With the test that usually gets skipped
Weekday patterns in traditional markets come from structure: exchanges close, settlement takes days, funds report on schedules, and news is released at set times. Crypto trades continuously with none of that machinery, so there is little mechanism for a day-of-week effect to arise from in the first place.
The statistics reflect that. Sample size is not the constraint — nine years gives roughly 470 observations per weekday — the problem is that a plausible effect of a tenth of a percent has to be detected inside daily swings of three to four percent. Even a large sample cannot separate those, which is why a t-statistic can sit near zero despite hundreds of observations.
Two further cautions apply to any weekday table, including this one. Seven simultaneous tests mean roughly a third of a false positive is expected by chance, so a single flagged day proves nothing. And because crypto returns are fat-tailed, the t-test overstates significance: comparing each mean against its median shows immediately when a few violent days are producing the result.
"Buy on Monday, sell on Friday" is repeated often enough to feel established, and it is easy to check. Nine years of Bitcoin gives roughly 470 observations of each weekday, which sounds like plenty. The difficulty is scale: an effect of a tenth of a percent has to be detected inside daily swings of three to four percent, and no amount of counting separates those.
That is why this page reports a t-statistic next to each average rather than a ranking. It also reports the median beside the mean, because the two disagreeing is the clearest sign that a handful of violent days are producing the result. A weekday whose mean is several times its median has an average, not a tendency — and since crypto returns are fat-tailed, the t-test itself overstates significance in exactly those cases.
The last caution is arithmetic rather than statistical. Seven weekdays tested at a 5% threshold produce about a third of a false positive by chance, so one flagged day is unremarkable and even two deserve scrutiny before belief. Crypto also lacks the machinery that creates weekday effects elsewhere — no closes, no settlement cycle, no scheduled reporting — so there is little for such an effect to arise from in the first place.
⚠️ Not investment advice. Weekday averages describe the past and are dwarfed by daily volatility even where they look favourable. Days are assigned by UTC close, so a different timezone convention will shift the table. All decisions and risks are your own.
The table shows what each weekday has averaged, but the honest answer is that any effect is far smaller than the daily noise it sits inside. A tenth of a percent tendency cannot be separated from three to four percent daily swings, however many observations you have.
Because sample size is not the constraint — effect size is. Nine years gives roughly 470 observations of each weekday, and the t-statistic can still sit near zero because the signal being looked for is two orders of magnitude smaller than the variation around it.
Because when they disagree, the mean is being carried by outliers. A weekday whose mean is several times its median has one or two enormous days in it rather than a consistent tendency, and that is precisely the case where a t-test overstates significance on fat-tailed returns.
No. Seven weekdays tested at a 5% threshold produce about a third of a false positive by chance, so a single flagged day is unremarkable. Even two deserve scrutiny, especially if the mean-versus-median check suggests outliers are responsible.
Because the mechanisms are absent. Weekday patterns in traditional markets come from exchange closes, multi-day settlement, scheduled fund reporting and timed news releases. Crypto trades continuously with none of that structure.
Yes. Days here are assigned by UTC close, which is the convention Binance candles use. A table built on a local timezone would split returns differently and could shift which weekday looks strongest.