Shown with the sample size that usually gets left out
Named seasonal patterns get repeated because they are memorable, and the arithmetic that would deflate them is rarely shown. Nine Octobers is nine observations. Seven of nine going up sounds convincing until you work out how often a fair coin does that — often enough that it needs no explanation at all.
The problem compounds because twelve months are examined together. Testing twelve things at a 5% threshold produces about 0.6 apparent hits by chance, so finding one striking month in a calendar year is the expected outcome of looking, not a discovery. That is why the p column sits next to every return here, and why this page draws no conclusion from the table.
This page groups a coin's daily closes into calendar months and reports the median return for each, along with how many years were up. That much is standard. What is usually missing is the count beside it: a monthly seasonality figure does not rest on thousands of days of data, it rests on the number of times that month has occurred. Bitcoin's Binance history covers about nine years, so every monthly figure here is built from roughly nine numbers.
Nine is small enough that ordinary randomness produces striking-looking patterns. Seven up years out of nine reads as a strong tendency and happens by chance often enough to need no explanation, which is why a p-value sits beside each month — the probability of seeing a split that lopsided, or more, from a fair coin. Most named crypto seasonal patterns do not survive that column.
The second problem is that twelve months are examined at once. Testing twelve things against a 5% threshold yields about 0.6 apparent hits from randomness alone, so finding one dramatic month in the calendar is the expected result of looking rather than a finding. Altcoins are worse still: a coin listed three years ago has three observations per month, which is not seasonality but three coincidences. The table is shown so that thinness is visible, not so it can be traded.
⚠️ Not investment advice. Monthly groupings describe a handful of past observations per month and carry no forecast. Returns are computed from Binance daily closes and exclude any period before the coin listed there. All decisions and risks are your own.
The table shows what each calendar month has actually done, but the honest answer is that the sample is far too small to establish it. Bitcoin has around nine Octobers on Binance, so an October figure rests on nine numbers rather than on years of daily data.
Because a lopsided win rate looks convincing until you calculate how often chance produces it. Seven up years out of nine happens frequently with a fair coin. The p-value is the probability of a split that uneven or more, and most named crypto seasonal patterns do not survive it.
Testing twelve things at a 5% threshold produces about 0.6 apparent hits from randomness alone. So finding one dramatic-looking month somewhere in the calendar is the expected outcome of looking rather than a discovery, and a single flagged month should not be read as a pattern.
October does show the strongest median in Bitcoin’s Binance history, and the sample behind it is roughly nine years. Whether that clears the bar for evidence is exactly what the p column answers, and the page reports the number either way rather than promoting the name.
You can select any Binance coin, but the sample gets much thinner. A token listed three years ago has three observations per month, which is three coincidences rather than a seasonal effect. The page flags coins with too few years rather than presenting the table as meaningful.
From the first to the last daily close within that calendar month, using Binance data. The first and last months of a coin’s history are excluded because they are usually partial, which would otherwise distort those two months.