MEASURED · 16 SEPTEMBER 2026

Is gold seasonality real?

January averages +3.61% and rose in nine of the last twelve years. That is the strongest gold seasonality claim there is, and this article is going to argue against it using the data that appears to support it.

The chart everyone has seen

Search for gold seasonality and you get a bar chart of average returns by month. Some bars are taller. The taller ones get a narrative — January buying from Asian physical demand, a weak September, a year-end rally. The chart is usually arithmetically correct and almost never states two things: how many observations sit behind each bar, and whether the differences between them mean anything.

The answer to the first is twelve. Even with 3,901 daily bars going back to 2014, splitting by calendar month leaves about 12.6 observations per bar. Monthly gold returns have a standard deviation over 4%, so the standard error on a twelve-sample average is more than a full percentage point. Most of the height difference between those bars is measurement noise.

What the numbers actually are

151 complete months, 2014-01-14 to 2026-09-16. Partial months at each end are excluded — a half-month masquerading as a full one contaminates its calendar slot. Returns are logarithmic so they aggregate correctly.

Look at the SE column against the mean column. January’s +3.61% carries an error bar of ±1.22 points. Most months cannot be told apart from zero, let alone from each other.

Testing it properly

Rather than assume a bell curve — monthly returns are not normally distributed — the test here is a permutation. Shuffle the calendar labels across the 151 observed returns, recompute every monthly average, and repeat 20,000 times. That builds the distribution of seasonality patterns produced by data with no seasonality in it, and the real result can be compared against it.

On its own, January returns p = 0.006. By the conventional 5% threshold that is not just significant, it is strong. Published as-is it would make a compelling article.

It is also wrong, for a reason that has nothing to do with gold.

Twelve tests, not one

January was not chosen in advance. It was chosen by looking at twelve months and picking the most extreme. Running twelve tests at a 5% threshold means expecting 0.6 false positives from noise alone. Finding one is not evidence of a pattern — it is the textbook description of what a dataset with no pattern produces.

The honest question is therefore not “is January unusual?” but “is January more unusual than the most unusual month noise would hand us?” The permutation answers it directly, by recording the strongest monthly average in every single shuffle:

Gold return seasonality does not survive contact with its own sample size. Not in this data, over 12.6 years. Whether the effect exists and is simply too small to detect with twelve samples per month is a fair objection — but a pattern you cannot detect is also one you cannot trade.

The part that is real

Change the question from which way gold moves to how far it moves, and the statistics transform. Daily range is measured from every one of the 3,871 days in the sample — roughly 323 observations per calendar month rather than twelve.

March runs 42% wider than September, and the permutation test on that spread never once produced a gap that large across 20,000 shuffles. The same data that cannot tell you which month gold rises in tells you very clearly which months it moves in.

That has a practical consequence a direction forecast never delivers. A stop set at a fixed dollar distance is a far tighter stop in March than the identical number is in September, and a position sized without reference to volatility carries materially different risk in each. The fix is not a seasonal calendar — it is sizing from measured current volatility, which gets this right automatically and keeps working in a month that breaks the pattern. It is the same reasoning behind measuring how the gold spread and range move by hour.

Why this keeps happening

Seasonality is unusually good at generating false positives. The data is free, the twelve buckets are handed to you, and every bucket comes with a ready-made story — Indian wedding season, summer liquidity, year-end positioning. When a month looks strong there is always an explanation available, and an explanation feels like evidence.

The defence is deciding what counts as proof before looking. Here that meant fixing the correction for twelve simultaneous tests in advance, which is what turned a publishable January result into a null one. We apply the same rule to the trading system itself: an earlier version scaled its risk-reward ratio by market regime, validated well in-sample, and was removed after out-of-sample testing showed it was fitting noise. Removing a feature improved results. See also how many trades a track record needs before it means anything, which is the same problem in a different costume.

Limits

One symbol, one broker, 2014-01-14 to 2026-09-16. Gold spent much of this window in a strong bull trend, which lifts most months and could mask a genuine seasonal effect underneath. A longer history would test more — though note it would also span periods when gold traded under entirely different monetary regimes, so more years is not straightforwardly more evidence.

The volatility result is much better powered than the return result, but it is still one instrument. And none of this says seasonality is impossible — only that in this sample, the return pattern cannot be distinguished from chance once the multiple-testing problem is handled honestly. The full numbers for every month are published alongside our other studies, including the columns that make this argument falsifiable.

Common questions