharpe ratio that is higher in two of three eras and materially lower in the third. Anyone reading this should not take “60/40 beats risk parity” as a rule.
Result 5: the estimation window mattered more than the scheme
The parent dataset page stated plainly that the 36-month covariance window was fixed in advance and never varied. So here it is varied, on the same 537 months, at each leverage level.
| Window | Sharpe (1×) | Realised vol (1×) | Vol overshoot (1×) | Sharpe (2×) | Sharpe (3×) |
|---|---|---|---|---|---|
| 24 months | 0.525 | 13.12% | +31.2% | 0.374 | 0.324 |
| 36 months | 0.621 | 12.15% | +21.5% | 0.458 | 0.403 |
| 60 months | 0.672 | 11.41% | +14.1% | 0.498 | 0.440 |
| 120 months | 0.754 | 10.74% | +7.4% | 0.570 | 0.508 |
The pattern is consistent and it points the same way at every leverage level: a longer estimation window produces a better book. The Sharpe ratio rises monotonically with the window, and the volatility overshoot falls from +31.2% at 24 months to +7.4% at 120 months.
The mechanism is straightforward. A short window reacts quickly to a change in volatility, and volatility is mean-reverting, so a reactive sizer systematically leans into the volatility that is about to fade and shrinks right before a calm stretch. The parent dataset page measured this directly: the ex-ante volatility estimate that sets the position correlates −0.492 with the volatility that actually arrives over the following year. The estimate is a contrarian indicator, not a leading one. A long window is sluggish, so it leans less hard against that mean reversion, and on this data the sluggishness paid: by the time a 120-month estimate has absorbed a regime change, the worst of it is behind you. The parent page’s own diagnostic showed the book was under-levered rather than over-levered in the majority of months — the sizing rule was not systematically chasing the wrong number, it was simply too twitchy.
Two cautions. A 120-month window means ten years before the strategy has produced a single live decision, which is a real operational cost that the backtest does not charge. And this sample contains one crisis of each kind, so “long windows are better” is a statement about 1980 and 2008, not a law.
The comparison is also a warning about how much of any of these results is an artefact of unexamined parameters. The scheme was worth 0.008 of Sharpe ratio. The window was worth 0.23. An unexamined parameter mattered roughly thirty times more than the parameter everyone argues about.
What this study is not
It measures monthly average prices, so every position is assumed to be entered exactly at a monthly average and held to the next one. A real order fills at some intraday price and pays a spread. No transaction costs are charged on the monthly rebalance, which flatters the more actively rebalanced books.
It is price-only. Dividends on the Nasdaq sleeve are excluded, which understates the equity sleeve’s return — so the 60/40 comparison is, if anything, conservative. Commodity storage, futures roll and refining costs are also excluded, which flatters the commodity sleeves inside the risk-parity book. A real risk-parity implementation would hold futures rather than spot, and the roll yield is not in here.
The Sharpe ratio throughout is computed against a zero risk-free rate. That is a deliberate choice — it matches the parent dataset page and makes the two comparable — but it systematically flatters every leveraged run and every equity-heavy run relative to a cash-funded one. The 2× and 3× rows would look worse still against a positive cash rate.
And it does not test a strategy you could trade. It tests monthly rebalancing of a long-only book with no costs, no taxes, no roll, no execution risk and no broker. It does not test whether the correlations that made diversification work in the past will hold in the next crisis.
If you are going to use risk parity anyway
- The diversification is the part that works. The ERC book beat owning gold by 0.156 of Sharpe ratio. The equal-variance construction was what bought that, and it is the part worth keeping.
- The equal-risk-contribution weighting is the part that does not. It was worth 0.008 of Sharpe ratio against equal weight and was beaten by inverse volatility. If you are paying for a risk-parity product, ask what you are actually paying for.
- Do not lever it. Every leverage step in this study reduced the Sharpe ratio and deepened the drawdown, at every window length and every estimation window tested. There is no window at which 2× looks good.
- Size to a volatility target you can actually fund. The drawdown numbers here are the binding constraint: −36.87% at 1× is something a trader can survive, −92.96% is not.
- Use a long estimation window if you use one at all. 120 months produced both the best Sharpe ratio and the smallest volatility overshoot here. The cost is that you spend a decade before the strategy has a live opinion.
- Remember what the comparison was against. The risk-parity book was measured against holding gold and against a 60/40 book, on the same data, at the same cost. Any pitch that only shows you the scheme table is showing you the easy half of the question.
Related reading
- Gold and financial stress: a weekly test, 1971-2026 - the crisis-hedge claim measured against the Chicago Fed NFCI and the St. Louis Fed STLFSI4: 2,912 weekly observations where the correlation between tightening conditions and gold’s weekly return is -0.063, and the tightest 58 weeks were gold’s worst weeks, not its best
- How much gold should you hold? Tested, 1971-2026 - the same question answered with drawdown and Sharpe instead of leverage
- Risk parity on leverage, 1972-2026 — the full dataset page: every scheme at every leverage level, the volatility overshoot control test, the procyclicality decomposition and the published CSV files
- Reading an Equity Curve: Drawdown, Recovery, and What They Hide — how to read the nineteen-year underwater period in the 3× row
- Gold Position Sizing: How Much to Risk per Trade — sizing against a drawdown you can actually fund
- Gold Multi-Account Capital Management — running several accounts against one risk budget
- EA Backtest Traps: Overfitting, Look-Ahead, and Out-of-Sample Testing — why the estimation-window sensitivity in this article matters for any parameter you fit
- Gold and Real Yields: The Mechanism Behind the Dollar Channel — why gold behaves the way it does across these four decades
Educational content only. Not investment advice. Past performance is not indicative of future results. Trading forex and CFDs carries substantial risk of loss.
Frequently asked questions
Does risk parity reduce risk on gold?
Not in the way it is usually claimed. Risk parity lowers the ex-ante variance contribution of each asset by construction, but on this sample the ERC book at 1x leverage still realised 12.15% volatility against a 10% target, a 21.5% overshoot. At 2x leverage the overshoot was 22.0% and at 3x it was 22.2%. Diversifying seven assets did not deliver the risk reduction the construction promised.
Does leverage improve a risk parity portfolio?
No, it makes it worse, monotonically. The ERC Sharpe ratio fell from 0.621 at 1x leverage to 0.458 at 2x and 0.403 at 3x, while the maximum drawdown deepened from -36.87% to -72.81% and then -92.96%. The extra return does not compensate for the extra risk taken to get it.
Is equal-risk-contribution weighting better than just splitting money equally?
Barely. On the common 537-month sample, ERC at 1x produced a Sharpe ratio of 0.621 against 0.617 for equal weight and 0.624 for inverse-volatility weighting. The entire spread between the three allocation schemes was 0.008. Equal weight was slightly worse on both return and drawdown, and delivered the lowest realised volatility of the three.
What actually beat the risk parity book?
A plain 60/40 gold and Nasdaq portfolio, monthly rebalanced and unlevered, with a Sharpe ratio of 0.843 against the ERC book's 0.621. It held a smaller drawdown too, -35.71% against -36.87%. It won in two of the three eras tested and lost badly in 2000-2009, the decade when gold itself was strongest.
How sensitive are these results to the estimation window?
Very, which the parent dataset page never tested. Re-running ERC at 24, 36, 60 and 120-month covariance windows on the same 537 months, the Sharpe ratio at 1x moved from 0.525 at 24 months to 0.754 at 120 months, and the volatility overshoot fell from +31.2% to +7.4%. A choice made in advance and never varied was worth more than 0.22 of Sharpe ratio.
What are the important caveats?
This is a price-only study. Dividends on the Nasdaq sleeve, commodity storage, futures roll and refining costs, and transaction costs on monthly rebalancing are all excluded, so the equity and commodity sleeves are flattered. The 60/40 comparison uses one equity index and one commodity, not a portfolio, and the sample contains exactly one crisis of each kind. The Sharpe ratio is computed against a zero risk-free rate, which flatters every leveraged and every equity-heavy run.
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
- FRED — NASDAQCOM, Nasdaq Composite Index — accessed 2026-11-11
- FRED — DFF, Effective Federal Funds Rate — accessed 2026-11-11
- Spinu, F. (2013) — An Algorithm for Computing Risk Parity Weights — accessed 2026-11-11
- Asness, C., Frazzini, A. & Pedersen, L. H. (2019) — Quality Minus Junk, Journal of Financial Economics 134(1), 193-215 — 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.