Monthly Returns Heatmap

For every (year, month) pair where both endpoints have a daily close on record, the cell is ret(y, m) = close(last day of m, y) / close(last day of m−1, y) − 1. The current month uses the most recent close in place of an end-of-month close, so the latest row always carries a live month-to-date cell. The annual column is December close over prior-December close minus one, falling back to compounding the defined months when December is unavailable — the first and current years.

The colour scale clamps at ±40% so the early-2010s moonshots don't bleach the rest of the grid; the printed number in each cell is always the raw value. Beyond the grid, the page computes per-calendar-month sample size, mean, median, standard deviation, and a t-statistic against zero: t = mean / (stdev / √n). That last number is the whole point. A column average looks like a season; the t-statistic asks whether the season is bigger than the month-to-month scatter. With 17 years on record — ten to sixteen observations per calendar month — the two-tailed p < 0.05 critical value sits near 2.14. The full derivation, clamp and month-end conventions live on the methodology page.

Sorted by average return — the line the seasonality crowd quotes — the biggest calendar months are November (+34.7%, n=16, t=1.22); April (+32.1%, n=16, t=1.50); October (+26.4%, n=16, t=1.92); May (+17.8%, n=16, t=1.59). Read the t-column, not the mean column. The four largest averages on the whole record carry t-statistics that never reach 2.14, the threshold a sample this size needs to call a mean distinguishable from zero. Their standard deviations run from +114.1% to +44.9% — several times the size of the average itself. That is the signature of noise, not season.

The table below scores all twelve. A tone badge fires only when |t-stat| ≥ 2.14; 0 of the twelve qualify. October is the closest miss — the popular “Uptober” trade, statistically a coin-flip dressed in a strong record.

The heatmap reads three ways. By row you see how a single year unfolded — whether the bear was front- or back-loaded, whether the rally came in June or November. By column you see how a calendar month behaved across cycles, and how often the colour flips. By the bottom row you get the per-month average, which is the line the seasonality crowd quotes — and the line the significance table above exists to caveat. The grid is a cycle-context lens, best read alongside the halving-cycle overlay, which explains more of the green clusters than the calendar ever does: the melt-up months tend to be the ones twelve-to-eighteen months past a halving, not the ones called October.

Pulling the annual total alongside each cycle's defining year shows the same data in a shape the monthly grid hides. 2013 and 2017 were both order-of- magnitude bull years; 2014 and 2018 were the symmetric capitulation years; 2020 compounded the Covid flush and the recovery into a positive annual; 2021 stalled on the year despite a November cycle top. The best- and worst-month columns expose each year's volatility regime: a +50% best month sitting next to a −30% worst month is a normal Bitcoin year, not a calm one.

The seasonality claims on this chart fail in a specific, datable way: the column averages are hostage to one or two prints, and the live record is already eroding the patterns people quote most.

October — the strongest column — is one outlier wide. October 2010 returned +211%, the largest single contribution to the famous +26.4% average. Strip that one print from the sixteen and the “Uptober” edge falls by nearly half. Then October 2025 closed −5.3% — the first red October since 2018 — directly into the year the trade was most crowded. The historical tilt is real (11 of 16 green); the forward edge is a coin-flip with a good story, exactly what t = 1.92 says it is.

September weakness was the cleanest pattern — and just broke. September ran red in 2017, 2018, 2019, 2020, 2021, 2022 and 2023 — seven losing Septembers in a row, the most reliable calendar signal Bitcoin ever produced. Then September 2024 closed +11.0% — the first green September since 2016 — and September 2025 +5.1%, snapping the streak the moment it became consensus. At t = -1.14, the test never endorsed it; the tape confirmed the test.

November's average is a single 2013 print. November averages +34.7% with a standard deviation of +114.1% — the widest of any month — almost entirely because November 2013 returned +453%, the best Bitcoin month on record. Outside that one print, recent Novembers split both ways: +34.7% in 2024, but −16.1% in 2025, −20.3% in 2022, −32.4% in 2018. The mean describes a month that effectively no longer exists.

Daily closes only, and month-end snapshots drift. A month that spiked intra-month but closed flat reads as flat; a month that wicked down but recovered by the last close reads as green. And different providers snapshot month-end at slightly different times, so a handful of historical months can read one-to-ten percentage points off the figures other sites publish — direction always correct, magnitude occasionally not. The merge provenance is on the data sources page.

The data-mining warning compounds the sample problem. The canonical critique of calendar effects in markets — Sullivan, Timmermann & White (2001) — shows that calendar tendencies which look strong in isolation largely vanish once you account for how many patterns were tested. With Bitcoin's sample a fraction of an equity index's, the warning is stronger, not weaker.

If you accumulate on a schedule, ignore the calendar entirely. The per-month signal is far too weak relative to monthly volatility for a month-of-year rule to add value, and a steady weekly buy lands at roughly the long-run average anyway. The path-dependent variance across years dwarfs anything the bottom row can offer. Rolling CAGR is the honest horizon: it shows where the monthly noise actually averages out.

If you trade the cycle, read the grid by row for shape — two- stage vs single-stage advances, front- vs back-loaded declines — and let cycle position, not the month name, set the bias. The halving-cycle overlay and drawdown carry the context the heatmap can't: a green October in year two post-halving and a green October at a cycle top are not the same trade, even though the grid colours them identically.