Research · Dataset

Gold and the calendar

"Gold is strongest in January." "September is the month to be flat." These claims get repeated every year without ever being measured properly. This page measures them over 54 post-peg calendar years and then tries hard to prove them wrong — because the point is not to find a seasonal edge, it is to find out whether one exists.

Data through 2026-10 · Source data updated November 03, 2026

Source and method

Publisher
World Bank Commodity Price Data (The Pink Sheet)
Series
Monthly Prices sheet, columns 70 (Gold, $/troy oz) and 72 (Silver, $/troy oz)
Frequency
Monthly average price, nominal USD
Coverage
802 months, 1960-01 to 2026-10 (65 complete calendar years)
Primary sample
54 complete years, 1972-2025 (post-peg)
Download
gold-seasonality.csv (month, both prices, both returns, both monthly means)
Source link
https://www.worldbank.org/en/research/commodity-markets → CMO-Historical-Data-Monthly.xlsx
Reproduction
scripts/gold_seasonality_analysis.py; independently recomputed by scripts/verify_gold_seasonality.py

Prices are monthly averages, so a "month return" here is the change in the monthly average price — not an entry and exit you could have executed. January's return is the move from December's average to January's average. No transaction cost, financing or storage cost is included anywhere on this page.

Why 1972 and not 1960. US convertibility was suspended in August 1971, so gold was a fixed $35/oz price through 1971. Monthly "returns" in those years are the trading noise around the peg plus one regime break, not market price discovery. The primary table is therefore computed on the 54 complete years from 1972; the full-sample table (including the peg years) is shown next to it so the effect of the choice is visible.

The baseline you are comparing against

Mean monthly return +0.70% arithmetic, over 801 months
Actual compounded 7.40% per year, the price path itself
Annualised volatility 15.8% monthly σ × √12
Up months 43.4% AR(1) = 0.269

A note worth reading: the arithmetic mean monthly return annualises to 8.68%, but the price path actually delivered 7.40% a year. The gap is the compounding drag from monthly returns being negatively skewed — the big down months are rarer but larger than the up months they come with. Any seasonal claim that ignores this is measuring the wrong baseline.

Monthly returns are serially correlated (AR(1) = 0.269), which means a month is not an independent draw from the year. Every test below is built to respect that.

The month table

Mean change in the monthly average gold price, by calendar month, over the 54 complete years from 1972-2025. The confidence interval is a year-block bootstrap: whole years are resampled, so the dependence between the twelve months of the same year is kept.

Month Mean Trimmed mean Bootstrap 95% CI Median Up years t Naive p Best Worst
January +2.66% +1.65% 0.85 to 5.06 +1.51% 64.8% 2.45 0.0144 +48.35% -6.34%
February +1.39% +1.08% 0.17 to 2.69 +0.73% 51.9% 2.18 0.0296 +16.28% -10.23%
March +0.08% +0.27% -1.26 to 1.39 +0.33% 51.9% 0.11 0.91 +13.51% -16.69%
April +0.55% +0.40% -0.47 to 1.60 -0.58% 46.3% 1.02 0.31 +9.69% -6.68%
May +0.62% +0.14% -0.44 to 1.76 -0.34% 38.9% 1.08 0.28 +12.24% -5.23%
June +0.40% -0.13% -0.81 to 1.77 +0.00% 48.1% 0.59 0.555 +17.65% -11.70%
July +0.13% +0.25% -0.94 to 1.21 +0.00% 48.1% 0.23 0.82 +7.62% -11.28%
August +0.90% +0.85% -0.20 to 2.10 +0.57% 55.6% 1.52 0.129 +11.82% -10.83%
September +1.47% +1.04% 0.09 to 2.97 +0.29% 50.0% 1.99 0.0464 +20.05% -11.66%
October +1.06% +0.61% 0.04 to 2.19 +0.00% 48.1% 1.91 0.056 +17.36% -5.98%
November +0.05% -0.29% -1.09 to 1.32 -0.34% 48.1% 0.09 0.933 +14.47% -9.25%
December +0.65% +0.53% -0.54 to 1.87 +0.42% 51.9% 1.05 0.293 +16.07% -13.92%

The trimmed mean drops the top and bottom 10% of the years, so one outlier month cannot carry a whole row. January survives it (+1.65% trimmed vs +2.66% raw). "Up years" is the share of years the month finished higher than the month before it.

The shape is familiar: a strong start to the year, a soft middle, and a November that carries the "worst month" reputation while sitting almost exactly flat (+0.05% mean, and -0.29% after trimming). March is the true low point at +0.08%. But look at the p-values before reading anything into the table. Twelve monthly means are twelve simultaneous tests, and under a pure-chance null about 0.6 of them would look significant at the 5% level.

Test 1: the whole table against a shuffled calendar

The right null hypothesis is not "each month is zero" but "the months carry no calendar information at all". So the twelve months are shuffled within each year — each year keeps its macro environment, its 1973 spike and its 1987 crash, but the calendar labels are destroyed. The statistic compared is the R² of calendar-month dummies on the monthly return.

Gold R² = 0.0208 permutation p = 0.207 over 2,000 shuffles
Silver (control) R² = 0.0228 permutation p = 0.164

The month table as a whole: does not survive.

Calendar month explains about 2.1% of the month-to-month variance in gold. Against the shuffled-calendar distribution that is p = 0.207 — not distinguishable from a random calendar. The same is true of silver (p = 0.164).

That does not settle it on its own. An omnibus test can miss a single real month hidden inside eleven noisy ones, which is exactly why the next test targets January alone.

Test 2: is January specifically real?

The focused version of the same shuffle. In each of 20,000 shuffles, the mean of whichever month happens to land in January's slot is recorded, building a null distribution for "the mean of a randomly chosen month". The real January mean is then placed against it.

Gold January +2.66% permutation p = 0.0042 · 35 of 54 years positive
Silver January +2.97% permutation p = 0.0310
Best month ever +48.35% worst -6.34% · 8 years down 3%+

January survives; the rest of the table does not.

Gold's January mean of +2.66% sits at p = 0.0042 against the shuffled calendar — the strongest single-month result on this page, and one that survives the trimmed mean (+1.65%). 8 of the 54 years had January as the single best month of the year.

But January is also the strongest month for silver (+2.97%, p = 0.0310), and the two metals' twelve monthly-mean vectors correlate at 0.84. Gold and silver have completely different supply and demand stories. A pattern they share is much more likely a shared macro calendar — Chinese New Year demand, month-end and fund-reporting flows — than a property of gold itself. Whatever the cause, it is not a gold-specific edge.

And it is not free. In 8 of the 54 years, January was down 3% or more, with a worst case of -6.34%. A month that is up 2.7% on average and down 6.3% in the worst year is not a low-risk trade.

Test 3: would you have known January in advance?

For each calendar year, buy gold once, in the calendar month with the highest mean monthly return over all preceding full years (minimum 30 years of history), and hold that month. The choice uses only information available before the year starts. The rule is applied with no look-ahead: the month chosen for year y is chosen from years before y only.

Approach Years Total CAGR Max drawdown Positive years
Seasonal rule, out of sample 24 (2002–2025) +84.38% +2.582% -4.07% 20 / 24
Buy and hold, same window 24 (2002–2025) +1,461.23% +12.132% -39.28% —

The seasonal rule did something real: it took 20 positive years out of 24 and cut the drawdown from -39.28% to -4.07%. And it made far less money than doing nothing — +2.58% a year against +12.13% for holding gold.

That is the honest shape of a seasonal effect: it is a risk overlay, not a return overlay. Being in gold only in the historically strongest month throws away most of the compounding to avoid a few of the drawdowns.

Two things the backtest exposes that the month table hides.

  • The rule degenerates. The expanding-window selection picks January in 24 of 24 years. There is no rotating seasonal trade to find — only one month that has historically been the strongest.
  • The calendar has to be wrong somewhere. Eleven rotated calendars are run as placebos: each one knows a real pattern but applies it to the wrong months. Their CAGRs run from -0.238% to 1.921%, and the real calendar comes out on top at 2.582% (rank 12 of 12). So the calendar signal beats every wrong calendar — but note the scale of the wrong calendars themselves: a rule that is systematically off by one month still returned 1.921% a year, against -0.238% for the worst of them. The whole placebo range is narrow because the strategy holds gold only one month a year; the ranking is what carries the information, not the absolute levels.

Illustrative only. The monthly AVERAGE price is not an executable entry price, no cost of any kind is included, and the whole sample is one market with a changing regime (fixed $35 gold before 1971, free prices after). The point of the table is to show how much of an apparent edge survives being tested out of sample.

Test 4: does the winning month stay the winning month?

If a calendar effect is real, it should repeat in overlapping windows. These are 30-year rolling windows over the post-peg sample.

Window Best month Its mean Worst month Its mean Months with positive mean
1972-2001 January +2.68% November -0.62% 9 / 12
1977-2006 January +2.55% March -1.17% 8 / 12
1982-2011 September +1.53% March -0.74% 10 / 12
1987-2016 January +0.99% June -0.48% 9 / 12
1992-2021 January +1.54% March -0.43% 8 / 12

January wins 4 of the 5 windows and September wins 1. The best month's mean never falls below 0.99% and never rises above 2.68% — which is the practical argument for the January effect, and also the reason it is worth far less than the headlines suggest: it is a tilt, not an edge.

Test 5: the silver control

Silver comes from the same workbook column family and the same calculation. If the calendar pattern is a gold phenomenon, silver should not show it. It does.

Month Gold mean Silver mean Gold bootstrap CI Silver bootstrap CI
January +2.66% +2.97% 0.85 to 5.06 0.53 to 6.33
February +1.39% +3.27% 0.17 to 2.69 1.03 to 5.96
March +0.08% -0.75% -1.26 to 1.39 -2.92 to 1.32
April +0.55% +0.90% -0.47 to 1.60 -1.68 to 3.32
May +0.62% +0.04% -0.44 to 1.76 -1.89 to 1.93
June +0.40% -0.65% -0.81 to 1.77 -2.42 to 1.15
July +0.13% +0.41% -0.94 to 1.21 -1.35 to 2.20
August +0.90% +0.77% -0.20 to 2.10 -0.95 to 2.62
September +1.47% +2.06% 0.09 to 2.97 -0.39 to 4.89
October +1.06% +0.76% 0.04 to 2.19 -1.26 to 2.75
November +0.05% -0.08% -1.09 to 1.32 -1.44 to 1.31
December +0.65% +1.13% -0.54 to 1.87 -0.78 to 3.30

The two monthly-mean vectors correlate at 0.84, and the months whose bootstrap interval excludes zero for both metals are January and February. Highlighted rows are those.

Silver is the control. If the same calendar months stand out for both metals, the pattern is more likely a shared macro calendar (month-end flows, fund reporting) than something specific to gold.

What this dataset supports, and what it does not

Supported: gold's January is genuinely stronger than a random month (+2.66% average, permutation p = 0.0042, surviving a trimmed mean and eleven placebo rotations). The seasonal tilt is real enough to reduce drawdown if used as a risk overlay. And the same tilt is visible in silver, which points to a shared macro calendar rather than a gold-specific supply story.

Not supported: a month table that is distinguishable from a shuffled calendar as a whole (p = 0.207); any claim that November or December is reliably the worst month (November averages +0.05% and March is the actual low point, and neither has a negative mean); or a seasonal strategy as a way to make money (+2.58% a year against +12.13% for holding). And no month passes the Bonferroni-corrected threshold for a 12-test family at the 5% level (0 of 12), which is the correct statistical reading of a table where four rows happen to have intervals that exclude zero.

The practical takeaway is narrow: if you run a gold position, the historical case for reducing it in the weakest months is weak, and the case for adding to it in January is real but small and shared with another metal. Neither is a reason to change a risk framework.

Frequently asked questions

Is gold really strongest in January?

January has the highest mean of the twelve months: 2.66% per month over 54 post-peg years, positive in 64.8% of them (35 of 54). A focused permutation test that shuffles the twelve months within each year gives p = 0.00415. So yes, January stands out on its own terms — but read the next answers before using it, because the omnibus test and the control metal both complicate it.

Does the whole month table survive multiple-testing correction?

No. Testing all 12 monthly means at once, a Bonferroni-corrected alpha of 0.00417 requires |t| > 2.807, and 0 months pass. The omnibus within-year permutation test gives R² = 0.0208 with p = 0.207 over 2,000 permutations, which does not reject. Under a pure null you would expect about 0.6 of the twelve months to look significant by chance. The month table as a whole is not distinguishable from noise; the focused January result is a separate, narrower test and it does hold.

If January is real, why not trade it?

Because the out-of-sample test kills the trade even if the month is real. Using an expanding window that picks the best month from all prior history (minimum 30 years), the strategy returned 2.582% a year over 24 years with a -4.07% drawdown, against buy-and-hold at 12.132% a year and -39.28% drawdown — roughly 4.7 times worse. The rule also degenerates: it picked January in 24 of 24 years, so there is no rotating seasonal trade to find. The eleven placebo rotations that apply the same knowledge to the wrong months span -0.238% to 1.921%, and the real calendar beats every one of them at 2.582% — it ranks highest of the twelve. But the whole range is narrow because the strategy holds gold for one month a year, so the ranking carries the information, not the level.

Is the January effect specific to gold?

No, and that is the most important finding on the page. Silver's January mean is 2.97% with its own permutation p = 0.03095, and the two monthly-mean vectors correlate at 0.8441. Two metals with different supply-demand stories share the same calendar pattern, which points to a shared macro calendar — Chinese New Year demand, month-end and fund-reporting flows — rather than a property of gold. Whatever the cause, it is not a gold-specific edge.

Is the seasonal pattern stable over time?

No. Across 5 rolling 30-year windows the best month was January in 4 of them and September in 1, and the best month's mean drops from 2.68% in the earliest window to 1.54% in the latest. A real effect should repeat; this one changes identity. The sample also excludes the fixed $35/oz years (US convertibility was suspended in August 1971), so the primary table uses 54 complete years from 1972-2025, with the 65-year full-sample table shown alongside so the effect of that choice is visible.

What was the arithmetic versus geometric return over this sample?

The gap is large and it matters. The arithmetic mean monthly return was 0.6957%, which annualises to 8.68% — but the path actually delivered 7.4% compounded, because the monthly returns are negatively skewed and the compounding drag is real. Annualised volatility was 15.84% and only 43.4% of months were positive. Monthly returns are also serially correlated (AR(1) = 0.2688), which is exactly why a naive t-test on a single month is not enough and why the bootstrap and permutation tests on this page block by year.

Can I use this to time entries into gold?

Not as a rule. The monthly average price is not an executable entry price, no transaction, financing or storage cost is included anywhere, and the whole sample is one market with a changing regime. What the page supports is narrower: the calendar month by itself carries little information, and the one month that stands out is shared with silver, so it is a macro calendar effect rather than a tradable gold edge. If you want to act on historical structure, the drawdown and position-sizing tools use the price history directly, which is a different and more defensible starting point.

Sources

  • World Bank Commodity Price Data (The Pink Sheet), CMO-Historical-Data-Monthly.xlsx — Monthly Prices sheet, columns 70 (Gold) and 72 (Silver), updated November 03, 2026. https://www.worldbank.org/en/research/commodity-markets
  • Analysis script: scripts/gold_seasonality_analysis.py. Independent recomputation from the published CSV: scripts/verify_gold_seasonality.py.
  • Method references for the testing approach used here: block bootstrap (Künsch, 1989), permutation testing within blocks, and the multiple-comparisons correction for a family of 12 tests.
  • For the broader correlation caveats that apply to any month-level reading of this data, see Gold vs. real yields and the dollar and Gold price history.
  • The same calendar question at daily resolution, where the weekday is actually defined: Gold day of week.

Data accessed 2026-11-09. Prices are nominal US dollars and include no carry, financing or storage cost. Nothing on this page is investment advice or a forecast.

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