Research · Dataset

How much gold should you hold?

Every other page on this desk tests gold against something — silver, oil, the dollar, real yields, equities, Bitcoin, housing. Not one of them answers the question a reader actually asks before opening a position: how much gold should I hold? That question has a folklore answer ("ten percent in gold") and a large retail literature, and as far as we could find it has never actually been measured against a real return series with a real drawdown. This page does that: it builds portfolios, measures the damage, finds the best weight in each decade, and then shows you why the best weight does not stay still.

669 monthly observations, 1971-02 to 2026-10 · Gold source updated November 03, 2026

The answer, first

Drawdown improvement 4.4 pp from a 10% gold sleeve, 95% CI 2.1–8.8
Sharpe change +0.07 95% CI 0.04–0.12
Optimal weight 10–100% of the equity sleeve, depending on the decade
Gold / equity correlation -0.0011 monthly, over the full sample

The short version: gold reliably reduces the depth of a portfolio drawdown, it costs almost nothing to hold, and there is no defensible single number for how much to hold — because the number that was best in each decade ranges across most of the entire range.

Source and method

Gold
World Bank Commodity Price Data (The Pink Sheet), Monthly Prices sheet, column 70 "Gold", $/troy oz, monthly average
Equity
FRED NASDAQCOM — Nasdaq Composite index, daily closes; monthly value is the mean of the daily closes inside the calendar month
Cash
FRED TB3MS — 3-month Treasury bill secondary market discount rate, annualised percent; monthly simple return = (1+a/100)^(1/12)-1
Inflation
FRED CPIAUCSL
Coverage
669 months, 1971-02 to 2026-10. The sample starts in 1971-02 because that is where the daily equity series begins.
Portfolio rules
Monthly rebalanced, unlevered, frictionless, no dividend or withholding-tax drag on the equity leg. Every figure on this page is nominal unless it is labelled real.
Significance
Paired monthly series, moving-block bootstrap with 2,000 resamples in 12-month blocks, seed 20,261,115.
Download
gold-allocation-monthly.csv · gold-allocation-episodes.csv
Reproduction
scripts/gold_allocation_analysis.py; independently recomputed from the raw sources by scripts/verify_gold_allocation.py (59 checks).

Two honest caveats before the numbers

  • FRED offers no long daily S&P 500 price series (SP500 starts 2016-11) and WILL5000PR returns HTML rather than CSV, so the Nasdaq Composite is the only long daily equity series that actually downloads. It is growth-heavy and more volatile than a broad index, so the equity leg here is a conservative test of diversification rather than a flattering one.
  • The equity leg is a PRICE index (Nasdaq Composite, not total return). Dividends are therefore excluded, which understates the equity leg and flatters gold.

The three legs, over the whole sample

Leg CAGR Total return Final multiple Ann. vol Max drawdown Sharpe (rf=4%) Worst 12m
Gold 100% 8.71% 10,426% 105.3× 17.1% -62.1% 0.568 -42.3%
Nasdaq Composite 100% 10.52% 26,353% 264.5× 17.4% -74.1% 0.662 -82.4%
Cash (3-month T-bill) 100% 4.34% 971% 10.7× 0.9% 0.0% 4.607 0.0%

Two things are visible immediately. First, the correlation between gold and the Nasdaq over 669 months is -0.0011 — the two assets barely move together, which is the entire reason a gold sleeve can reduce drawdown without simply reducing return. Second, gold is not the safe asset that the phrase "safe haven" implies: on its own it fell 62.1% peak to trough and lost 42.3% over its worst twelve months. Its best month was +48.4%.

Test one: replace part of your equity with gold

The cleanest experiment. One pot of money, split between the Nasdaq Composite and gold, rebalanced monthly, with no cash leg at all.

Portfolio CAGR Ann. vol Max drawdown Sharpe (rf=4%) Final multiple
0% gold / 100% equity 10.52% 17.4% -74.1% 0.662 264.5×
5% gold / 95% equity 10.59% 16.6% -72.0% 0.691 273.1×
10% gold / 90% equity 10.63% 15.8% -69.6% 0.720 279.5×
15% gold / 85% equity 10.66% 15.0% -67.2% 0.750 283.7×
20% gold / 80% equity 10.67% 14.4% -64.5% 0.779 285.5×
25% gold / 75% equity 10.67% 13.8% -61.7% 0.806 285.0×
30% gold / 70% equity 10.65% 13.2% -58.6% 0.831 282.2×
40% gold / 60% equity 10.56% 12.5% -51.9% 0.866 269.8×
50% gold / 50% equity 10.41% 12.2% -44.3% 0.871 249.8×

What the table says

  • Equity alone: 10.52% a year with a -74.1% drawdown.
  • 80/20: 10.67% a year with a -64.5% drawdown. The drawdown is 9.6 points shallower and the return is +0.15 points a year higher — because over this particular half-century gold compounded faster than the Nasdaq. That is a statement about the sample, not a property of the mixture.
  • 50/50: 10.41% a year with a -44.3% drawdown. Now the return is -0.11 points a year below equity alone, and the drawdown is only 29.8 points shallower. Half the portfolio in gold is a materially different risk, not a free lunch.

Test two: add a gold sleeve to a 60/40

The realistic retail question — you already hold a 60/40, how much gold do you add? Here the weight is a fraction of the equity sleeve, so a 10% sleeve means 6% of the total portfolio, and both numbers are stated. This is the allocation tested for significance below.

Gold sleeve (of equity) Gold (of portfolio) Equity Cash CAGR Ann. vol Max drawdown Sharpe (rf=4%)
0% 0% 60% 40% 8.42% 10.5% -52.2% 0.825
5% 3% 57% 40% 8.42% 9.9% -50.1% 0.862
10% 6% 54% 40% 8.41% 9.5% -47.8% 0.900
15% 9% 51% 40% 8.40% 9.0% -45.5% 0.938
20% 12% 48% 40% 8.38% 8.6% -43.0% 0.976
25% 15% 45% 40% 8.35% 8.2% -40.5% 1.013
30% 18% 42% 40% 8.32% 7.9% -37.9% 1.045
40% 24% 36% 40% 8.24% 7.5% -32.3% 1.092
50% 30% 30% 40% 8.14% 7.3% -26.4% 1.103

Note the shape of the return column: it falls monotonically as the gold sleeve grows, from 8.42% with no gold to 8.14% with the largest sleeve. Over this sample the Nasdaq out-competed gold over most of the period, so gold was bought as insurance and paid for out of the return.

Is the difference real? A paired bootstrap

A 4.4-point drawdown improvement sounds like it should be obvious, but with 669 months of history most differences are not distinguishable from noise. Both comparisons below are paired — the same months, the same rebalancing, one portfolio with gold and one without — and tested with a moving-block bootstrap on the monthly difference (12-month blocks, 2,000 resamples, fixed seed).

Comparison (60/40 vs 60/40 + 10% gold sleeve) Base With gold Difference 95% bootstrap CI Distinguishable from zero?
Maximum drawdown -52.24% -47.80% +4.44 pp 2.05 to 8.79 pp Yes
Sharpe ratio (rf=4%) 0.825 0.900 +0.0749 0.0393 to 0.1178 Yes

Both intervals exclude zero. The drawdown improvement is the more robust of the two: it is a path property that depends on where the worst month sits, while the Sharpe difference is an average that a single lucky or unlucky decade can move.

Robustness: against the realised cash return instead of a flat 4% risk-free rate, the 60/40 Sharpe is 0.421 and with the gold sleeve 0.453 (realised cash averaged 4.26% a year). Using month-end equity prices instead of monthly averages gives a 0.635 Sharpe and a -70.5% drawdown for the 10% sleeve, against 0.720 and -69.6% on the monthly-average convention.

The finding that matters: there is no stable best weight

This is the part that gets left out of the folklore. For each decade, find the gold weight that maximised the Sharpe ratio of the 60/40 portfolio — the weight that was, retrospectively, the best one to hold.

Decade Months Best gold sleeve Sharpe there Sharpe with no gold Sharpe with a 20% sleeve
1970s 107 65% 1.661 0.603 1.082
1980s 120 10% 1.098 1.086 1.079
1990s 120 25% 1.495 1.467 1.493
2000s 120 100% 1.197 -0.037 0.144
2010s 120 25% 1.386 1.296 1.378
2020s 82 60% 1.642 1.134 1.333

Five different answers across six decades

  • The 2000s picked 100% — the Nasdaq fell roughly two-thirds and gold rose. Holding nothing but gold was "optimal" in that decade, which is a statement about a bear market, not a portfolio rule.
  • The 1980s picked 10% — gold collapsed from its peak while equities compounded, and a 10% sleeve was the most the sample could bear.
  • The 1990s and 2010s picked 25% and 25% — middling weights in decades where equities did the heavy lifting.
  • Rolling the same search over 92 ten-year windows: the median best weight was 45%, the range was 0% to 100%, and 2.2% of windows said the best weight was zero gold.

What this means. If the optimal weight ranges from 10% to 100% depending on when you started, then any specific recommendation is really a bet on which decade you are in. The honest reading of this dataset is that the risk benefit of a moderate gold sleeve is reasonably stable, and the return benefit is not.

When equities fall hardest: every drawdown of 15% or more

The diversification argument only matters during the bad days, so this is where it gets tested. Every peak-to-trough Nasdaq drawdown of 15% or more in the sample, with what gold and cash did over the same window.

Peak → trough Months Equity Gold Cash 60/40 + 10% gold sleeve
2000-03 → 2002-10 31 -71.8% +5.7% +10.0% -45.3%
1973-01 → 1974-09 20 -56.0% +137.5% +13.4% -27.6%
2021-11 → 2022-10 11 -27.4% -6.4% +1.3% -15.3%
1987-08 → 1987-12 4 -26.7% +7.8% +2.5% -13.6%
1989-10 → 1990-10 12 -28.0% +5.2% +8.3% -12.7%
1983-06 → 1984-07 13 -23.2% -20.5% +10.9% -10.5%
1981-06 → 1982-08 14 -22.7% -24.2% +16.1% -8.6%
2020-02 → 2020-03 1 -15.8% +2.0% +0.2% -8.3%

The largest episode was 2000-03 to 2002-10, 31 months, with the Nasdaq down 74.1%. Over that same window gold returned +5.7% and cash +10.0%. The 60/40 with a 10% gold sleeve returned -45.3% against -71.8% for equities alone.

The pattern in the table is worth stating plainly: gold is not a one-day crash hedge. Over the 31 months of the worst episode it made +5.7% — it did not save anybody. What it did was slow the decline and shorten the depth of the hole. Every episode in the full table is in the published CSV.

What does the insurance actually cost?

A sleeve that helps in a crisis but costs you in good years is only worth holding if the cost is smaller than the benefit. Here it is.

Item Value
Annual portfolio turnover, monthly rebalanced 12.2% of the portfolio
Gold actually traded per year (one-way notional) 2.8% of the portfolio
Assumed round-trip cost 0.30 USD per ounce
Gold price used 4,105 USD per ounce
Estimated annual cost 0.02 basis points of the portfolio
Return cost of the 10% sleeve -0.01 percentage points a year

The rebalancing cost is about 0.02 basis points of the portfolio a year. The return cost of the sleeve is 0.01 percentage points a year. So the trading is close to free and what you are actually paying for the insurance is the opportunity cost of the sleeve — roughly 50 times the trading cost.

Inflation-adjusted, the gap is larger than it looks

CPI rose 746% over this sample. Gold's 8.71% nominal CAGR is 4.63% in real terms, and the cash leg's 4.34% nominal is 0.42% real — inflation took almost all of it. Any allocation comparison that stays in nominal terms is comparing gold against a number that mostly shrank.

Gold sleeve Nominal CAGR Real CAGR
0% of the equity sleeve 8.42% 4.35%
10% of the equity sleeve 8.41% 4.34%
20% of the equity sleeve 8.38% 4.30%

What this page is not

It is not a recommendation of a gold weight. It is a measurement of what a gold weight does, and of how much the answer moves depending on the window. Your own allocation depends on your time horizon, your currency, your tax position, your drawdown tolerance and the rest of your balance sheet, none of which are in this dataset.

It is also not a study of gold as a trading instrument — that is what the rest of this desk is for. This page measures holding a sleeve of physical metal. If you are trading gold with leverage, the position sizing question is a different one entirely, and volatility-target sizing covers it.

No transaction cost, financing, storage or tax is included in any portfolio figure, so the results are if anything flattered. The equity leg excludes dividends. Nothing here is investment advice, and past results do not predict future returns.

Current context

Gold was $4,105/oz in 2026-10, +1.2% over the trailing twelve months, against +16.1% for the Nasdaq and +3.7% for cash, with CPI running at 4.1% year on year. The most recent decade in this sample picked a 60% gold sleeve as its most favourable weight. An allocation decision made now inherits that price level and that decade — it does not inherit a free option.

Frequently asked questions

How much gold should I hold?

This sample does not support a single answer, and saying otherwise would be dishonest. Over 669 months (1971-02 to 2026-10), adding a 10% gold sleeve to a 60/40 portfolio improved the worst drawdown by 4.4 percentage points, from -52.2% to -47.8%, with a 95% bootstrap interval of 2.1 to 8.8 points. The Sharpe ratio moved by 0.07, with a 95% interval of 0.04 to 0.12. But the optimal weight is not stable: the best gold weight ranges from 10% to 100% of the equity sleeve depending on the decade, and across 92 rolling 10-year windows the most favourable weight was zero in 2.2% of them. Treat any single number as an assumption about your own risk tolerance, not a finding.

Does gold actually reduce portfolio risk, or does it just look that way?

It measurably reduces drawdown and it does not do it by accident. The monthly correlation between gold and the Nasdaq Composite over this sample is -0.0011, essentially zero, and the gold leg's own worst drawdown (-62.1%) is far shallower than the equity leg's (-74.1%). Put a 10% gold sleeve into the 60/40 portfolio and the worst drawdown improves by 4.4 points, with a bootstrap interval of 2.1 to 8.8 that excludes zero. The risk reduction is real; what is not real is any promise about how much risk-adjusted return it adds, since that number moves with the sample.

What is the actual cost of holding gold in a portfolio?

Small, and measurable. A monthly-rebalanced 60/40 with a 10% gold sleeve turns over 12.2% of the portfolio a year, but almost all of that is the equity and cash legs rebalancing among themselves. The gold itself trades about 2.81% of the portfolio per year in one-way notional. At the assumed 0.3 USD per ounce round trip and a gold price of 4,105 USD, that works out to about 0.02 basis points of the portfolio a year. Against a return difference of -0.01 pp percentage points a year for the 10% sleeve, the trading is close to free and what you are really paying is the opportunity cost of the sleeve, not the execution.

Is there really no stable best allocation?

No, and this is the most important finding on the page. The per-decade best gold weight (of the equity sleeve, maximising Sharpe) is 65% in the 1970s, 10% in the 1980s, 25% in the 1990s, 100% in the 2000s, 25% in the 2010s, 60% in the 2020s. Five different answers across six decades. Rolling that over 120-month windows, the median best weight was 45% with a range of 0% to 100%, and 30.4% of windows landed between 5% and 25%. The 2000s picked 100% because gold rose and the Nasdaq fell; the 1980s picked 10% for the opposite reason. Anyone who tells you there is one right number is describing a period, not a principle.

What happens to gold during the equity crashes, when the hedge is supposed to matter?

This page lists 12 Nasdaq drawdown episodes of 15% or more. The largest was 2000-03 to 2002-10, an equity fall of -74.1%; over that window gold returned +5.7% and cash +10%. The 60/40 portfolio with a 10% gold sleeve returned -45.3% over the same episode against -71.8% for equities alone. Gold was not a reliable one-day crash hedge; what it did was limit the depth and the duration of the damage across the full episode. The episode table is published as a CSV so you can read every case rather than the summary.

Does this hold up in real, inflation-adjusted terms?

Yes, and the inflation-adjusted numbers are the more honest ones. CPI rose 746% over this sample, so the 8.7% nominal gold CAGR becomes 4.6% in real terms. The 60/40 portfolio without gold compounded at 8.4% nominal and 4.3% real; with a 20% gold sleeve, 8.4% nominal and 4.3% real. Cash compounded at 4.3% nominal but only 0.4% in real terms — inflation took almost all of it. Any allocation argument that only compares nominal returns is really comparing gold against a number that shrank.

How reliable is this, given the data sources and the model?

There are real limitations, and they are stated on the page rather than buried. The equity leg is the Nasdaq Composite, a price index that excludes dividends, so the equity leg's return is understated and gold's relative contribution is flattered. There is no transaction cost, financing, storage or tax in any portfolio figure. The cash leg is a 3-month T-bill rate, which assumes you can roll it at that rate with no funding friction. The sample starts in 1971-02 because that is where the daily Nasdaq series begins, so it includes the 1970s inflation shock and the 2000s equity bust but not the 1994 bond shock as a standalone event. Every number on this page is recomputed independently by a second script with a different code path, and both the monthly series and the episode table are published as CSV.

What is the current starting point?

Gold was 4,105 USD per ounce in 2026-10, +1.2%% over the trailing twelve months, against +16.1%% for the Nasdaq and +3.7%% for cash, with CPI running at 4.1% year on year. The 2020s in this sample picked a 60% gold sleeve as its most favourable weight. A new allocation decision inherits that level and that decade; it does not inherit a free option.

Related reading

Sources

  • World Bank Commodity Price Data (The Pink Sheet) — Monthly Prices sheet, gold column, monthly averages in USD per troy ounce. Data updated November 03, 2026.
  • FRED NASDAQCOM — Nasdaq Composite daily closes; monthly value is the mean of the daily closes inside the calendar month. Last observation 2026-11.
  • FRED TB3MS — 3-month Treasury bill secondary market discount rate, annualised percent.
  • FRED CPIAUCSL — Consumer Price Index for All Urban Consumers.
  • scripts/gold_allocation_analysis.py — production script, writes src/data/research-gold-allocation.json and both CSVs.
  • scripts/verify_gold_allocation.py — independent second implementation; recomputes every published statistic from the raw sources with numpy and a different code path.

Last updated: 2026-11-15

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