Research · Dataset

Gold falls on Mondays. It dumps on Fridays. Right?

Both claims are repeated constantly in retail gold trading, and neither survives a proper test. Across 14,723 daily benchmark fixes spanning 1968-04-01 to 2026-11-10, the difference between the best and the worst weekday is 0.0490 percentage points a day — and 0 of the five weekdays clear a Bonferroni threshold, an in-year permutation test and a Newey-West autocorrelation-robust test at the same time.

1968-04-01 to 2026-11-10 · 14,723 daily benchmark fixes · 14,722 daily returns · overall annualised volatility 19.31%

The answer, in four numbers

Weekdays surviving every test 0 of 5 Bonferroni + permutation + Newey-West
Best-to-worst weekday spread 0.0490 pp per day, Wednesday against Tuesday
Sell-Friday rule, after costs 0.82x buy-and-hold over the same window: 109.3x
Turn-of-month edge -0.0301 pp against a length-matched control, p = 0.787

The direction of the weekday means is not the point, and neither is the volatility spread. The point is that every one of them is the size of noise: daily gold moves 19.3% annualised, so a weekday "edge" worth 0.0490 points a day is a rounding error on a market whose own daily movement is roughly 0.85%.

What the claim actually is, and what the data says

The convention matters, so it is stated first. Each fix is stamped with the weekday of its own date, and returns are simple returns between consecutive fixes. The return stamped to a weekday is therefore the move into that weekday's fix from the previous fix, whatever day that was. Monday's bucket is the Monday-over-Friday move, which contains the weekend gap. There are five buckets, not seven: the series contains no Saturday or Sunday fixes.

Weekday Mean daily return Annualised Mean absolute move Bootstrap 95% CI t-test p Permutation p Newey-West p Verdict
Monday +0.0335% +8.45% 0.849% -0.0143 to 0.0807 0.180 0.604 0.791 not significant
Tuesday +0.0169% +4.26% 0.778% -0.0263 to 0.0624 0.437 0.870 0.266 not significant
Wednesday +0.0659% +16.61% 0.780% 0.0228 to 0.1108 0.003 0.089 0.179 mixed
Thursday +0.0175% +4.41% 0.774% -0.0246 to 0.0570 0.422 0.866 0.273 not significant
Friday +0.0621% +15.64% 0.773% 0.0201 to 0.1079 0.004 0.128 0.234 mixed

2,763 to 2,929 observations per bucket. A single weekday is significant at p<0.05 by chance one time in twenty, and testing five is looking for the best one — which is why each bucket carries a Bonferroni threshold of 0.01, an in-year permutation test that relabels observations inside each calendar year (preserving the autocorrelation a global shuffle destroys), and a Newey-West HAC test with five lags. A weekday is only called significant when all three agree.

Read the table honestly: the plain t-test calls Wednesday and Friday significant, and that is exactly what you would expect to happen roughly two times in five by chance when you look at five buckets. Once the multiplicity is corrected and the autocorrelation is handled, nothing survives. The permutation test, which is the most honest of the three because it preserves the year-level structure, rejects every weekday.

The Monday selloff and the weekend gap

The Monday-over-Friday move averages 0.0335% against 0.0335% for the Tuesday-to-Thursday moves — a difference of 0.0001 points with a p-value of 0.998. There is no Monday selloff in the data.

What does show up is dispersion. The average absolute move is largest on Monday at 0.849% and smallest on Friday at 0.773%, a spread of 0.076 points. The weekend gap is where gold moves most, but it moves in both directions, and a wider spread of outcomes is not an exploitable edge — it is a larger loss waiting for anyone who has to be right.

Why "the weekend gap is bigger" is not an edge

A bigger average absolute move means more variance, not more predictability. If the Monday move were larger and tilted in one direction, the sell-Friday rule would have something to work with. The second condition is the one that fails: the Friday-to-Monday leg averages 0.0632% a week, and the sign is not stable enough to trade.

The rule that does not work: sell Friday, buy Monday

This is the claim in its tradable form, so it is tested in its tradable form. The rule means being flat over the weekend and long from the Monday fix to the Friday fix. The rule's weekly return is therefore the Monday-to-Friday leg only, and the Friday-to-Monday leg it gives up is forgone. It runs over 2,741 weeks, from 1968-04-08 to 2026-11-06.

Version Mean weekly Annualised Bootstrap 95% CI Final multiple Note
Rule, gross of costs +0.1471% +7.65% 0.0636 to 0.2274 27.40x Monday fix to Friday fix, 2,741 weeks
Rule, $0.30/oz round turn +0.0192% +1.00% -0.0672 to 0.1060 0.82x one round turn per week
Rule, 10 bps round turn -0.0530% -2.75% -0.1328 to 0.0296 0.11x proportional cost variant
Rule, 1990 onwards, gross +0.1204% +6.26% 0.0324 to 0.2054 5.64x 1,717 weeks
Rule, 1990 onwards, $0.30/oz +0.0692% +3.60% -0.0216 to 0.1607 2.34x cost applied on top of a $4,000/oz price
Buy-and-hold, same window — 9.31% CAGR — 109.3x from 1968-04-08 to 2026-11-06

Gross of costs the rule earns 0.1471% a week, 7.65% annualised, with a bootstrap interval of 0.0636% to 0.2274% and a p-value of <0.001. That is a real positive number, and it is the trap: the long side of the week is positive simply because gold drifts up over fifty-eight years. Buy-and-hold over the identical window returned 109.3x, a 9.31% annualised rate, against 27.4x for the rule.

Add the cost and the rule disappears. A flat $0.30 per ounce round turn, charged once a week, takes it to 0.0192% a week and 0.82x over the sample — a loss. A proportional 10 basis point round turn, which is what a modern spread actually looks like, takes it to -0.0530% a week and 0.11x.

And the cost model is not neutral in time, which is why the modern-era numbers are given separately. The same $0.30 is 0.863% of the cheapest price in the sample and 0.0074% of the last price. Restricted to 1990 onwards the rule returns 0.1204% a week gross and 0.0692% after the flat cost, ending at 2.34x against 10.0x for buy-and-hold over the same modern window. Under every cost model tested, the rule loses to simply holding.

Turn of the month

The turn-of-month effect is the last of the three calendar stories, and it is the cleanest test on this page because the comparison can be made length-matched: the window from the last trading day of a month through the third trading day of the next, against a window of the same length starting at the beginning of the same month.

Window Observations Mean Bootstrap 95% CI Positive
Turn of the month 703 +0.1549% -0.0350 to 0.3314 51.2%
Length-matched control 703 +0.1850% — 52.9%
Difference — -0.0301 pp p = 0.787 —

The turn-of-month window is, if anything, marginally worse than a random window of the same length: -0.0301 points with a p-value of 0.787. The turn of the month is not special in gold.

Decade by decade: the best weekday keeps changing

The single clearest sign that a weekday effect is noise rather than structure is that the best weekday is not the same weekday twice.

Decade Mon Tue Wed Thu Fri Best
1960s +0.0046% +0.0080% +0.0095% -0.0536% -0.0359% Wed
1970s +0.1930% +0.0695% +0.1777% +0.0465% +0.0994% Mon
1980s -0.1130% -0.0353% +0.1261% -0.0191% +0.0509% Wed
1990s -0.0675% -0.0085% -0.0240% +0.0019% +0.0459% Fri
2000s +0.0934% -0.0299% +0.0473% +0.0747% +0.1168% Fri
2010s +0.0812% +0.0289% -0.0139% -0.0104% +0.0102% Mon
2020s +0.0101% +0.1062% +0.1035% +0.0262% +0.0677% Tue

The best weekday by decade was Wednesday, Monday, Wednesday, Friday, Friday, Monday, Tuesday. Monday and Friday also flip sign relative to each other in some decades — Monday's mean is consistent in sign across all 7 decades: no; Friday's: no. A genuine effect would not need a new winner every ten years.

What this dataset supports, and what it does not

Supported: that gold returns do not depend on the day of the week in any way that survives a multiplicity correction, an autocorrelation-robust test and a decade-by-decade stability check. It is also supported that the retail rule built on the claim loses money once a realistic round-turn cost is charged, and that the apparent gross edge of the rule is simply the long side of a drifting market minus the weekend leg it gives up.

Not supported: any claim that this is a forecast. It is one historical path through one benchmark, and a rule that fails on it has not been proven to fail on every path. Nor is anything here evidence about intraday behaviour — a benchmark fix is an auction price, and the path inside a day is invisible in this series.

The practical takeaway is narrow: a weekday is not a reason to enter or exit a gold position, and a rule built on one is paying costs to capture a share of a drift that simply holding captures for less. If you are sizing a position on a view about gold, size it on the view and the risk, not on the calendar.

Frequently asked questions

Is there a day of the week on which gold reliably falls?

No. Across 14,722 daily benchmark fixes from 1968-04-01 to 2026-11-10, the five weekday buckets have means between 0.0169% and 0.0659% per day — a total spread of 0.049 percentage points. The largest weekday mean is Wednesday, but 2 of the five weekdays clear a Bonferroni threshold of 0.01, 0 clear an in-year permutation test, and 0 clear a Newey-West HAC test once autocorrelation is handled. The number of weekdays that pass all three at once is 0.

What about the "Monday selloff" — is the weekend gap bigger?

The Monday-over-Friday move averages 0.0335% against 0.0335% for the Tuesday-to-Thursday moves, a difference of 0.0001 points with a p-value of 0.998. The gap is real in size but not distinguishable from noise. What does stand out is dispersion rather than direction: the average absolute move on a Monday is 0.849% against 0.773% on Friday, a spread of 0.076 points, and Monday carries the largest average absolute move of the five.

Does the "sell Friday, buy Monday" rule make money?

Gross of costs the rule returns 0.1471% a week, or 7.65% annualised over 2,741 weeks, with a bootstrap interval of 0.0636% to 0.2274%. That is the long side of the week, from the Monday fix to the Friday fix, with the weekend left flat. But the weekend is not flat for free: the Friday-to-Monday leg the rule gives up averages 0.0632% a week. Buy-and-hold over the identical window returned 109.3x against 27.4x for the rule. Once a $0.30 per ounce round turn is charged, the rule compounds to 0.82x — it loses money.

What happens if the cost assumption is changed?

The cost model matters more than the effect being hunted, so it is reported three ways. With a flat $0.30 per ounce round turn the rule returns 0.0192% a week and ends at 0.82x. With a proportional 10 basis point round turn it returns -0.053% a week and ends at 0.11x. Restricted to 1990 onwards, where a flat dollar cost is 0.1% or less of the price, the rule returns 0.1204% a week gross and 0.0692% after the flat cost, ending at 2.34x against 10x for buy-and-hold. The conclusion does not depend on which cost model is used: the rule does not clear buy-and-hold.

Is gold more volatile at the turn of the month?

No measurable excess. The window from the last trading day of a month through the third trading day of the next month averages 0.1549%, against 0.185% for a length-matched control window in the same month — a difference of -0.0301 points with a p-value of 0.787. The turn-of-month window is positive 51.2% of the time and the control window 52.9% of the time, over 703 month turns.

Has the same weekday ever been the best one?

No. Across 7 decades the best weekday was Wednesday, Monday, Wednesday, Friday, Friday, Monday, Tuesday. The identity of the best weekday changes from decade to decade, and Monday and Friday change sign relative to each other in some decades. That instability is the clearest single sign that the weekday means are noise rather than a stable structure: a real effect would not need a new winner every ten years.

How big are the weekday effects compared with daily noise?

The overall annualised volatility of daily gold returns is 19.31%. The largest weekday mean is 0.0659% per day, and the whole spread between the best and worst weekday is 0.049 percentage points. The weekday structure is a rounding error next to the day-to-day movement of the price itself.

What are the limits of this analysis?

Four. First, a benchmark fix is an auction price, not a tradable bid/ask, so the intraday path inside a day is invisible and a real account sees more dispersion than this series shows. Second, no spread, slippage, commission, financing, swap or tax is modelled beyond the stated round turn, so the cost-aware results are optimistic in absolute terms. Third, the flat $0.30 per ounce cost is not neutral in time — about 0.863% of the cheapest price in the sample ($34.75 per ounce) and about 0.0074% of the last price ($4050.9 per ounce) — which is why the results are also given for 1990 onwards and with a proportional cost. Fourth, the series contains no Saturday or Sunday fixes, so "the weekend" is represented only by the Friday-to-Monday gap. There is no live account behind any number here: no broker statement, no server-side execution, no requotes and no swaps.

Sources and method

  • LBMA Gold Price PM benchmark fixes, 1968-04-01 to 2026-11-10 — 14,723 daily observations, built by scripts/build_daily_gold.py. lbma_csv and wgc_fsapi feeds spliced with a verified seam.
  • Analysis script: scripts/gold_day_of_week_analysis.py. Independent recomputation from the raw daily file with a separate numpy implementation: scripts/verify_gold_day_of_week.py (104 checks, 0 failures).
  • Method: Five weekdays are tested, so a single weekday is significant at p<0.05 one time in twenty by chance. Each weekday carries a Bonferroni threshold (alpha = 0.01), an in-year permutation test that relabels observations within each calendar year, a Newey-West HAC t test (lag 5) of a weekday dummy on the daily return, and a moving-block bootstrap confidence interval. A weekday is called significant only when the conservative tests agree.
  • Rule definition: The retail rule is "sell Friday, buy Monday", which means being flat over the weekend and long from the Monday fix to the Friday fix. The rule's weekly return is therefore the Monday-to-Friday leg only; the Friday-to-Monday leg is flat and earns nothing. Buy-and-hold is measured over the identical calendar window.
  • For the monthly counterpart of this test, see Gold and the calendar. For why a monthly series cannot answer a question about a single day, see the intra-month drawdown dataset and the sampling-frequency comparison.

Prices are nominal US dollars per troy ounce 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.