gold long term holding period

Gold Long-Horizon Holding Periods: What Actually Happens If You Buy and Hold for a Decade

Buying gold and holding for ten years failed 132 times out of 680 possible start dates, and those failures arrive in long runs rather than scattered years. This article recomputes the run length, the conditional probability of a second losing decade, and the waiting time to break even from the published monthly dataset.

By Auriatech Research Desk · · Last updated 2026-11-15 · 3 min read
goldholding periodlong termhistorical returnsinflationdrawdown Contains our own backtest data
What this is. The parent dataset page (Gold long-horizon holding periods) already answers the basic questions: how many ten-year windows lost money, where they are, how long the 1980 peak took to recover. This article adds the four questions that page does not answer, because they are the ones a person actually deciding on a holding period has to ask:
  1. Do losing decades arrive one at a time, or in runs?
  2. If a decade has just gone wrong, is the next one likely to go wrong too?
  3. What was the spread of outcomes, not just the average?
  4. If I buy at a bad price, how long until I break even?

Every number below is computed from the published monthly and window CSVs of the parent dataset, so this article cannot drift away from the dataset page. The analysis is in scripts/gold_long_horizon_article_stats.py; a second, independent script (scripts/verify_long_horizon_article.py) re-reads the same published CSVs, recomputes every statistic from scratch with numpy, and fails the build if any published number in this article disagrees with the recomputation.

The short version

  • 132 of 680 ten-year windows are negative in nominal terms. 19.41%.
  • 261 of 680 are negative after inflation. 38.38%.
  • Losing decades do not arrive independently. 93.94% of the time a negative decade is followed by another negative decade, against a base rate of 19.44%.
  • The longest losing run is 76 consecutive ten-year windows, from a 1986-09 start to a 1992-12 start, averaging -23.57% each.
  • A buyer at the January 1980 peak waited 316 months (26.3 years) for nominal breakeven and 541 months (45.1 years) for real breakeven.
  • Median ten-year CAGR is 6.39% nominal and 2.51% real. The 5th percentile is -3.45% nominal.

Why the base rate on its own is misleading

The dataset page leads with the share of negative windows: 19.41% nominal over ten years. That number is correct and it is also the wrong number to plan with, for a structural reason that has nothing to do with gold.

Adjacent ten-year windows overlap. The window starting 1980-01 runs to 1990-01; the window starting 1980-02 runs to 1990-02. They share 119 of their 120 months. Two adjacent windows are not two observations of the same phenomenon — they are essentially the same observation, offset by a month. So the 680 windows are not 680 independent draws, and treating them as if they were understates how much the outcomes cluster.

The clustering is measurable. Order the ten-year windows by start month and count how often a negative window is followed by another negative window:

Measure All start dates Starts from 1990 Starts from 2000
Adjacent window pairs 679 319 199
Negative first window 132 43 2
Also negative next window 124 40 1
P(next negative | current negative) 93.94% 93.02% 50.00%
P(next negative), unconditional 19.44% 13.17% 1.01%
Lift over base rate 4.83x 7.07x 49.75x

In real terms the same statistic is starker: 98.08% conditional against a 38.29% unconditional base rate, a lift of 2.56x. If your decade is underwater in purchasing-power terms, the next one almost certainly is too.

Read the first column carefully. If you are midway through a ten-year window that is currently underwater, the historical evidence says the ten-year window that starts one month later is very likely to be underwater too — 93.94% of the time. The unconditional rate is 19.44%. A negative decade is roughly 4.8 times more likely to be followed by another negative decade than the base rate suggests.

That is the serial-dependence result. It is also why the third column is worth reading rather than dismissing: after 2000 there are only 2 negative windows in 200, so the conditional probability is computed on a single observation. A “50%” figure built on one event carries no information. The honest statement is that the modern sample is too small to estimate the conditional probability at all, not that it fell to 50%.

The runs

The serial dependence shows up more concretely as streaks. Grouping the negative ten-year windows into consecutive runs:

Run length (windows) Nominal runs Real runs
1 2 1
2 1 –
3 1 –
5 1 –
9 – 1
17 1 –
22 – 1
27 1 –
76 1 –
202 – 1

There are 8 nominal runs in total, of which 6 run for 2 windows or more, 4 run for 5 or more, and 3 run for 12 or more. The distribution is not “occasional bad decades scattered across six decades” — it is a small number of long, expensive episodes.

The longest nominal run is 76 windows, from a 1986-09 start to a 1992-12 start, with the last window ending 2002-12. The worst single window inside that run returned -40.54%, and the mean across the run was -23.57%. Every one of those 76 windows is a distinct person buying gold, holding for a decade, and losing money.

In real terms the picture is worse and the longest run is longer still: 202 consecutive negative ten-year windows, running from a 1979-02 start to a 1995-11 start, with the last window ending 2005-11. That run’s worst window returned -64.27% and the mean was -34.48%.

The cascade from the worst start month

The single worst ten-year window starts 1980-07 and returns -43.55%. From that start, the cascade continues for 20 consecutive negative ten-year windows, running from 1980-07 to a 1982-02 start, with the final window ending 1992-02. The best return achieved anywhere inside that cascade is -5.35%; the mean is -24.68%.

This is the part that matters for planning. A buyer entering at the worst moment in the series did not face one bad decade. They faced two decades in which every possible re-entry date still lost money, and the best available result was a double-digit annual loss.

The distribution, not the average

An average is a poor summary of a distribution with this shape, because the average ten-year return on gold is dominated by a handful of spectacular starts in the early 1970s.

Ten-year holding-period CAGR, nominal, all 680 start dates:

Statistic Value
Mean 7.57%
Median 6.39%
5th percentile -3.45%
25th percentile 0.66%
75th percentile 13.74%
95th percentile 21.41%
Minimum -5.56%
Maximum 34.44%

Real (inflation-adjusted) ten-year CAGR, same 680 windows:

Statistic Value
Mean 3.46%
Median 2.51%
5th percentile -6.49%
25th percentile -2.08%
75th percentile 8.91%
95th percentile 16.04%
Minimum -9.78%
Maximum 25.08%

Two things to take from these tables.

First, the median nominal ten-year CAGR of 6.39% is a materially more honest planning number than the mean of 7.57%, and both are far above the real median of 2.51%. A nominal decade is a good decade. A real decade, over the full sample, is a flat decade.

Restricting the nominal figures to starts from 2000 raises the mean to 8.86% and the median to 7.77%, with a 5th percentile of 0.84% — a positive number, on a sample of 200 windows.

Second, the 25th percentile of the real distribution is -2.08%. A quarter of all ten-year real holding periods lost money. That is the number to compare against any allocation you are considering.

The five-year horizon, for contrast, has a median CAGR of 4.59% nominal with a 5th percentile of -6.56%, and 184 of its 740 windows are negative. Shortening the horizon does not remove the risk; it moves it.

The waiting time

The dataset page reports the recovery from the January 1980 peak. The more general question — how long does a buyer wait, from any given purchase month, before the price comes back — can be answered the same way, for each purchase month:

Purchase month Purchase price Nominal breakeven Real breakeven
1980-01 $675.00 2006-05 (316 months) 2025-02 (541 months)
1980-07 $643.00 1980-09 (2 months) 1980-09 (2 months)
1981-01 $557.00 2006-03 (302 months) 2011-04 (363 months)
1985-01 $303.00 1985-03 (2 months) 1985-03 (2 months)
1990-01 $410.00 1990-02 (1 month) 1990-02 (1 month)
2011-09 $1,772.00 2020-07 (106 months) 2024-08 (155 months)
2013-04 $1,488.00 2019-08 (76 months) 2020-04 (84 months)

The 1980-01 row is the important one, and it is the same statistic the dataset page publishes: 316 months, or 26.3 years, to come back to $675 in nominal terms, and 541 months, or 45.1 years, in real terms. The real number means the January 1980 buyer did not break even against inflation until February 2025 — 45 years after the purchase.

Two rows deserve a note. The 1980-07 and 1985-01 rows show breakeven in 1-2 months because gold was still climbing out of the 1979-1980 spike at that point; the price fell back to those levels quickly. This is the mirror image of the 1980-01 row and it is the same phenomenon: the spike had not yet peaked. A buyer in July 1980 bought a still-rising price; a buyer in January 1980 bought the top.

The 2011-09 row is the modern version of the problem. 106 months of nominal breakeven and 155 months of real breakeven — 8.8 years and 12.9 years respectively — for a buyer at the post-crisis high.

The modern era is different, and the sample is small

Everything above is dominated by the 1970s and 1980s. Restricting to ten-year windows starting in 2000 or later:

Measure Nominal Real
Windows 200 200
Negative windows 2 23
Share negative 1.00% 11.50%
Worst window -4.75% (2012-10 start) -25.96%
Median 10y return +111.34% —

Only 2 of 200 modern ten-year windows lose money nominally. In real terms 23 of 200 lose money, and the worst is -25.96%.

The temptation is to read the modern column as “gold no longer has losing decades”. The honest reading is narrower: the specific catastrophe that dominated the historical record — a price spike followed by a multi-decade real-terms bleed — has not repeated since 2000. The modern sample also cannot support the conditional probability, because with only 2 negative windows there is nothing to condition on. Twenty-five years of history is not evidence of a changed process; it is a shorter sample of the same one.

The daily benchmark fix gives the same answer from an independent source. Over 14,723 fixings from 1968-04-01 to 2026-11-10, the 2,520-trading-day (10-year) window is negative for 22.41% of the 12,203 windows, with the worst at -50.46% starting 1980-01-21. Over the 5,040-day (20-year) window, 11.47% of 9,683 windows are negative and the worst is -64.44%, also from a 1980 start. The 20-year negative windows all begin before 1990; there are 0 negative 20-year windows starting from 1990 onward.

What this does and does not tell you

This is an analysis of a price series, not a recommendation. Nobody should read a 19.41% ten-year loss rate and conclude that gold is unsuitable as a holding, or that the 1.00% modern loss rate means it is suitable. Both conclusions over-read the data.

What the numbers do support:

  • A ten-year gold holding period has a materially different risk profile from a five-year one, and the difference is not the average return, it is the tail.
  • The distribution is clustered, so “what are the odds my decade loses money” is not a coin flip with a 19.41% probability. If the decade you are in is underwater, the next one probably is too.
  • Real, inflation-adjusted, the median decade has returned 2.51% a year with a 25th percentile of -2.08%, over a sample that contains a genuine gold mania and a genuine price collapse.

What they do not support:

  • Any inference about a purchase made today. The 1980 spike and the 2011 spike are different objects, and the modern sample of 200 windows is far too small to estimate the tail.
  • Any statement about a leveraged position. These returns are unlevered price returns. Leverage multiplies both the wait and the loss; a levered holder of the January 1980 peak had a 26-year wait for a position that was already subject to a margin call.
  • Any claim about total return. Storage, insurance, spread, tax and financing costs are not in these numbers. For a physical holder, spot gold is the worst-performing way to hold gold over long horizons.

How the numbers were verified

The analysis lives in scripts/gold_long_horizon_article_stats.py. It reads the published CSVs of the parent dataset:

  • public/data/gold-long-horizon-monthly.csv — 800 monthly rows, 1960-01 to 2026-09, with the nominal price and the CPI-deflated index (1980-01 = 100)
  • public/data/gold-long-horizon-windows.csv — every rolling window for each horizon and series

and writes:

  • src/data/research-gold-long-horizon-article.json — the run lengths, the conditional probabilities, the CAGR distributions, the waiting times, the cascade and the modern-era summary
  • public/data/gold-long-horizon-streaks.csv — one row per negative run

The verification script scripts/verify_long_horizon_article.py is a separate implementation. It does not import the analysis script and does not read the article’s JSON for any recomputation. It re-reads the published CSVs, rebuilds the run structure, recomputes the conditional probabilities, the CAGR percentiles, the recovery months for every purchase month, the cascade and the modern-era figures with numpy, cross-checks the published streaks CSV against its own run reconstruction, confirms the parent dataset’s own 1980 recovery figures use the same recovery definition, and then checks every literal number in this article against those recomputations.

The last run: all checks passed, 0 failures.

The parent dataset’s own verification, scripts/verify_gold_long_horizon.py, recomputes the window statistics from the same published CSVs with numpy and runs as part of this article’s build gate.

Frequently asked questions

Has gold ever had a negative ten-year return?

Yes, and it has done so repeatedly. On the published monthly dataset, 132 of the 680 possible ten-year windows are negative in nominal terms, a share of 19.41%. The worst is -43.55% for the window starting 1980-07. The common claim that gold never had a negative ten-year return is false as stated.

Are the losing ten-year windows independent events?

No, and this is the part the base rate hides. Of the 679 adjacent pairs of ten-year windows, 132 have a negative first window, and 124 of those 132 are followed by another negative window. That is a conditional probability of 93.94% against an unconditional base rate of 19.44%, a lift of about 4.8x. Once a decade goes wrong, the next one is very likely to go wrong too, because the two windows share 119 of their 120 months.

How long were the losing stretches?

The longest run of consecutive negative nominal ten-year windows is 76 windows, running from a 1986-09 start to a 1992-12 start, with the last of them ending 2002-12. The mean return across that run is -23.57%. There are 8 separate runs in total: two of length 1, and then runs of 2, 3, 5, 17, 27 and 76.

What did a buyer who bought at the 1980 peak have to wait for?

316 months, or 26.3 years, for the nominal price to come back to the $675 purchase price, and it did so in 2006-05. In real terms the wait was 541 months, or 45.1 years, and the real breakeven did not arrive until 2025-02. The 1980 peak was $2,910 per ounce in today's CPI terms when this analysis was run.

Does any of this apply to a buyer starting in the modern era?

The historical risk is concentrated in the 1970s and 1980s. Among the 200 ten-year windows starting in 2000 or later, only 2 are negative, a share of 1.00%, and the worst is -4.75% for the 2012-10 start. In real terms the modern era is materially worse: 23 of the same 200 windows are negative, a share of 11.50%, with the worst at -25.96%. Starting after 2000 also breaks the serial dependence, because there are too few losing windows to estimate it.

What are the important caveats?

Monthly gold prices are monthly averages, so the windows are not executable quotes, and no storage, insurance, spread, tax or financing cost is included. The 1960-1971 price was administratively pegged, which removes genuine price discovery from the early sample. The CPI-deflated series pairs a monthly average price with a monthly CPI index, so real returns are an approximation. The ten-year windows overlap by 119 months, so the 680 windows are far fewer independent observations than the count suggests, and every conditional probability in this article inherits that dependence. Nothing here is a recommendation to hold gold or to use any particular allocation.

Sources

  1. World Bank Pink Sheet, monthly average gold price (USD per troy ounce) — accessed 2026-11-15
  2. FRED CPIAUCSL — US CPI-U, index 1982-84=100 (used to deflate the gold series) — accessed 2026-11-15
  3. LBMA Gold Price PM benchmark fix, daily (sampling-frequency cross-check) — accessed 2026-11-15
  4. Gold long-horizon holding periods, 1960-2026 (parent dataset page this article extends) — accessed 2026-11-15

Use the free tools

Open-source calculators — no signup required.

Use the free tools
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.

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