Back to ResearchTopic: On-Chain Data, Metrics and Market Cycles

Bitcoin Rolling Holding-Period Returns, 1-5 Years

C

Research Desk • Organizational attribution

Source Standard
6 source notes
Last Reviewed
No edits since publication
Not Advice
Educational research, not a recommendation to buy or sell

No 4-year bitcoin holding window that has finished ended below its start price: 0 of 4,491 daily starts between July 18, 2010 and November 2, 2022, on Coin Metrics' daily reference price. The same series has 27.9% of 1-year windows below start, 15.8% of 2-year windows, 33 of 4,856 three-year windows and 9 of 4,126 five-year windows. Pulled November 3, 2026.

"Never lost money over four years" is therefore true of this record, and it is narrower than it sounds. It holds at exactly four years and fails at three and at five: three-year windows starting between November 22, 2013 and January 15, 2014 or in February 2021, and five-year windows starting between December 7, 2017 and January 6, 2018, ended below their start. It also rests on about four independent windows, not 4,491. And CoinDesk's four-year compound annual growth rate of -6% in March 2025 belongs to the ether-to-bitcoin ratio in that article, not to bitcoin priced in dollars. Bitcoin's own four-year rate in the same article is +8%.

Everything below is descriptive: completed windows from one price series, in dollars, before fees and taxes. The result with the most weight is not a percentage but a count. The record holds about four separate four-year outcomes, and the thousands of daily windows are overlapping views of those few.

What Counts as a Holding Window Here

Every figure comes from Coin Metrics' bitcoin PriceUSD metric, the same series behind the year-by-year price history. Coin Metrics' metric catalog defines it as "The fixed closing price of the asset as of 00:00 UTC the following day (i.e., midnight UTC of the current day) denominated in USD", generated by its fixing and reference-rate service. The pull used here has 5,952 daily values from July 18, 2010 to November 2, 2026 with no missing days. The price-history page was pulled on October 28 and stops at October 27; this pull adds six more days. Its closes on October 27, 2026, December 31, 2025 and the record dates match to the cent, so a reader moving between the two pages is looking at one series.

A holding window is defined by three rules:

Start is a daily close on a date the rule selects: every day, the last day of each month, January 1, December 31, or a day that set a new record close.
End is the close on the same calendar date N years later, for N of 1, 2, 3, 4 and 5. A start on February 29 ends on February 28 when the target year has no leap day. A window counts only if its end date is inside the data, so windows that have not finished are left out and listed separately near the end.
Below start means the end close is strictly lower than the start close. Returns are plain price changes in US dollars: no fees, no tax, no inflation adjustment, no income.

One consequence of those rules is the number of windows. A 1-year rule has 5,587 daily windows, because starts run from July 18, 2010 to November 2, 2025. A 5-year rule has 4,126, because starts stop at November 2, 2021. Longer horizons have fewer windows, and each window overlaps its neighbours almost completely. That overlap is the main reason the headline shares need care, and a later section counts it.

The price-history page measures a calendar year from the last close of the previous December, so a one-year window that starts on a December 31 reproduces its yearly column exactly: 2014 at -56.0%, 2018 at -73.5%, 2022 at -64.4% and 2025 at -6.3%. A window that starts on January 1 differs, because the price moves in a day; the 2018 window from January 1, 2018 to January 1, 2019 is -71.7%. Both are correct for the question each asks. The December 31 version is used wherever a figure has to agree with that table.

The Distribution for One to Five Years

Table 1 puts all daily starts together, using every complete window in the series.

Table 1. Bitcoin holding periods, all daily starts

Holding periodComplete windowsEnded below startShare belowMedian changeWorst window (start to end)Best window, starts since 2013 (start to end)
1 year5,5871,55827.9%+95.2%2017-12-16 to 2018-12-16: $19,640.51 to $3,195.41, -83.7%2013-01-06 to 2014-01-06: $13.50 to $914.46, +6,672%
2 years5,22282415.8%+263.7%2013-12-04 to 2015-12-04: $1,134.93 to $363.24, -68.0%2015-12-16 to 2017-12-16: $454.14 to $19,640.51, +4,225%
3 years4,856330.7%+474.1%2013-11-29 to 2016-11-29: $1,128.73 to $731.65, -35.2%2015-01-14 to 2018-01-14: $175.64 to $13,681.07, +7,689%
4 years4,49100.0%+1,289%2021-04-13 to 2025-04-13: $63,445.64 to $83,568.95, +31.7%2013-01-04 to 2017-01-04: $13.45 to $1,134.50, +8,334%
5 years4,12690.2%+2,378%2017-12-16 to 2022-12-16: $19,640.51 to $16,606.11, -15.4%2013-01-06 to 2018-01-06: $13.50 to $17,103.59, +126,551%

Calculated by CryptosEyes from Coin Metrics PriceUSD daily closes, 2010-07-18 to 2026-11-02, pulled November 3, 2026. Starts on every day with a complete window.

The shares fall fast and then stop: 27.9% at one year, 15.8% at two, 0.7% at three, nothing at four and 0.2% at five. The medians are enormous because bitcoin traded below a dollar in 2010 and below a thousand dollars until 2017. A median 4-year change of +1,289% is driven by early starts. The best windows over the whole series are sub-dollar 2010 and 2011 starts with gains in the hundreds of thousands of percent, so the last column restricts the best case to starts since 2013, and a later section shows how every statistic shifts when the early years are cut.

The worst windows are the more useful column. The worst 1-year window began on the record close of December 16, 2017 ($19,640.51) and ended December 16, 2018 at $3,195.41. The ratio is 0.1627, a change of -83.7%. The worst 2-year window began on the record close of December 4, 2013 ($1,134.93) and ended two years later at $363.24, a change of -68.0%. Both started at the top of a run.

The 4-year row reads differently. The thinnest window in the entire series began April 13, 2021 at $63,445.64 and ended April 13, 2025 at $83,568.95, a gain of +31.7%. There is no loss to report, so the worst case is the smallest gain. Annualised, +31.7% is +7.1% a year, and that is the lowest four-year growth rate of any start in the record. It matters for the CoinDesk comparison below.

Why Four Years Is Clean While Three and Five Are Not

If the history only showed that windows get safer as they lengthen, the 5-year row would be at least as clean as the 4-year one. It is not: nine 5-year windows ended below start, and 33 of the 3-year windows did. The pattern is not a smooth decline in risk. It is a record of where specific tops fall relative to specific lows.

Table 2 lists the bad windows as episodes. Consecutive daily starts that ended below start are merged into one episode when fewer than 60 days separate them, so 1,558 losing 1-year windows become six episodes.

Table 2. Episodes of windows that ended below start

Holding periodBelow-start windowsEpisodesStart dates of each episode (windows below start, worst window)
1 year1,55862011-05-10 to 2011-08-24 (91; worst 2011-06-09 to 2012-06-09, -80.7%); 2013-11-14 to 2014-12-04 (377; worst 2014-01-14 to 2015-01-14, -78.5%); 2017-10-31 to 2018-05-22 (203; worst 2017-12-16 to 2018-12-16, -83.7%); 2019-06-20 to 2019-09-13 (49; worst 2019-06-26 to 2020-06-26, -28.8%); 2021-02-03 to 2022-06-12 (491; worst 2021-11-09 to 2022-11-09, -76.5%); 2024-11-19 to 2025-11-02 (347; worst 2025-08-13 to 2026-08-13, -48.6%)
2 years82442013-11-11 to 2014-08-09 (195; worst 2013-12-04 to 2015-12-04, -68.0%); 2017-11-20 to 2018-05-10 (149; worst 2017-12-17 to 2019-12-17, -65.7%); 2020-11-13 to 2022-01-19 (420; worst 2021-03-11 to 2023-03-11, -64.4%); 2024-03-08 to 2024-08-01 (60; worst 2024-06-05 to 2026-06-05, -14.3%)
3 years3322013-11-22 to 2014-01-15 (27; worst 2013-11-29 to 2016-11-29, -35.2%); 2021-02-08 to 2021-02-22 (6; worst 2021-02-21 to 2024-02-21, -10.0%)
4 years00none
5 years912017-12-07 to 2018-01-06 (9; worst 2017-12-16 to 2022-12-16, -15.4%)

Calculated by CryptosEyes from the same series. An episode is a run of daily start dates whose windows ended below start, merged when the gap to the next run is under 60 days.

The 3-year row has two episodes. The first runs from November 22, 2013 to January 15, 2014, starts within weeks of the December 4, 2013 record close, and the worst of them is November 29, 2013 ($1,128.73) to November 29, 2016 ($731.65), a change of -35.2%. The second is six starts from February 8 to 22, 2021, and the worst is February 21, 2021 ($57,500.72) to February 21, 2024 ($51,744.32), or -10.0%. All 33 are in those two bursts. The first burst spans 54 days and the second 14 days.

The 5-year row has one episode, with nine starts from December 7, 2017 to January 6, 2018. The worst is the December 16, 2017 record close of $19,640.51 against the December 16, 2022 close of $16,606.11, a change of -15.4%. The December 2017 top was followed by a lowest close of $3,185.07 in December 2018, so the same starts were deeply negative at one and two years. At five years the window ended in December 2022, after the price had gone up to $67,541.76 on November 8, 2021 and back down to a lowest close of $15,758.29 on November 9, 2022. The December 16, 2022 close of $16,606.11 was 15.4% under the record close of five years earlier.

The 4-year row has no episodes. The reason is not that four years is a magic length. Look at the spacing of the record closes: December 4, 2013 to December 16, 2017 is 1,473 days, December 16, 2017 to November 8, 2021 is 1,423 days, and November 8, 2021 to October 6, 2025 is 1,428 days. A 4-year window is 1,461 days. So a 4-year window beginning at a record close ended within 38 days of the next record close. The 2013 start ended December 4, 2017 at $11,562.13, twelve days before the next record, for +918.8%. The 2017 start ended December 16, 2021 at $47,663.47, 38 days after the November 2021 record, for +142.7%. The 2021 start ended November 8, 2025 at $102,335.02, for +51.5%.

That alignment is a feature of the cycle spacing in this particular sample, and it is the most important caveat on the whole result. Three peak-to-peak gaps of roughly four years each line up with the 4-year horizon, which is why it is clean in this sample. A window of three years from the same tops ended at -32.5%, +8.8% and +13.3%; a window of two years ended at -68.0%, -65.0% and -47.0%. The halving-cycle spacing is covered in the halving cycles history, and the depth and duration of each decline is in the bear market table. The holding-window view adds how often a given length of ownership ended under water for any start date, not only the tops.

Buying at a Record Close: The Ladder from Three Peaks

A start on a record close is an entry at the highest price the series had seen up to that day. Table 3 follows the three earlier record closes through four years, and the current one so far.

Table 3. Windows that start at a record close

Start (record close)CloseAfter 1 yearAfter 2 yearsAfter 3 yearsAfter 4 years
2013-12-04$1,134.932014-12-04: -67.5%2015-12-04: -68.0%2016-12-04: -32.5%2017-12-04: +918.8%
2017-12-16$19,640.512018-12-16: -83.7%2019-12-16: -65.0%2020-12-16: +8.8%2021-12-16: +142.7%
2021-11-08$67,541.762022-11-08: -72.6%2023-11-08: -47.0%2024-11-08: +13.3%2025-11-08: +51.5%
2025-10-06$124,824.452026-10-06: -31.5%not completenot completenot complete

Calculated by CryptosEyes from Coin Metrics PriceUSD, pulled November 3, 2026. The October 6, 2025 row has one finished window, the 1-year window to 2026-10-06, which is counted in Tables 1, 2 and 4 and in the sixth 1-year episode. Its 2-year to 4-year windows are still open, and the -38.6% at November 2, 2026 (day 392) used elsewhere on this page is a mark to the latest close, not a result.

The first two columns show how deep and how long the damage ran. From the December 2013 close, the price was 67.5% lower after one year and 68.0% lower after two, then 32.5% lower after three, and the fourth-year jump to +918.8% came from a market that moved from about $766 to about $11,562 in twelve months. From the December 2017 close, the one-year figure was -83.7% and the two-year -65.0%; the three-year point was +8.8% because the price had passed the old record on November 30, 2020 (the price-history page dates that crossing). From the November 2021 close, the figures were -72.6%, -47.0%, +13.3% and +51.5%.

Across the three completed ladders, the 1-year change at a record close ran between -67.5% and -83.7%. The 4-year change was positive each time, and it shrank each time: +918.8%, +142.7%, +51.5%. That shrinking is a fact about the three samples. With a larger asset the same percentage move needs far more capital, and the market value table on the price-history page shows year-end market value growing from $8.9 billion in 2013 to $1.62 trillion on October 27, 2026. The fourth point of the current ladder does not exist yet, so there is nothing yet to compare.

How the Start-Date Definition Moves the Answer

A holding-period table needs a start rule, and the rule changes the answer. Table 4 applies five rules to the same prices.

Table 4. Share of windows that ended below start, by start-date rule

Start-date definition1 year2 years3 years4 years5 years
Every day27.9% (1,558 of 5,587)15.8% (824 of 5,222)0.7% (33 of 4,856)0.0% (0 of 4,491)0.2% (9 of 4,126)
Month-end28.3% (52 of 184)16.9% (29 of 172)0.6% (1 of 160)0.0% (0 of 148)0.0% (0 of 136)
First day of year (Jan 1)26.7% (4 of 15)28.6% (4 of 14)0.0% (0 of 13)0.0% (0 of 12)0.0% (0 of 11)
Last day of year (Dec 31)26.7% (4 of 15)28.6% (4 of 14)0.0% (0 of 13)0.0% (0 of 12)0.0% (0 of 11)
Days that closed at a record34.8% (87 of 250)30.7% (67 of 218)5.7% (12 of 212)0.0% (0 of 212)1.9% (4 of 211)

Calculated by CryptosEyes from the same series. Month-end uses the last close of each month; Jan 1 and Dec 31 use the close on those dates; the record-close rule uses days that closed above every earlier close.

Table 4 gives three results.

Daily and month-end rules agree. Month-end starts give 28.3%, 16.9%, 0.6%, 0.0% and 0.0% for the five horizons, against 27.9%, 15.8%, 0.7%, 0.0% and 0.2% for every day. One start per month is a thin sample of the daily record, yet it lands within about one percentage point of it at one and two years. The difference at five years is the nine December 2017 and January 2018 days, none of which is a month-end: the December 31, 2017 close of $13,921.48 ended five years later at $16,524.22, up 18.7%, and no other month-end start ended below start either. A month-end table says "never" at five years where the daily table has nine losses, and both are right about their own rule.

January 1 and December 31 give the same shares and different worst cases. Each has 4 losing 1-year windows of 15 and 4 losing 2-year windows of 14, and none at three, four or five years. Yet the worst 1-year window is -71.7% from January 1 and -73.5% from December 31, and the worst 2-year window is -46.7% from January 1 and -48.5% from December 31, because a one-day shift changes which price is the start. The shares are identical because the same four years were bad under either rule. The worst-case dates are in Table 5.

Calendar-year starts can overstate the 2-year share. At two years, January 1 starts show 28.6% below start (4 of 14), against 15.8% on daily starts. With 14 windows, each one is worth 7.1 points. The four losing Januaries (2014, 2018, 2021 and 2022) all fell inside three of the four daily episodes, so the rule happened to sample those episodes more often than their share of the calendar, which is about one start day in six. A table built on calendar-year starts shows a 2-year loss share nearly twice the full daily record.

The record-close rule gives the highest shares at every horizon but four years: 34.8% at one year, 30.7% at two, 5.7% at three, 0.0% at four and 1.9% at five. This is the "bought at the top" case, and it has a few hundred starts at each horizon rather than thousands. At four years it still shows none lost, with a thinnest gain of +31.7%.

Table 5 gives the exact date windows for each worst case under each rule.

Table 5. Worst window under each start-date rule

Holding periodEvery dayMonth-endJan 1Dec 31
1 year2017-12-16 to 2018-12-16: -83.7%2017-12-31 to 2018-12-31: -73.5%2018-01-01 to 2019-01-01: -71.7%2017-12-31 to 2018-12-31: -73.5%
2 years2013-12-04 to 2015-12-04: -68.0%2013-11-30 to 2015-11-30: -66.3%2018-01-01 to 2020-01-01: -46.7%2017-12-31 to 2019-12-31: -48.5%
3 years2013-11-29 to 2016-11-29: -35.2%2013-11-30 to 2016-11-30: -33.7%2014-01-01 to 2017-01-01: +32.5%2013-12-31 to 2016-12-31: +32.8%
4 years2021-04-13 to 2025-04-13: +31.7%2021-03-31 to 2025-03-31: +40.3%2022-01-01 to 2026-01-01: +86.5%2021-12-31 to 2025-12-31: +88.8%
5 years2017-12-16 to 2022-12-16: -15.4%2021-03-31 to 2026-03-31: +16.0%2018-01-01 to 2023-01-01: +23.3%2017-12-31 to 2022-12-31: +18.7%

Calculated by CryptosEyes from the same series. The worst case under the record-close rule is the same window as the daily worst case at every horizon. A positive entry means no window under that rule ended below start, and the figure shown is the smallest gain.

Under the daily rule, the worst 1-year window is -83.7%; under month-end, Jan 1 and Dec 31, it is -73.5%, -71.7% and -73.5%. Moving from a daily rule to a calendar rule can lift the worst-case outcome by 10 to 12 points without changing the data. At three and five years the rules disagree about whether there was a loss at all: the daily rule finds -35.2% and -15.4%, while the Jan 1 and Dec 31 rules find only gains (+32.5% and +32.8% at three years, +23.3% and +18.7% at five).

Two notes on the dates. The 2-year worst case begins on the December 4, 2013 record close, while the 3-year worst case begins five days earlier, on November 29, 2013, when the close of $1,128.73 was 0.55% below that record. And the 1-year worst case begins on December 16, 2017, the day before the December 17 intraday high that the price-history page shows for Bitstamp, which is why the close versus intraday high distinction matters when matching dates between sources.

Does the Sample Era Matter as Much as the Date Rule?

It matters more at one and two years. Bitcoin's 2010 to 2012 prices are fractions of a dollar to a few dollars, with thin markets, and the price-history page treats those rows as approximate. Table 6 repeats the daily rule for three starting eras.

Table 6. Share below start and median change, by sample era

Holding periodStarts from 2010-07-18: share below / medianStarts from 2013-01-01: share below / medianStarts from 2017-01-01: share below / median
1 year27.9% (1,558 of 5,587) / +95.2%31.3% (1,467 of 4,689) / +70.1%33.8% (1,090 of 3,228) / +48.2%
2 years15.8% (824 of 5,222) / +263.7%19.1% (824 of 4,324) / +174.9%22.0% (629 of 2,863) / +124.7%
3 years0.7% (33 of 4,856) / +474.1%0.8% (33 of 3,958) / +332.6%0.2% (6 of 2,497) / +182.8%
4 years0.0% (0 of 4,491) / +1,289%0.0% (0 of 3,593) / +806.5%0.0% (0 of 2,132) / +293.1%
5 years0.2% (9 of 4,126) / +2,378%0.3% (9 of 3,228) / +1,143%0.5% (9 of 1,767) / +518.5%

Calculated by CryptosEyes from the same series. Share below start and median change for daily starts on or after the stated date.

Two things happen as the early years drop out. The medians fall steeply, which is expected: a 1-year median of +95.2% with all starts is +70.1% from 2013 and +48.2% from 2017, and the 4-year median goes from +1,289% to +806.5% to +293.1%. And the loss shares at one and two years rise, from 27.9% to 31.3% to 33.8% at one year and from 15.8% to 19.1% to 22.0% at two. Windows from the early years rarely ended lower. In the 2017 column, which covers the most recent nine years of starts, roughly a third of 1-year starts and a fifth of 2-year starts ended lower, and none of the 2,132 four-year starts did.

The 3-year and 5-year shares do not move in step with those. From 2017 the 3-year loss share is 0.2%, six windows, because the cluster of 27 losing starts from November 2013 to January 2014 drops out. The 5-year share rises from 0.2% to 0.5% as the denominator shrinks and the nine December 2017 and January 2018 starts stay. Cutting the early years does not lower every number; it changes which episodes carry the weight.

The era choice also decides which tops sit inside the sample. The 2017 column leaves out the late 2013 starts, which produced all 27 of the 3-year losses before 2017, and keeps the December 2017 starts that produced all nine 5-year losses. A short record makes those choices decisive, and the next section counts how short it is.

How Many Independent Windows Does the History Contain?

Rolling windows are not independent. A 4-year window that starts on June 1 and one that starts on June 2 share all but one day of their span, and two windows 30 days apart share all but 30 days. Counting 4,491 four-year windows and then saying "zero lost" is counting one outcome thousands of times.

The honest unit is a set of windows that do not overlap. The series covers 16.29 years, from July 18, 2010 to November 2, 2026. That fits four non-overlapping 4-year windows at most, and it fits only three for 257 of the 365 possible places to start the first window. The four from the series start are July 18, 2010 to July 18, 2014 (from $0.09 to $630.75), July 18, 2014 to July 18, 2018 (to $7,376.04), July 18, 2018 to July 18, 2022 (to $22,305.71) and July 18, 2022 to July 18, 2026 (to $64,777.66). All four ended above start: +734,698%, +1,069.4%, +202.4% and +190.4%. Four separate outcomes, all positive, are the evidence behind "never lost over four years", far fewer than 4,491. Starting the chain on any other day of the first year gives three or four windows, and Table 7 shows that none of the 365 starting days produces a loss.

Table 7 repeats the count for every horizon. For each of the 365 days on which the first window could begin, it builds the non-overlapping chain and records the share that ended below start.

Table 7. Non-overlapping windows

Holding periodOverlapping daily windowsNon-overlapping windows that fitShare below start across all 365 start phases
1 year5,58715 or 1613.3% to 40.0%
2 years5,2227 or 80.0% to 14.3%
3 years4,85650.0% to 20.0%
4 years4,4913 or 40.0% in every phase
5 years4,12630.0% in every phase

Calculated by CryptosEyes from the same series. Each chain starts on one of the first 365 days of the series and steps forward by exactly N years.

At one year, 15 or 16 non-overlapping windows fit, and the share below start ranges from 13.3% to 40.0% depending on where the chain starts, bracketing the 27.9% from the overlapping count. At two years the chains hold 7 or 8 windows and 0 or 1 of them lost, which is a share of 0% to 14.3%. At three years the chains hold 5 windows and 0 or 1 lost: 0% to 20%. One window changes the answer by 20 points. The 4-year chains hold 3 or 4 windows and none lost in any of the 365 phases, and the 5-year chains hold 3 and none lost in any of them.

The ledger of bad windows in Table 2 gives a second count of independence. Six episodes produce the 1,558 losing 1-year windows. Four episodes produce the 824 losing 2-year windows. Two produce the 33 losing 3-year windows. One produces the 9 five-year losses. Each loss count reflects a handful of market events, not hundreds or thousands of separate draws, so a statement of the form "X% of windows lost" should be read with the episode count beside it.

The record also holds few peaks. The record-close table on the price-history page lists the highs that stood at the end of each run of new records: 2010, 2011, 2013, 2017, 2021 and 2025. Since 2013 that is four peaks and three completed gaps between them. Three or four outcomes can describe what happened; they are too few to estimate a long-run frequency. Drawdown size and recovery, the other side of holding-period outcomes, are framed in the volatility survival guide.

The CoinDesk Four-Year Growth-Rate Figure, Checked Against the Series

On March 12, 2025, CoinDesk published an article titled "Bitcoin's Four-Year Compounded Annual Growth Rate Drops to Record Low of 8%" (Wayback copy, saved April 8, 2025; the figures below are quoted from it). Its summary reads: "The Ethereum-to-Bitcoin ratio (ETH/BTC) has dropped to 0.022, its lowest level since 2020. As the four-year CAGR has turned negative at -6%, Bitcoin's four-year CAGR remains positive at 8% but has slipped to record lows."

The summary sentence is awkward on its own, but the rest of the article is not ambiguous: -6% is the four-year CAGR that "has turned negative" for the ETH/BTC ratio, and bitcoin's own is "positive at 8%". The two charts in the article are captioned "BTC: 4yr Compound Annual Growth Rate (Glassnode)" and "ETH/BTC 4yr CAGR (Glassnode)", and the body says the same thing in two places. It states that "Bitcoin's (BTC) four-year compound annual growth rate (CAGR) has dropped to its lowest recorded level of 8%, according to Glassnode data", and later that "The ether (ETH)-to-bitcoin (ETH/BTC) ratio has also entered negative CAGR territory at 6%, reflecting the underperformance of ethereum's native token compared to bitcoin." The body also reports the ratio as 0.024 where the summary says 0.022, and it drops the minus sign on the 6%; the article's footer says parts were produced with AI tools and reviewed by its editors. None of that changes which asset the -6% belongs to. A page that quotes it as a bitcoin return misreads the source.

Table 8 converts the two rates to total change over four years, which is the figure a holder would see.

Table 8. Annual rate and four-year total change

Annual rate over four yearsTotal change over four years
-6.0% a year-21.9%
+0.0% a year+0.0%
+8.0% a year+36.0%

Calculated by CryptosEyes: (1 + annual rate) to the fourth power, minus one.

A -6% compound rate is a -21.9% total change. A +8% rate is +36.0%. The series gives no bitcoin window that ended below start at four years, so no window comes near -21.9%. Bitcoin's 4-year changes in the series run from +31.7% upward.

Does the series support the article's "record low"? Table 9 recomputes the window the article describes and its neighbours.

Table 9. Four-year change around the CoinDesk dates

WindowStart closeEnd closeFour-year changeAnnual rate
2021-03-11 to 2025-03-11$57,847.25$82,747.16+43.0%+9.4%
2021-03-12 to 2025-03-12$57,334.64$83,630.16+45.9%+9.9%
2021-03-13 to 2025-03-13$61,288.79$81,110.29+32.3%+7.3%
2021-04-13 to 2025-04-13$63,445.64$83,568.95+31.7%+7.1%

Calculated by CryptosEyes from Coin Metrics PriceUSD, pulled November 3, 2026. Glassnode's own series is a different data source.

For March 12, 2021 to March 12, 2025 the series gives +9.9% a year. Move the start one day later and the rate is +7.3%, because the March 13, 2021 close was $61,288.79 against $57,334.64 the day before. Glassnode's +8% sits inside that range, and the sensitivity is the point that CoinDesk itself makes: "this metric is highly dependent on the reference points. In 2021, Bitcoin was experiencing a blow-off top early in the cycle, whereas in March 2025, $80,000 could be marking a cycle bottom." The lowest four-year rate in the whole series is +7.1%, for April 13, 2021 to April 13, 2025, a month after the article. So "record low, still positive" is consistent with this record: the growth rate was at its lowest ever and still above zero. Once the -6% is put back with the ratio it belongs to, the 4-year result and the CoinDesk article do not conflict.

What the reconciliation does not show is that the ETH/BTC number is unimportant. If the -6% is exact, a 4-year rate of that size for a ratio means ether lost about 22% of its value against bitcoin over the period, which is a fact about relative price, not about dollars. It sits outside this page's subject.

Windows That Have Not Finished

The windows excluded from every table above are the ones with no end date yet: 365 daily starts at one year, 730 at two, 1,096 at three, 1,461 at four and 1,826 at five. For 4-year windows, the starts run from November 3, 2022 to November 2, 2026.

The record close of October 6, 2025 at $124,824.45 is the one that matters most. The November 2, 2026 close of $76,680.75 is 38.6% below it, 392 days in. That is a mark to the latest price, not a result: the window is open until October 2029 under a 4-year rule and the series says nothing about how it ends. Of the 1,461 open 4-year starts, 549 have a start close above the latest close; they run from November 9, 2024 to November 1, 2026 and include 29 days that set record closes between November 9, 2024 and October 6, 2025. The other 912 open starts are at or below the latest close.

On the finished windows, the 1-year row of Table 2 shows that the sixth episode began on November 19, 2024 and has 347 losing 1-year windows from starts through November 2, 2025; the worst, August 13, 2025 ($123,442.68) to August 13, 2026 ($63,395.01), is -48.6%. The 2-year row has 60 losing windows from starts between March 8 and August 1, 2024, the worst June 5, 2024 ($71,087.44) to June 5, 2026 ($60,907.61) at -14.3%. Those are finished windows, already counted in the distributions above.

What a Rolling Window Table Leaves Out

A holding-window table answers a narrow question: what happened if the whole amount was invested on one day and left alone. It does not describe what most people actually do. Someone who buys monthly over four years has an average cost between the highest and lowest prices in the period; the dollar-cost averaging guide covers that case with its own arithmetic. Someone who sold during a drawdown has a different result from the window that ended later. A window table also assumes the holder never needed the money before the end date.

Three further limits apply to these figures. First, they are price changes. Fees, spreads, custody costs, taxes on gains and inflation would all reduce a holder's result, and none is modelled here; a gain of +31.7% over four years is a price change, not a take-home figure. Second, the series is a daily reference close at midnight UTC, not a price anyone could necessarily trade at; the price-history page shows an intraday low on Bitstamp 22.4% under the lowest daily close of the same year (March 2020). Third, the earliest closes come from a very small market, and Coin Metrics' catalog entry does not say which venues its earliest values draw on, so the 2010 to 2012 windows are best treated as approximate. That is why the era table above sets the 2013 and 2017 cuts beside the full record.

"No four-year window lost" is a statement about 4,491 overlapping windows that reduce to three or four separate outcomes. It is not offered as an argument in either direction. Bitcoin has fallen 83.7% in one year on this series, from the December 2017 close to the December 2018 close, and the price at the latest close sits 38.6% under the October 2025 record. How a particular holder would experience a drawdown of that size depends on their horizon, their other assets and their obligations, none of which a table can see. The realised price guide covers an on-chain average of what each coin was worth on the day it last moved, which is a different lens on how far the market sits from that average at a given price.

Rebuilding the Table

Everything above can be reproduced from the free Coin Metrics community API with no key:

1.Request https://community-api.coinmetrics.io/v4/timeseries/asset-metrics?assets=btc&metrics=PriceUSD&frequency=1d&start_time=2010-07-01&page_size=10000. The response for this page holds 5,952 rows, from 2010-07-18 to 2026-11-02, in one page.
2.Keep each row's date label and PriceUSD value. There are no gaps between the first and last date.
3.For each start date the rule selects, find the close N calendar years later (February 28 when a leap day does not exist). If the end date is past the last row, skip the start.
4.Return = end close divided by start close, minus one. A window is below start when the return is negative.
5.For the independence counts, step forward by exactly N years from each of the first 365 days of the series and tally windows and losses. For the episode table, merge consecutive losing start dates whose gap is under 60 days.

If your numbers differ from this page by a cent, check rounding. If they differ by more, check the series, the time zone and whether the date labels are shifted by one day. The bear market and halving pages in the on-chain and cycles hub use the same series.

Frequently Asked Questions

Has bitcoin ever lost money over four years?

Not on the complete windows in this record. Of 4,491 four-year windows with daily starts from July 18, 2010 to November 2, 2022, none ended below its start close. The thinnest was April 13, 2021 ($63,445.64) to April 13, 2025 ($83,568.95), a gain of +31.7%. Windows that began after November 2, 2022 have not finished.

Did bitcoin's four-year CAGR turn negative in March 2025?

No. CoinDesk's March 12, 2025 article reports bitcoin's four-year CAGR at +8%, a record low, and attributes the -6% to the ETH/BTC ratio. On Coin Metrics closes, bitcoin's four-year rate for the nearby dates was between +7.3% and +9.9%, and the lowest of any start was +7.1%.

How often did bitcoin end a one-year holding period lower?

On this series, 1,558 of 5,587 daily one-year windows, or 27.9%, ended below the start close. They come from six episodes: 2011, late 2013 to 2014, late 2017 to 2018, mid-2019, 2021 to 2022, and November 2024 to November 2025. From 2017 onward the share is 33.8%.

What was the worst bitcoin holding period?

Over one year, December 16, 2017 ($19,640.51) to December 16, 2018 ($3,195.41), a change of -83.7%. Over two years, December 4, 2013 to December 4, 2015, -68.0%. Over three years, November 29, 2013 to November 29, 2016, -35.2%. Over five years, December 16, 2017 to December 16, 2022, -15.4%.

Why do some tables show fewer losing windows than others?

The start-date rule changes the sample. Daily starts show nine losing five-year windows, all within December 7, 2017 to January 6, 2018, while month-end, January 1 and December 31 starts show none. January 1 starts show 28.6% of two-year windows lower, against 15.8% for daily starts, because only 14 windows exist.

How many independent four-year windows does bitcoin's history contain?

At most four, and three for most start dates. The series spans 16.29 years, so it holds at most four non-overlapping four-year windows. The 4,491 daily windows overlap almost completely. The four from the series start all ended above their start price.

Does this predict what the next four years will do?

No. The table describes completed windows from one price series. It rests on about four independent four-year outcomes, and the open window from the October 6, 2025 record close is not complete. Nothing here is a forecast or a reason to buy or sell.

CryptosEyes publishes general educational research, not financial, investment, tax or legal advice. Figures describe what one price series did in the past, in US dollars, before fees and taxes. Bitcoin has lost most of its value within a single year, as the 2017 to 2018 and 2021 to 2022 windows on this page show.

Sources

Primary documents, market context, and methodology references behind this article.

Reviewed against source notes and calculations
On-Chain
Research note: This article is educational market research, not financial advice. Crypto and public equity data can change quickly; see our methodology and editorial policy for sourcing, review, and correction standards.
Important: Educational Purposes OnlyThe data, charts, treasury tracking metrics (including mNAV and SPS), and research provided on CryptosEyes.com are for informational and educational purposes only. They do not constitute certified financial, investment, or trading advice. Digital assets like Bitcoin and Ethereum are highly volatile. Always conduct your own research and consult with a registered financial advisor before making investment decisions.