Research · Dataset

Risk parity on leverage, 1972–2026

Risk parity is sold on one promise: allocate by risk, not by capital, and no single asset dominates the outcome. It is usually demonstrated on an unlevered book. Real risk parity products run at two to four times that, and the argument almost never survives the multiplier. So we ran the thing ourselves — seven assets, monthly rebalancing, a 36-month estimation window, a 10% volatility target, and the same allocation at 1×, 2× and 3× leverage. Here is what breaks, and one place where the criticism turns out to be overstated.

657 monthly observations, 1972-02 to 2026-10 · Source data updated November 03, 2026

What we ran

The universe is 7 liquid-ish sleeves: gold, silver, copper, crude oil, wheat, coffee and the Nasdaq Composite. Prices come from the World Bank Pink Sheet monthly averages; the equity sleeve is FRED NASDAQCOM daily closes averaged to calendar months; the cash rate is the FRED effective federal funds rate. The common sample runs 1972-02 to 2026-10.

Each month we estimate a 36-month covariance matrix, compute weights three ways, then scale the whole book to a 10% annualised volatility target times the leverage multiplier:

  • Equal weight — one seventh each. The control group.
  • Inverse volatility — weight proportional to 1/σ. The popular retail approximation of risk parity.
  • ERC (risk parity proper) — weights solved so every asset contributes exactly 1/7 of portfolio variance. Computed with the Newton-type cyclic algorithm of Spinu (2013), and independently re-solved by a different method in the verification script.

Excess capital is financed at the fed funds rate and the cash left over earns the same rate, so financing is priced rather than ignored. Every number below is produced by scripts/gold_risk_parity_leverage_analysis.py and independently recomputed from the published CSV by scripts/verify_gold_risk_parity.py, which uses numpy and a different solver and a different code path.

The answer, first

ERC Sharpe, 1× → 3× 0.74 → 0.48 falls with every step of leverage
ERC max drawdown, 3× -97% vs -37% at 1×
Realised vol vs target (2×) 23% 24.6% against a 20% target
Financing cost 49% of the mean gain, ERC 2×
The short version. Risk parity's diversification promise is real on paper — the ERC book does spread variance evenly and does diversify better than equal weight. But scaling it up degrades the risk-adjusted return monotonically while the drawdown compounds faster than the leverage, the volatility target is missed by roughly a quarter, and financing takes close to half the gain. The one place the usual criticism is too strong: the "it adds leverage right before volatility rises" charge is mostly the arithmetic of the sizing rule, not a separate defect.

1. The diversification is real — and it does not pay for the leverage

The case for risk parity starts from a measurement, and that measurement holds up here. The ERC book's diversification ratio — the weighted average asset volatility divided by the portfolio's own volatility — is 2.130, against 1.892 for equal weight over the same period. And the defining property holds exactly: in every rebalance, each of the 7 assets contributes 14.29% of portfolio variance. The largest residual gap between the highest and lowest risk contribution across the whole sample is 1.5e-11 — the solver is converged, not approximately converged.

The problem is what happens when you turn this into a levered product. Three schemes, three leverage levels, everything else held fixed:

Scheme Leverage Realised vol vs target CAGR Sharpe (rf=0) Max drawdown Worst month Median gross lev
Equal weight 1× 10.9% +8.7% +7.55% 0.725 -36.9% -14.2% 0.96
Equal weight 2× 21.8% +8.9% +9.11% 0.510 -69.7% -28.5% 1.91
Equal weight 3× 32.7% +9.0% +9.36% 0.438 -93.8% -42.7% 2.87
Inverse volatility 1× 11.8% +17.9% +8.44% 0.748 -35.0% -14.2% 1.14
Inverse volatility 2× 23.6% +18.1% +10.69% 0.550 -80.7% -28.5% 2.28
Inverse volatility 3× 35.5% +18.3% +11.35% 0.483 -97.6% -42.8% 3.43
ERC (risk parity) 1× 12.3% +22.8% +8.61% 0.736 -37.3% -16.3% 1.21
ERC (risk parity) 2× 24.6% +23.1% +10.86% 0.545 -83.1% -33.3% 2.42
ERC (risk parity) 3× 37.0% +23.2% +11.29% 0.481 -97.1% -50.3% 3.63

Read the Sharpe column down each block. For equal weight it falls 0.725 → 0.510 → 0.438. For ERC it falls 0.736 → 0.545 → 0.481. For inverse volatility it falls too. There is no leverage level at which any of the three schemes produces a better risk-adjusted return than its own unlevered version. The CAGR column does edge up with leverage, but the extra return is smaller than the extra volatility it requires — which is exactly what a falling Sharpe ratio means.

The drawdown column is where the asymmetry lives. Going from 1× to 2× roughly doubles the drawdown; going from 2× to 3× takes it from -83% to -97%, where a book that loses 97% of its equity has left very little room for anyone still holding. The Calmar ratio, which divides CAGR by the absolute drawdown, is 0.231 at 1× and 0.116 at 3× — a threefold deterioration for a book whose headline growth rate went up.

The promise is kept on diversification and broken on risk-adjusted return. Risk parity makes the portfolio's risk contribution more even. It does not make the portfolio's risk-adjusted return better once you lever it.

2. The volatility target is missed, and the ERC book misses it worst

A volatility-targeting rule is supposed to deliver the target. Here the ERC book at 1× leverage produced 12.3% against a 10% target — a 22.8% overshoot. At 2× it produced 24.6% against a 20% target, an overshoot of 23.1%. At 3×, 37.0% against 30%.

The interesting part is that the overshoot is not the same for every scheme. Equal weight over-shot by 8.7%, inverse volatility by 17.9%, and ERC by 22.8%. The more the allocation is driven by the covariance estimate, the worse the realised volatility tracks the target — which is the first sign that estimation error is the binding constraint, not the allocation scheme.

The ERC 2× book also shows how far the realised volatility strays from the target when the regime turns. Its trailing 12-month volatility has a median of 20.3%, a 90th percentile of 33.5%, a 99th percentile of 44.9% and a maximum of 51.2%. In 32 of 610 months (5.2%) the book's own trailing volatility was more than double the 10% target it was supposed to be holding.

A target you miss by a quarter is not a target. The position size is set from a 36-month covariance estimate; the risk that arrives is generated by a different covariance matrix.

3. Is the procyclicality charge fair?

The standard criticism of volatility targeting is that it is procyclical: it leans up when volatility looks calm and de-leans when volatility looks ugly, which is precisely backwards. The raw evidence here supports the charge. Gross leverage in the ERC 2× book correlates 0.522 with the realised volatility of the following twelve months, and the bootstrap 95% interval on the high-versus-low leverage tercile difference in forward volatility is 6.0% to 10.9% — it does not cross zero.

But this needs a control, because leverage is a deterministic function of the volatility estimate that produced it. A book with low estimated volatility is by construction a book with high leverage, so some positive correlation is arithmetic, not a defect. We conditioned the leverage/forward-volatility correlation on that estimate directly. The partial correlation drops from 0.522 to 0.203. For reference, conditioning on trailing realised volatility instead — the standard control — leaves 0.339, because trailing realised volatility and the correlation estimate are nearly the same thing here (they correlate 0.044).

The more informative number is the estimate's own forecast quality. The ex-ante volatility estimate that set the position correlates -0.492 with the volatility that actually arrived over the next year — a negative correlation. The median ratio of arriving volatility to the estimate was 1.257, and in 36.0% of months the volatility that arrived was lower than the estimate, so the book was under-levered rather than over-levered. The sizing rule is not systematically chasing the wrong number.

The second piece of evidence points the same way. Comparing forward returns rather than forward volatility, the months with the highest leverage returned 1.25% per month on average against 0.98% for the lowest-leverage months — but the bootstrap interval on that difference runs from -1.23% to 1.75% and crosses zero. High leverage did not reliably buy better returns.

Verdict: partly overstated. The procyclicality is real but smaller than the raw correlation implies, and the risk here is estimation error rather than a perverse feedback loop. That said, "not a perverse feedback loop" is not a recommendation — the leverage still degrades the Sharpe ratio and deepens the drawdown.

4. Where the losses actually came from

The ERC 2× book drew down -83.1% from its 1980-02 peak to its 1991-01 trough. The published CSV carries every month of that path, with the gross leverage and the equity curve on each row.

Month Gross leverage Portfolio return Equity
1980-02 (peak) 1.86 +0.99% 1.0099
1991-01 (trough) 3.37 -3.64% 0.1708

At the trough the book was running 3.37 times equity. Its worst single month was -33.3% in 1986-02. Comparing the worst decile of months against the rest of the sample, median gross leverage was 2.62 in the worst decile against 2.37 elsewhere — a difference of 0.25 times equity. The sizing rule was leaning lower into the bad months, not higher; the damage came from the size of the position, not from the direction of the tilt.

The era breakdown shows how much of any headline number depends on which window you look at:

Era Months CAGR Realised vol Worst month Median gross lev Max gross lev
1972-1984 119 +11.73% 24.2% -18.8% 1.95 3.29
1985-1999 180 +1.27% 26.2% -33.3% 3.05 3.94
2000-2009 120 +18.35% 26.3% -32.6% 2.32 2.80
2010-2019 120 +5.08% 19.1% -11.9% 2.17 3.15
2020-2026 82 +31.38% 25.9% -26.8% 2.03 3.04

The best stretch was 2020-2026 at +31.38% a year; the worst was 1985-1999 at +1.27% a year. In the worst era the book ran its highest median leverage (3.05) and produced almost nothing. That is the pattern that matters for anyone sizing a real book: the leverage the rule chooses is high in the regimes where the payoff is worst.

5. Financing takes nearly half the gain

A book at 2.42 times equity is financed, and financing is not free. Cash earns the effective federal funds rate and the borrowed amount is charged the same rate. Over the 621 simulated months the ERC 2× book paid 3.43 in cumulative return units — equal to 49.5% of the mean gain — at a mean monthly drag of 0.553% of equity per month.

This is not an argument that leverage is never worth it; it is an argument that the comparison has to be made after financing. A rule that targets 10% volatility and runs at 2.42 times equity is running a materially different strategy from the same rule at 1×, and the return ranking between them is close enough that the financing bill decides it.

6. What the weights actually did

ERC weights are not stable. Over the sample the gold weight ranged from 10.3% to 105.8% of equity, with a median of 42.2%; the Nasdaq sleeve ranged from 16.2% to 119.1%, median 44.0%. Weights above 100% of equity are what leverage looks like on a sleeve — the equity is borrowed and the sleeve is the larger position.

Asset Min weight Median weight Max weight Share of time within 1% of its max
gold 10.3% 42.2% 105.8% 0.64%
silver 6.2% 21.4% 69.2% 0.32%
copper 9.3% 29.3% 57.2% 0.16%
oil 4.2% 21.7% 118.6% 0.16%
wheat 12.2% 37.3% 80.6% 0.16%
coffee 11.5% 25.5% 111.5% 0.16%
nasdaq 16.2% 44.0% 119.1% 0.16%

Three snapshot rebalances show the mechanism:

Month goldsilvercopperoilwheatcoffeenasdaq
1975-01 15.4%9.3%13.6%3.4%14.6%18.9%24.8%
1999-05 30.2%14.4%12.1%11.1%11.9%8.1%12.3%
2026-10 16.1%5.0%14.5%10.9%14.7%15.3%23.4%

Note that the weights are near-identical in every snapshot even though the underlying assets have wildly different volatilities. That is the point of the scheme: the covariance estimate moves the allocation enough to equalise risk contributions. It also means the ERC book's capital exposure to a given asset moves a long way over time — the sleeve that carried 44.0% of equity at the median carried as little as 16.2% at times.

The diversification comes from the correlation structure, and that structure is not constant. Median pairwise 36-month correlations across the sample:

Pair Min Median Max Share of windows negative
Gold / Nasdaq -0.605 -0.044 0.482 56.8%
Coffee / Nasdaq -0.476 0.025 0.525 45.4%
Oil / Wheat -0.510 0.040 0.402 39.8%
Wheat / Nasdaq -0.350 0.064 0.495 33.8%
Oil / Coffee -0.431 0.073 0.573 36.6%
Oil / Nasdaq -0.714 0.104 0.752 41.5%
Gold / Wheat -0.365 0.104 0.406 25.4%
Silver / Wheat -0.381 0.104 0.457 21.6%
Silver / Nasdaq -0.288 0.112 0.557 25.3%
Gold / Coffee -0.411 0.115 0.467 22.4%
Gold / Oil -0.318 0.121 0.579 26.6%
Silver / Coffee -0.415 0.124 0.558 21.4%
Copper / Coffee -0.442 0.127 0.504 28.5%
Copper / Wheat -0.311 0.140 0.495 33.7%
Silver / Oil -0.259 0.142 0.581 16.9%
Wheat / Coffee -0.281 0.143 0.683 22.5%
Copper / Oil -0.172 0.157 0.846 21.9%
Gold / Copper -0.203 0.215 0.770 11.6%
Copper / Nasdaq -0.563 0.233 0.813 20.1%
Silver / Copper -0.243 0.381 0.752 21.4%
Gold / Silver 0.161 0.710 0.925 0.0%

Gold and silver are the closest pair by a wide margin — median correlation 0.710, never negative across the 36-month windows. Gold and oil are the least reliable diversifier of the group: median 0.121 but 26.6% of windows negative and a minimum of -0.318. A risk-parity book that leans on the gold/oil relationship is leaning on the least stable leg of the set.

What this page is not

This is a price-only study. The Nasdaq sleeve carries no dividends, the commodity sleeves carry no storage, roll or refining costs, and no transaction cost is applied to the monthly rebalancing. Every one of those is a real cost, and every one of them falls harder on a levered book than an unlevered one, so the true drag on this strategy is larger than the 49.5% financing figure shown above.

Three methodological choices were fixed in advance and never varied: the 36-month estimation window, the 10% target volatility, and monthly rebalancing. We did not test 24-month or 60-month windows, we did not test alternative targets, and we did not test weekly rebalancing. The results are conditional on those three choices, and some of them could reasonably be better than the ones used here.

The drawdown statistics rest on a single survivor path. The sample contains one 1980s inflation bust and one 2008 crisis, and a strategy that survives both tells you it survived those two, not that it survives everything. The 5 era rows make the point directly: the same book produced +31.4% a year in one era and +1.3% in the next.

Nothing on this page is investment advice, and past results do not predict future returns. A risk-parity allocation that behaved this way over fifty-four years can behave differently over the next five, and the drawdown column above is a reminder of how much of the answer depends on which years you happened to get.

Related reading

Frequently asked questions

Does risk parity actually reduce risk on these assets?

On an ex-ante basis, yes: the ERC book's diversification ratio is 2.130 against 1.892 for equal weight, and each asset contributes an equal 14.29% of portfolio variance by construction. But that is a statement about the estimate, not about what the portfolio then experiences. Realised volatility for the ERC book at 2× leverage was 24.6% against a 20% target — a 23.1% overshoot.

Does leverage improve the risk parity portfolio?

No. The ERC Sharpe ratio (excess return over zero, annualised) falls from 0.736 at 1× to 0.545 at 2× and 0.481 at 3×, while the maximum drawdown deepens from -37.3% to -97.1%. The same monotone decline appears for equal weight and inverse volatility, so it is a property of scaling this book, not of the risk-budget scheme.

Is volatility targeting procyclical here — does it add leverage right before volatility rises?

The raw correlation between gross leverage and the following 12 months of realised volatility is 0.522. But leverage is a deterministic function of the volatility estimate that produced it, so part of that correlation is arithmetic. Controlling for that estimate directly, the partial correlation falls to 0.203, and the volatility estimate itself correlates -0.492 with the volatility that follows. Most of the apparent procyclicality is the sizing rule working as designed, not evidence of a separate failure.

What did financing cost?

The book holds gross leverage above 1, so the shortfall is financed. Cash earns the effective federal funds rate and the borrowed amount is charged the same rate, over 621 simulated months. Total financing cost was 3.43 in cumulative return units, equal to 49.5% of the mean gain, at a mean monthly drag of 0.553% of equity per month.

How large was the worst stretch?

The ERC 2× book drew down -83.1% from its 1980-02 peak to its 1991-01 trough, a span of 132 months. Its worst single month was -33.3% in 1986-02. Gross leverage at the trough was 3.37 times equity.

What are the important caveats?

This is a price-only study. Dividends on the Nasdaq sleeve, commodity storage, futures roll and refining costs are all excluded, so the drag on the equity and commodity sleeves is overstated and the real risk-parity portfolio would look somewhat better than this one. The window is 36 months, the target volatility is 10%, and rebalancing is monthly; none of those three choices were varied, so the results are conditional on them rather than general. A single survivor path through 1980 and 2008 dominates the drawdown statistics, and the sample contains exactly one crisis of each kind.

Sources

  • World Bank Commodity Price Data (The Pink Sheet) — Monthly Prices sheet: gold, silver, copper, crude oil average, US HRW wheat, coffee arabica. Data updated November 03, 2026.
  • FRED NASDAQCOM — Nasdaq Composite daily closes, averaged to calendar months.
  • FRED DFF — effective federal funds rate, monthly average. Used for both cash and borrowing rates.
  • F. Spinu, An Algorithm for Computing Risk Parity Weights, SSRN 2297383, 2013 — the Newton-type cyclic solver used for the ERC weights.
  • S. Maillard, T. Roncalli and J. Teiletche, On the Properties of Equally-Weighted Risk Contributions Portfolios, Journal of Portfolio Management 36(4), 2010, pp. 60-70 — the risk-contribution definition and the diversification-ratio measure used here.
  • C. Harvey, J. Campbell, N. Hugonnier and B. Rattray, The Impact of Volatility Targeting, Journal of Portfolio Management 45(1), Fall 2018, pp. 14-33 — the reference result that volatility scaling raises Sharpe ratios for risk assets but has negligible effect for commodities, which is consistent with the Sharpe column falling here.
  • scripts/gold_risk_parity_leverage_analysis.py — production script, writes src/data/research-gold-risk-parity.json and the CSV.
  • scripts/verify_gold_risk_parity.py — independent second implementation; recomputes every published statistic from the published CSV with numpy, a different ERC solver and a different code path.
  • gold-risk-parity-monthly.csv — every month: all seven prices, all seven returns, the fed funds rate, and the ERC, equal-weight and inverse-volatility 2× returns, leverage and equity curve.

Last updated: 2026-11-11

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