Is the Bitcoin Four-Year Cycle Real? How to Test Cycle Claims
Are crypto cycles real? Testing cycle claims honestly. How to test a crypto cycle claim fairly: tiny samples, hindsight, multiple testing, forking paths, formal bubble tests, pre-registration and a worked test of halving rules on public data.
The evidence is consistent with a four-year rhythm but too thin to establish one: there are only four halving cycles, peaks and lows can only be dated in hindsight, and many rules fitted to past cycles failed on later ones. Formal bubble tests confirm explosive episodes, which is not the same as a regular schedule.
You’ll learn
- Restate a cycle claim as a testable rule with a fixed definition, horizon and comparison, written before outcomes are known
- Quantify how little four cycles can show, using intervals, base rates and the chance of a match by luck
- Recognize hindsight bias, look-ahead in dating peaks and lows, multiple testing and the garden of forking paths in cycle research
- Explain what the Phillips-Shi-Yu bubble tests detect, what Cheung, Roca and Su found for Bitcoin, and why explosive episodes are not evidence of a regular cycle
- Pre-register a cycle hypothesis and score it out of sample with runnable code
A three-for-three pattern that broke in March 2024
Before 2024, a simple cycle pattern had held every time. Bitcoin had set a new all-time high only after a halving, the scheduled cut in new issuance. The 2011 high of about $29 was first exceeded in February 2013, 83 days after the November 2012 halving. The December 2013 high was first exceeded in February 2017, 229 days after the July 2016 halving. The December 2017 high was first exceeded on November 30, 2020, 203 days after the May 2020 halving. On March 5, 2024, Fortune reported that Bitcoin had passed its November 2021 record, trading as high as $69,170 intraday, with spot exchange-traded funds drawing record inflows. On daily reference prices the old high was first exceeded on March 4. The halving came on April 20, 47 days later. A rule with a perfect record had failed on its fourth try. Three for three sounds strong. As this lesson shows, three matches of a pattern chosen after looking at the data are close to no evidence at all, and the same reasoning applies to every cycle rule, including the ones that have not failed yet.
Lessons 16 to 19 described the four-year cycle, on-chain cycle indicators, trend models and macro explanations. This lesson asks a narrower question about all of them: what kind of evidence could show that a cycle claim is true, and how much of that evidence exists? The answer is more useful than a yes or no. It tells a reader which claims can be tested at all, which have been tested fairly, and how to write a claim so that the future can test it.
What a cycle claim is
“Crypto moves in cycles” can mean at least three different things, and they need different tests.
| Kind of claim | Example | What would test it |
|---|---|---|
| Periodicity | Peaks recur about every four years, tied to the halving | Timing of turning points relative to a fixed clock, compared with chance |
| Regime alternation | Long rising phases alternate with long falling phases | Persistence of trends, measured against a random walk with the same volatility |
| Episodic bubbles | Prices periodically become explosive and then collapse | Formal tests for explosive behavior, such as the Phillips-Shi-Yu tests |
A dataset can support one of these and not the others. Evidence of bubbles is not evidence of a four-year clock.
Testable cycle rule: A statement that names: the clock (halving dates, calendar quarters, days since the last low); the event being timed (defined so it can be identified without later data, or with an explicit delay); the window or threshold; the comparison (chance, a base rate or a rival rule); and the cases it will be scored on, all written down before those cases are observed.
Most published cycle claims fail the last condition. They are statements about the cycles already seen, which the claim was shaped to fit. That does not make them false. It means the existing cycles are the training data, and only new cycles can test them.
Four cycles: the sample-size problem
Bitcoin has had four halvings (November 28, 2012; July 9, 2016; May 11, 2020; April 20, 2024) and four completed peaks after them. Every halving-cycle claim therefore rests on at most four observations, and each new observation takes about four years to arrive. Tversky and Kahneman (1971) named the intuition that small samples resemble the population they come from “belief in the law of small numbers”, and showed that even trained researchers hold it.
How much four for four can show
- A rule matches 4 of 4 cycles. The observed rate is 100%.
- The Wilson 95% interval for 4 of 4 is: center = (1 + 1.96²/8) / (1 + 1.96²/4) = 1.480 / 1.960 = 0.755; half-width = 1.96 × √(1.96²/64) / 1.960 = 0.245. Interval: 51% to 100%.
- If the event has a base rate of 72% on any random date, the chance of 4 of 4 by luck is 0.72⁴ = 0.27.
- If the rule were written after seeing the four cases, the chance it matches them is close to 1, whatever the truth.
The effective sample is smaller still, because the cycles are not independent experiments. They share one asset, one community of holders and a growing body of commentary about the cycle itself. If enough participants act on a four-year expectation, the pattern can partly cause itself, which makes it real in the data and fragile at the same time. And the setting has changed between cycles: market value, volatility, access through funds, and the share of new supply in total supply are all very different in 2024 from 2012.
Hindsight and look-ahead in dating cycles
Hindsight bias: The tendency, once an outcome is known, to judge it as having been more predictable than it was. Fischhoff (1975) showed experimentally that people told an outcome rate its prior likelihood higher than people not told, and are largely unaware of the shift.
In cycle work, hindsight enters through the dating of turning points. A cycle peak is the highest price before a large decline, so it can only be identified after the decline. A cycle low is the lowest price before a sustained recovery, so it can only be confirmed after the recovery. Any rule stated in terms of “the peak” or “the low” therefore uses information that was not available on the day. Lesson 9 called this look-ahead bias. Here it is built into the definitions.
| Statement | When it becomes knowable |
|---|---|
| “The peak came 17.5 months after the halving” (October 6, 2025) | Only after prices had fallen far enough, long enough, to rule out a higher high |
| “The low came about a year after the peak” | Only after a recovery confirms no lower low. The lowest daily price after the 2025 peak, through November 19, 2026, was $58,525 on June 30, 2026, and it is not yet known whether that is the cycle low |
| “Cycles have shown diminishing returns” | Each peak’s size is known only after it is a confirmed peak; the curve it is compared with is refitted using later data (Lesson 18) |
| “The 2026 bottom matched the midterm-year pattern” | Requires both the bottom date and the definition of the pattern to be fixed, the first only after the fact |
Peak and low dates use Coin Metrics daily reference prices. Intraday extremes on exchanges differ.
A fair test replaces hindsight events with rules that can be evaluated on the day. Instead of “the peak”, use a mechanical event such as “the first day price closes 50% below its highest close of the past year”, which is knowable the moment it happens. Such a rule is less elegant and usually performs worse than a hindsight story, which is exactly why it is honest.
Multiple testing and the garden of forking paths
Lesson 9 showed that testing many rules guarantees that some look good by chance, and introduced Bonferroni and Benjamini-Hochberg corrections. Harvey, Liu and Zhu (2016) argued for this reason that a newly proposed factor in asset pricing should clear a t-ratio above 3.0 rather than 2.0. In finance, one method that corrects for the full set of rules a researcher searched is White’s (2000) Reality Check. Sullivan, Timmermann and White (1999) applied it to the technical trading rules first studied by Brock, Lakonishok and LeBaron, quantifying how much data snooping inflates the apparent performance of the best rule.
Gelman and Loken (2013) pointed out a subtler version. A researcher may run only one test, honestly. But if the details of that test (which cycles to count, which price series, which window, which definition of a peak) were chosen after seeing the data, the test is effectively one path chosen from many. They called this the garden of forking paths, and stressed that it needs no conscious fishing.
| Fork | Choices typically available |
|---|---|
| Which cycles count | Include 2011 or start at the first halving; treat 2019 as a peak or not; count the 2021 April high or the November high |
| Which price | Daily close, daily reference rate, intraday high, weekly close; one exchange or an index |
| Which clock | Halving date, calendar quarter, days from the prior low, US election or midterm years |
| Which window | Peak within 12–18 months, or 10–20, or “in the fourth quarter” |
| Which asset | Bitcoin only, or Ethereum and other coins as well |
| Which outcome | Peak timing, peak size, drawdown depth, return over a fixed horizon |
Six forks with three or four options each give several hundred to a few thousand possible specifications. Four cycles cannot discipline that many.
A useful habit is to count the forks before quoting a result, and to ask whether the result survives on most of the reasonable paths, not just the one shown. If the four-year pattern were strong, it would appear on daily and weekly data, with or without 2011, and with windows that were not tuned to the cases.
Formal bubble tests: what they find and what they do not
Econometrics has a well-developed test for one kind of cycle claim: episodes in which prices grow explosively. The tests build on the augmented Dickey-Fuller (ADF) regression, which asks whether a series behaves like a random walk or something more explosive.
y is the (log) price. β = 0 corresponds to a random walk (a unit root). β > 0 means each period’s level pushes the next change in the same direction: explosive behavior. The test statistic is the t-ratio on β, and the bubble tests look for large positive values.
SADF and GSADF: The sup ADF test (Phillips, Wu and Yu, 2011) computes the ADF statistic on forward-expanding samples and takes the largest. The generalized sup ADF test of Phillips, Shi and Yu (2015) also varies the start of each window, which gives it more power when a sample contains several bubbles. A backward version, BSADF, is computed at each date and compared with critical values to date-stamp when an explosive episode began and ended, using only data up to that date.
Phillips, Shi and Yu (2015) applied the method to monthly S&P 500 data from 1871 to 2010 and identified well-known historical episodes of exuberance and collapse, more of them than earlier methods found. Cheung, Roca and Su (2015) applied it to Bitcoin prices on the Mt. Gox exchange from 2010 to 2014. They found several short-lived bubbles and three large ones between 2011 and 2013, each lasting 66 to 106 days, the largest coinciding with the collapse of Mt. Gox.
These results establish something real: Bitcoin’s price history contains episodes that are statistically explosive, not just volatile. They do not establish a four-year clock. A series can have repeated bubbles at irregular intervals. The bubble tests also come with their own assumptions, including the choice of minimum window, the lag length k, and critical values that depend on the error structure. Their date-stamping uses only past data, which is a strength. But the explosive root they detect is a description of the price path, not of its cause, so they cannot tell a halving-driven bubble from a leverage-driven or attention-driven one.
Pre-registration and out-of-sample checks
Nosek, Ebersole, DeHaven and Mellor (2018) argued that the central distinction in empirical work is between prediction, which tests an idea on observations not yet seen, and postdiction, which uses existing data to generate or refine an idea. Both are valuable. Confusing them is the problem, because a postdiction presented as a prediction carries P values and confidence that it has not earned. Their remedy is pre-registration: writing down the question and the analysis plan, with a date, before seeing the outcome data, and reporting any departures from it.
Cycle research is well suited to this, because the outcome data arrive slowly and on a known schedule. A claim registered today about the next halving cycle cannot be fitted to it. In November 2026 Baquero and Menezes did exactly this for the Bitcoin power-law trend: they publicly hashed their forecasts for 2027 to 2030 and their procedure before applying it once, so it can be scored against prices that do not yet exist. Whatever the scores turn out to be, that design produces evidence. A new chart of old cycles does not.
| Item | Example entry |
|---|---|
| Date written | The day the registration is made public, with a hash or timestamped post |
| Hypothesis | “The highest daily close of the halving cycle that begins in 2028 occurs between 12 and 18 months after the halving block.” |
| Data | Coin Metrics daily PriceUSD, with a named backup source if it is unavailable |
| Event definition | Highest daily close between the halving and the next halving, so the result is known only at the end, and stated as such |
| Comparison | Chance probability if timing were uniform across the 48-month cycle: 6/48 = 12.5% |
| Decision rule | Report hit or miss, the actual month, and the distance in months from the window |
| What would count against the claim | A peak outside the window; a peak in the window by less than one month, reported as such |
| Departures | Any later change, with the reason and the original result alongside |
An illustrative registration. It describes a test, not an expectation about prices.
Two weaker substitutes exist when waiting is not possible. A holdout sets aside the most recent cycle and builds the rule only from earlier ones, which is honest only if the researcher truly did not look. A walk-forward design refits the rule at each date using only data before it, as Lesson 18 described for trend models. Both are better than nothing. Neither is as clean as a public registration, because the person running them already knows how the later cycles turned out.
Base rates
A cycle claim is only informative relative to what happens anyway. Bitcoin rose over most long windows in its history, so a claim that it rose in some particular 12-month window is weak on its own. On Coin Metrics daily data from July 2010, 71.9% of all 365-day windows ended higher than they began (68.5% for windows starting in 2013 or later), and the median 365-day change across all windows was +94%, a figure dominated by the early years. Any rule about rising years has to beat that.
Calendar-quarter peaks against chance
- The four post-halving cycle peaks (December 2013, December 2017, November 2021, October 2025) all fell in the fourth calendar quarter.
- If peak timing were uniform across quarters, the chance of four Q4 peaks is 0.25⁴ = 0.0039, or about 1 in 256.
- But the claim was formed after the first two or three, so only the later ones are tests. Taking the 2021 and 2025 peaks as tests gives 0.25² = 0.0625.
- Adding the June 2011 peak, which preceded the first halving, makes it 4 of 5, and treating the June 2019 high as a peak makes it 4 of 6.
Benjamin Cowen of Into The Cryptoverse is a prominent defender of this timing pattern, stating in 2026 that Bitcoin “tops in the fourth quarter” every cycle. On the record so far the statement is accurate for the four post-halving cycles. The fair evaluation is the one above: the pattern is striking, the number of independent tests is small, and its strength depends on choices about which peaks count.
Worked exercise: testing halving rules on public data
The code below runs with standard Python and the free Coin Metrics community API, which needs no key at the time of writing. It tests two claims and then scores five rules that a person could have written after the first two full cycles. It was run for these notes and the output is shown below it.
Claim A. All four post-halving years rose, but so did 72% of all years, so four rises would happen by chance about 27% of the time. Ranked against every 365-day window, the four post-halving years sat at the 99th, 74th, 84th and 35th percentiles. The mean rank of 0.73 gives a rough z of 1.57, which is not strong evidence, and the calculation ignores the trend and the overlap between windows. The most recent case, which is the only one in the era of exchange-traded funds, ranked below the median. The supportable statement is that post-halving years were often strong in the past, with weakening evidence in the latest cycle.
Claim B. The peak-timing rule scored two out-of-sample hits, with a chance probability of about 1.6% under uniform timing. That looks impressive, and it is the best result in the exercise. Two cautions apply. The peak dates used to score it are hindsight events, so the rule could not have been checked on the day. And the 12–18 month window itself was written for these notes by someone who already knew all four peak dates, which is precisely the problem this lesson describes. A truly blind version of the test would need the window from a document dated before 2020.
The five rules. All five fitted the first two cycles perfectly, by construction. On the next two, they scored 4 of 10. The timing rule survived; the size rules (an 80–90% fall, a tenfold peak) failed on both later cycles; the other two split. If only the surviving rule is remembered and repeated, it looks like a law. Seen alongside the rules that could have been written from the same data, it looks like one success among several attempts. That selective memory is survivorship applied to ideas instead of tokens.
Common mistakes
- Counting cycles that were used to build a rule as evidence for the rule.
- Stating rules in terms of “the peak” or “the low”, which are only known long afterward.
- Reporting the one rule that worked without the others that were tried or could have been tried.
- Treating evidence of bubbles as evidence of a regular cycle.
- Ignoring base rates: an asset that rose in most years makes any “up year” rule look good.
- Assuming each cycle is an independent trial when the same holders and the same commentary carry over.
Is the four-year cycle real?
The case that it is. Four post-halving cycles have peaked in the fourth calendar quarter, 12 to 18 months after the halving, and a version of that timing rule written after two cycles matched the next two. Drawdowns and peak multiples have followed a consistent, shrinking pattern. Formal tests confirm repeated explosive episodes. A plausible mechanism exists in coordination: if many participants expect a four-year rhythm, their behavior can produce one. Defenders also argue that the burden of proof sits with the claim that this cycle is different. The case that it is not, or not reliably. Halvings are known years in advance, so an efficient market should price them gradually rather than react to them on schedule. There are four observations, they are not independent, and the most memorable rules were selected from many with hindsight. Several rules that fitted the first two cycles failed on the next two. The size of the halving’s supply effect has shrunk each time, since new issuance is now a small share of the existing supply, while demand from funds and companies has grown. The 2024 new high came before the halving, breaking a three-for-three pattern. What would resolve it. Not more charts of the same cycles. Pre-registered rules about the 2028 cycle, stated with mechanical event definitions and scored when the data arrive, are the only evidence that can separate a real rhythm from a pattern read into four points. Until then, “the evidence is consistent with a four-year rhythm, but too thin to establish one” is the most accurate summary.
Problem set
- A post says: “The Pi Cycle Top called 3 of the last 3 cycle tops.” Using Lesson 18’s table, restate the record precisely, then compute the Wilson 95% interval for 3 of 3 and explain what else is needed.
Show solution
On daily reference data, crosses matched the December 2013 and December 2017 peaks, matched the April 2021 local high rather than the November 2021 peak, and did not occur at all in 2024–25. So the record is two clear matches, one partial and one miss, about 2 of 4 by the strict cycle-peak definition. For 3 of 3 the Wilson interval is about 44% to 100%, consistent with a coin flip. Also needed: whether the indicator was designed after the cycles it is credited with, the base rate of a cross occurring near any high, and whether “called” means within days of the peak or within months. - A researcher tests 4 clocks (halving, calendar quarter, US midterm year, days since prior low) × 3 windows × 2 price series = 24 specifications for peak timing, and reports the best one with p = 0.01. Apply a Bonferroni correction at 5%. What else would make the test more honest?
Show solution
Threshold = 0.05 / 24 ≈ 0.0021, so p = 0.01 does not pass. The specifications are correlated (the same four peaks), so the effective number of tests is lower and Bonferroni is conservative, but the result is still weak with four observations. More honest: report all 24 results; state which specification was planned first; report the best result alongside the median result; and register the chosen specification for scoring on the next cycle. - Compute the chance that two independent cycle peaks both fall in a pre-specified 6-month window if peak timing were uniform over a 48-month cycle. Then explain why uniform timing may be the wrong null.
Show solution
(6/48)² = 0.125² = 0.0156, about 1.6%. Uniform timing assumes any month is equally likely. But peaks can only occur after a long rise, and lows tend to follow peaks with a lag, so timing is constrained by the dynamics of a trending, volatile series. A random walk with drift and Bitcoin’s volatility, simulated many times, would give a better null distribution of peak timing; the chance of a match under that null could be higher than 1.6%. A null that is too easy to reject overstates the evidence. - Write a mechanical, knowable-on-the-day replacement for “the cycle peak”, and explain one way it would score worse than the hindsight version.
Show solution
Example: “The first day that the daily close is at least 40% below the highest close of the preceding 365 days marks a confirmed top; the top date is the date of that highest close.” This is knowable on the confirmation day, which comes weeks or months after the top. It scores worse because (1) confirmation is late, so any rule that uses it to act is late; (2) it can trigger in mid-cycle corrections that later recover, such as May 2021, creating false tops. Both costs are real and are hidden when the hindsight peak is used instead. - The post-halving 365-day returns ranked at the 0.987, 0.739, 0.836 and 0.347 percentiles. Under no effect, ranks are uniform on 0–1 with variance 1/12. Compute the mean rank, its standard error and the z-score. Then compute the same using only the last two halvings.
Show solution
Mean = (0.987 + 0.739 + 0.836 + 0.347) / 4 = 2.909 / 4 = 0.727. Standard error = √(1/12 / 4) = √0.02083 = 0.1443. z = (0.727 − 0.5) / 0.1443 = 1.57. Last two: mean = (0.836 + 0.347) / 2 = 0.592; SE = √(1/12 / 2) = 0.2041; z = 0.092 / 0.2041 = 0.45. The evidence weakens sharply on the more recent cycles, and both calculations overstate precision because overlapping windows and the long upward trend make ranks less than independent. - A Phillips-Shi-Yu test on Bitcoin detects explosive episodes in 2013, 2017, 2021 and 2024. A writer concludes that this proves the four-year cycle. Write a two-sentence correction.
Show solution
The test shows that Bitcoin’s price was statistically explosive in those episodes, which is evidence of repeated bubble-like behavior, not of a schedule. Whether the episodes are tied to the halving clock is a separate timing question, and four episodes cannot settle it, especially when the timing rule was chosen after seeing them.
References
- Rowe, N. (2024). Bitcoin breaks all-time high of $69,000 as investors flock to spot ETFs ahead of halving. Fortune, March 5, 2024. Source
- Tversky, A. and Kahneman, D. (1971). Belief in the law of small numbers. Psychological Bulletin 76(2), 105-110. Source
- Fischhoff, B. (1975). Hindsight is not equal to foresight: The effect of outcome knowledge on judgment under uncertainty. Journal of Experimental Psychology: Human Perception and Performance 1(3), 288-299. Source
- Gelman, A. and Loken, E. (2013). The garden of forking paths: Why multiple comparisons can be a problem, even when there is no “fishing expedition” or “p-hacking” and the research hypothesis was posited ahead of time. Working paper, Columbia University. Source
- Harvey, C. R., Liu, Y. and Zhu, H. (2016). … and the Cross-Section of Expected Returns. Review of Financial Studies 29(1), 5-68. NBER Working Paper 20592. Source
- White, H. (2000). A Reality Check for Data Snooping. Econometrica 68(5), 1097-1126. Source
- Sullivan, R., Timmermann, A. and White, H. (1999). Data-Snooping, Technical Trading Rule Performance, and the Bootstrap. Journal of Finance 54(5), 1647-1691. Source
- Phillips, P. C. B., Shi, S. and Yu, J. (2015). Testing for Multiple Bubbles: Historical Episodes of Exuberance and Collapse in the S&P 500. International Economic Review 56(4), 1043-1078. Source
- Cheung, A., Roca, E. and Su, J.-J. (2015). Crypto-currency bubbles: an application of the Phillips-Shi-Yu (2013) methodology on Mt. Gox bitcoin prices. Applied Economics 47(23), 2348-2358. Source
- Nosek, B. A., Ebersole, C. R., DeHaven, A. C. and Mellor, D. T. (2018). The preregistration revolution. Proceedings of the National Academy of Sciences 115(11), 2600-2606. Source
- Baquero, C. and Menezes, R. (2026). Deceleration without identification: testing a power-law trend in Bitcoin prices under persistent noise. arXiv:2611.13962 (preprint, with pre-registered forecasts). Source
- Lyons, C. (2026). The debate over Bitcoin’s four-year cycle is over: Benjamin Cowen. Cointelegraph Magazine, March 6, 2026. Source
- mempool.space. Bitcoin block explorer (timestamps of halving blocks 210,000, 420,000, 630,000 and 840,000). Source
- Coin Metrics. Community API documentation (daily PriceUSD used in the worked exercise). Source
Exercise
- Run the code above. Then change the price series to weekly closes and the 12–18 month window to 10–20 and 14–20 months, and record how each result changes.
- Repeat Claim A for Ethereum, using its own launch date and the Bitcoin halving dates, and compare the ranks.
- Replace “the peak” with a mechanical confirmation rule of your own, and rescore Claim B with it.
- Write a dated pre-registration for one cycle rule about the 2028 halving cycle, using the table in this lesson as a template. Keep it, and score it when the data exist.
Check yourself
1. A rule fitted to four cycles matches all four. What is the main reason this is weak evidence?
2. Why does a rule stated in terms of “the cycle peak” contain look-ahead?
3. In the exercise, the 2024 post-halving year ranked at the 35th percentile of all 365-day windows. Which statement fits?
4. Gelman and Loken’s “garden of forking paths” describes:
5. A GSADF test detects several explosive episodes in Bitcoin. What does this establish?
6. Four peaks in Q4 has a chance probability of 1 in 256 under uniform timing. Why is that number too generous to the claim?
7. Five rules fit the first two cycles. Out of sample they score 4 of 10, and one rule scores 2 of 2. What is the risk in citing only that rule?
8. What makes a pre-registered cycle hypothesis stronger evidence than a new chart of past cycles?
Common questions
Why are four cycles not enough to prove a pattern?
A rule built from past cycles matches them by construction, and even four genuine matches give a wide confidence interval, roughly 51% to 100% for four of four. Only cycles observed after a rule is written can test it.
What is the garden of forking paths?
A term from Gelman and Loken for analysis choices made after seeing the data, such as which cycles, prices and windows to use. It creates a multiple-testing problem even when only one test is run.
How can a cycle claim be tested fairly?
Write the rule down with a date before the outcome data exist, define events so they are knowable on the day, compare the result with chance or a base rate, and report every rule tried.
This lesson teaches methods and describes past events. It is not investment advice. Live panels: tracked whale wallets, updated 2026-11-22 07:40 UTC.