Bitcoin Log Regression, Power Law and Moving Average Models Explained
Trend and regression models: moving averages, log curves and power laws. How moving averages, weekly bands, log regression curves, power laws and stock-to-flow are built, what their past record shows, and why curves fitted to a few cycles overfit.
Trend models smooth past prices (moving averages such as the 20-week SMA, 21-week EMA and 200-week SMA) or fit a curve to them, usually a straight line in log price against log time. These curves have described Bitcoin’s history closely, but they are refitted as data arrive, rest on a few cycles, and have missed badly out of sample, as stock-to-flow did in 2022.
You’ll learn
- Compute simple and exponential moving averages by hand and explain why each lags price by a different amount
- Describe the 20-week and 21-week band, the 200-week average, the Pi Cycle Top and the 2-Year MA Multiplier, with their creators and their record on public data
- Fit a log-log regression by least squares, convert its output to a price curve and read its residuals
- Explain why the power-law and stock-to-flow models fit history well, and what out-of-sample tests found
- Show with code how refitting moves a model’s “fair value” and why that matters for any claimed track record
A model with a 95% fit that missed by a factor of five
On March 22, 2019 a pseudonymous analyst known as PlanB published “Modeling Bitcoin’s Value with Scarcity”. The essay regressed the log of Bitcoin’s market value on the log of its stock-to-flow ratio, the existing supply divided by one year of new issuance, using 111 monthly points from December 2009 to February 2019. The fitted line was ln(market value) = 3.3 × ln(SF) + 14.6, with an R² of 95%. Because the May 2020 halving would double the ratio, the model implied a market value of about $1 trillion after it, or roughly $55,000 per coin. For a while the price moved toward the line. By June 2022 Bitcoin traded near $20,000, while the model’s range for 2022 was about $100,000 to $110,000. Vitalik Buterin called the model “harmful”. PlanB replied that it had “a good run” from March 2019 to March 2022. A high R² on past data, built from a handful of halving periods, had said little about the next few years. This lesson takes apart that class of model, and the simpler moving-average tools that traders use alongside it.
Every model in this lesson does one of two things. It smooths past prices into a trend line, or it fits a curve to past prices and treats distance from the curve as information. Neither is a theory of value in the economist’s sense. Both are descriptions of history, and the main question for each is how well that description held on data it had not seen. All price figures below use the free Coin Metrics daily reference price (PriceUSD, one value per day), which differs slightly from exchange closes and intraday highs. Figures run to November 19, 2026.
Moving averages: the arithmetic
Simple moving average (SMA): The unweighted mean of the last n observations: SMAt(n) = (Pt + Pt−1 + … + Pt−n+1) / n. Each new observation enters with weight 1/n, and the observation n periods back drops out entirely.
Exponential moving average (EMA): A weighted mean in which weights fall geometrically with age: EMAt = k·Pt + (1 − k)·EMAt−1. Under the common “span” convention, k = 2 / (n + 1). Every past observation keeps some weight, which shrinks by a factor of (1 − k) each period. This is the same recursion as simple exponential smoothing in forecasting texts.
Both averages describe the price about (n − 1)/2 periods ago. A 20-week SMA is centered about 9.5 weeks in the past. A 200-week SMA is centered nearly two years back.
The matching average age is why a 20-week SMA and a 21-week EMA track each other closely. They differ in shape. The SMA treats week 1 and week 20 equally and then forgets week 21 completely, so it can jump when a large old value drops out. The EMA reacts more to the newest weeks and never fully forgets. Neither is “more correct”. Both lag, and the lag is the price paid for smoothing out noise.
Updating both averages by hand (example numbers)
- A 21-week EMA stands at 70,000. The new weekly close is 63,000. With k = 2/22 = 0.0909: EMA = 0.0909 × 63,000 + 0.9091 × 70,000 = 5,727 + 63,636 = 69,364.
- A 20-week SMA stands at 72,000. The close leaving the window was 80,000 and the new close is 63,000. New SMA = 72,000 + (63,000 − 80,000) / 20 = 72,000 − 850 = 71,150.
- The EMA fell 636 and the SMA fell 850. The SMA fell more because the departing value was high, not because of anything about this week.
The research record on moving-average rules is older than crypto. Brock, Lakonishok and LeBaron (1992) tested moving-average and trading-range rules on the Dow Jones index from 1897 to 1986 and reported results inconsistent with a random walk and several standard time-series models. Sullivan, Timmermann and White (1999) later re-examined that universe with White’s Reality Check bootstrap, which adjusts for the full set of rules a researcher considered rather than the best one alone. The lesson of that sequence is less about whether moving averages “work” than about how easily rule-picking inflates apparent results. Lesson 20 returns to this.
Weekly bands and long averages
Three weekly constructions appear constantly in cycle analysis. They use weekly closes, so a single day’s spike does not count.
| Tool | Construction | How it is read by its users | Origin |
|---|---|---|---|
| 20W SMA + 21W EMA band | The zone between a 20-week SMA and a 21-week EMA of weekly closes | Price above the band as a bull-market regime, below as a bear-market regime | Widely associated with Benjamin Cowen (Into The Cryptoverse). Public implementations credit it only as “popularized in crypto analysis” |
| 200-week SMA | Mean of the last 200 weekly closes, about 3.8 years | A very slow floor; closes below it have been rare | A generic long average, not owned by anyone |
| 50-week averages | 50-week SMA or EMA, and crosses with shorter averages | Medium-term regime | Generic |
Descriptions of use are what practitioners say, not established findings.
Cowen’s framework, as reported in his recurring “bull market support band” videos and in press coverage, treats the band as support while a bull market lasts and as resistance during a bear market, and gives weight to weekly closes rather than intraweek moves. In 2026 coverage he described the same band acting as resistance in a midterm year. This is one practitioner’s approach. Its claims can be checked against data, which is the fair way to evaluate it.
How often the band changed hands, 2015 to November 2026
- Data: Sunday Coin Metrics reference prices as weekly closes. A week counts as “above” if the close is above both lines, “below” if under both. Weeks inside the band keep the prior state.
- From January 2015 to November 15, 2026 there were 41 changes of state, so 40 completed regimes.
- 15 of those 40 regimes lasted three weeks or fewer. Examples: February to March 2018 (five changes in six weeks), October 2023, July to September 2024 and October to November 2025.
- Long regimes did exist: above the band for 120 weeks from October 2015 to February 2018, above it for 54 weeks from May 2020 to May 2021, and below it from December 2021 to January 2023 apart from two weeks in April 2022.
- The close on November 15, 2026 was 65,189, below the 20W SMA (73,843) and the 21W EMA (75,500).
The 200-week SMA is slower still. On the same weekly data, closes below it happened in 58 of the 655 weeks for which it exists: briefly in 2015, one week in March 2020, during 2022–23, and in mid- and late 2026. The weekly close of November 15, 2026 (65,189) was below the 200-week SMA (about 67,970). The often-quoted line that Bitcoin “never closes below the 200-week average” was not true on this data, and the rarity of closes below it says nothing about what happens after one.
Ratio-of-averages tools: Pi Cycle Top and the 2-Year MA Multiplier
Two well-known indicators were created by analyst Philip Swift, founder of the LookIntoBitcoin charts site. They are often discussed by cycle analysts, including Cowen, but are not his tools.
Pi Cycle Top indicator: Flags a possible cycle top when the 111-day SMA of price crosses above two times the 350-day SMA. The ×2 multiplies the price level, not the window. The name comes from 350 / 111 ≈ 3.153, close to π. Promoters state that it picked out past cycle highs to within about three days.
2-Year MA Multiplier: Plots the 730-day SMA of price and five times that average. Price below the 2-year average has been read as historically cheap, and price above five times it as historically stretched.
| Cycle | Pi Cycle cross (111D above 2×350D) | Cycle high (daily reference) | Price above 5× 2-year SMA? |
|---|---|---|---|
| 2013 | April 6 and December 5, 2013 | April 9 local high; December 4 peak ($1,135) | Yes, March–April and October–November 2013 |
| 2017 | December 16, 2017 | December 16 ($19,641) | Yes, September 2017 to January 2018 |
| 2021 | April 12, 2021 | April 13 local high ($63,446); November 8 peak ($67,542) | No. Peak ratio 4.88 (February 21, 2021) |
| 2024–25 | No cross. The ratio peaked at 0.74 (mid-2024) | October 6, 2025 ($124,824) | No. Peak ratio 2.48 (March 13, 2024) |
Computed from Coin Metrics daily data. A cross in July 2011 is excluded because the 350-day average only became defined that month. Intraday highs on exchanges were higher than these daily reference prices.
The record is good for 2013 and 2017, and weak after that. The 2021 cross matched the April high, not the November high that marked the cycle. In the 2024–25 cycle neither tool reached its top condition at all. One reading is that each tool was designed with the 2013 and 2017 cycles in view, so its calibration fitted those two well. Another is that volatility has fallen, so fixed multiples such as ×2 and ×5 are now too far away. Both readings agree on the practical point: a threshold tuned on two or three episodes has very few independent tests behind it.
Log regression fair-value curves
Bitcoin’s price has risen by about six orders of magnitude, so a straight line in price makes no sense. A straight line in log price against log time is the usual alternative. It says growth slows over time in a specific way: the price is proportional to a power of the asset’s age.
Pt is price on day t, dt is days since a chosen origin (commonly the genesis block, January 3, 2009), a and b are fitted constants, and et is the residual. A residual of +0.3 means price was 100.3 ≈ 2 times the curve.
Ordinary least squares with x = log10 d and y = log10 P. R² = 1 − Σe² / Σ(y − ȳ)². Bands are usually drawn as the fitted line plus or minus a constant in log terms.
A three-point fit by hand (example numbers)
- Points (x, y): (3.0, 1.0), (3.5, 4.0), (4.0, 6.0). Means: x̄ = 3.5, ȳ = 3.667.
- Σ(x − x̄)² = 0.25 + 0 + 0.25 = 0.50. Σ(x − x̄)(y − ȳ) = (−0.5)(−2.667) + 0 + (0.5)(2.333) = 1.333 + 1.167 = 2.50.
- b = 2.50 / 0.50 = 5.0. a = 3.667 − 5.0 × 3.5 = −13.833.
- At x = 3.8 (about 6,300 days): y = −13.833 + 19.0 = 5.167, so the curve value is 105.167 ≈ 147,000.
Cowen’s framework, as described in summaries of his videos, fits a fair-value band through what he calls non-bubble data, adds parallel bands above it, and argues from them that each cycle peak reaches a lower band than the one before. Log-growth curves for Bitcoin predate his work, so this is best described as an approach he popularized and has used publicly for years, with coefficients he has not published. In November 2023 he argued that each cycle top had landed about two and a half bands below the previous one, and that the same step applied in late 2025 pointed to a top “just north of $100,000”; the same report also cites about $111,000. The October 2025 peak came in the fourth quarter of the post-halving year, as his framework expected, at about $124,800 on daily reference data (about $126,000 intraday as reported). The timing matched. The level was above his stated range. He has publicly named a different miss, a late-2023 expectation that the price would stall below $35,000.
Fitting the same kind of curve to all daily data from July 2010 to November 19, 2026, with the genesis block as origin, gives b = 5.63, R² = 0.962 and a residual standard deviation of 0.301 in log10 units. That last number means the typical gap between price and curve is a factor of about 2 in either direction. This is a full-sample fit, so it uses information that did not exist at the time of the early points.
Diminishing returns
| Cycle peak (date) | Price | Curve value that day | Price / curve | Multiple of prior peak |
|---|---|---|---|---|
| June 8, 2011 | $29 | $1.90 | 15.3 | n/a |
| December 4, 2013 | $1,135 | $102 | 11.1 | 39× |
| December 16, 2017 | $19,641 | $2,972 | 6.6 | 17× |
| November 8, 2021 | $67,542 | $22,750 | 3.0 | 3.4× |
| October 6, 2025 | $124,824 | $101,608 | 1.2 | 1.8× |
Full-sample log-log fit on Coin Metrics daily data. Cycle lows on the same curve: January 14, 2015 at 0.55, December 15, 2018 at 0.59, November 9, 2022 at 0.45, and June 30, 2026 (the lowest daily price after the 2025 peak, through November 19, 2026) at 0.45.
Each peak sat closer to the curve than the last, and each peak was a smaller multiple of the one before. This pattern, usually called diminishing returns, is the most robust-looking regularity in the lesson: it appears whether peaks are measured against a curve, against the prior peak or as drawdown depth. Peak-to-trough declines were about 93% (2011), 85% (2013–15), 84% (2017–18), 77% (2021–22) and 53% from October 2025 to June 2026.
Three cautions apply. First, it rests on four or five points, which is too few to estimate the rate of decay with any precision. Second, it is partly built in: a curve fitted to all the data is pulled toward recent prices, which lowers the ratio for recent peaks. Third, a smooth decline in the ratio is a description, not a mechanism. Larger market value, lower volatility, new kinds of holders and simple maturation are all candidate explanations, and they are hard to tell apart with this little history.
The power-law model and its 2026 evaluation
The “power law” is the same log-log regression, promoted as a structural description of Bitcoin’s growth rather than a convenient curve. It is most associated with the Italian physicist Giovanni Santostasi, whom a 2024 Cointelegraph report describes as the person who identified the relationship for Bitcoin. The most recent formal account is a 2026 preprint by Santostasi and Perrenod, which ties the exponent to network adoption. Supporters treat the corridor around the line as a long-run range. In the same 2024 report, Santostasi called the model not a perfect predictive tool and said a large, sustained move below or above the trend would disprove it.
Carlos Baquero and Raquel Menezes put the claim through formal tests in a May 2026 arXiv preprint, “Bitcoin’s Power Law: Weak Structure, Strong Forecasts”. It is not peer-reviewed. Their findings, on daily data from August 2010 to March 2026:
| Question | Finding |
|---|---|
| How good is the fit? | Exponent 5.64 with the genesis block as origin, R² 0.96, residual σ 0.30 in log10 |
| Is the exponent a stable constant? | No. Moving the time origin changes it from 5.65 to 16.49 across the range tested |
| Is it uniquely the right curve? | No. Residual diagnostics cannot separate it from a stack of S-shaped curves, which fits better in sample |
| Does it forecast? | In 11 yearly walk-forward tests, it beat standard baselines at 12–24 month horizons (p < 0.05). At 1–3 months a naive “no change” forecast was better |
| Size of the 12-month edge | Mean error in log10 price of 0.375 versus 0.510 for no change. At 24 months, 0.349 versus 0.675 |
Baquero and Menezes (2026), arXiv 2605.21316. The authors note that 11 cutoffs give tests of moderate power.
The same authors posted a second preprint in November 2026, “Deceleration without identification”. Using weekly data and an explicit model of persistent deviations with a four-year cycle, they could not reject a log-age coefficient of 5.7. They also could not reject a Gompertz curve with a ceiling, and they concluded that whether prices revert to the trend cannot be determined from the data. They pre-registered and hashed forecasts for 2027 to 2030 to be scored later. Whatever those scores turn out to be, the method is the right one: a claim fixed in public before the data arrive.
Two points survive both papers. A straight line on log-log axes is necessary for a power law but not sufficient, because regressions of a trending series on time produce high R² even when no stable relationship exists. And the forecasting result is real but modest: a typical 12-month miss of 0.375 in log10 is still a factor of about 2.4.
Stock-to-flow: why it failed out of sample
Stock-to-flow ratio (SF): Existing stock divided by annual new production. For Bitcoin, the flow falls by half at each halving, so SF roughly doubles in a step every four years.
The model in the opening story regressed log market value on log SF. Several problems explain why a 95% in-sample fit did not carry forward. SF is almost a deterministic function of time, so the regression was close to a regression of value on time, and its apparent explanatory power partly reflected the shared upward trend in both series. The model has no demand side: issuance is known decades in advance, so any information in it should already be in the price. The doubling at each halving creates a few big steps in the explanatory variable, so its effective sample is a handful of halving periods, not 111 months. Work cited by Baquero and Menezes reports that the stock-to-flow variable was about 80.6% correlated with log time and lost significance once time effects were controlled, which is the standard warning sign of a spurious regression. The 2022 miss was then a factor of about five relative to the model’s range.
Overfitting, refitting and how to evaluate a curve
Every fitted curve faces the same two problems. Overfitting means choosing a model, origin, window and band width that match the past closely, including its noise. Refitting means the curve is re-estimated as data arrive, so the “fair value” for a given date keeps moving. A chart drawn today with today’s fit always looks better than the curve that existed at the time. The code below shows both effects on a small labeled dataset: Bitcoin’s reference price on January 1 of each year, 2011 to 2026. It runs with standard Python and no packages.
Read the refit lines carefully. The curve value for the same date, January 1, 2026, was about $857,000 using data through 2015, $310,000 using data through 2018, $180,000 using data through 2021 and $126,000 using all of it. The slope fell each time and R² rose each time. A chart that shows only the final curve hides that history, and it makes the model look as if it always knew where the price would be. The out-of-sample block is a fairer test: fitted through 2020, the curve missed 2023 by a factor of 100.48 ≈ 3. Its out-of-sample error was smaller than its in-sample spread here only because the earliest years were the noisiest, which is a reminder that one split proves little. The full daily fit gives a slope of 5.63 rather than 5.79, because this small dataset samples one day per year.
| Practice | Why |
|---|---|
| Freeze the model and record its output with a date before scoring it | Only predictions made in advance test the model |
| Use walk-forward evaluation: refit only on data before each forecast date | Removes look-ahead from the fit itself |
| Compare with a naive baseline, such as no change or a random walk with drift | A curve that cannot beat “no change” has not earned its complexity |
| Report errors in log terms and as a factor | A 0.3 log10 error is a factor of 2, which matters more than a high R² |
| Count the degrees of freedom: origin, window, exclusions, band width | Each choice is a parameter, even when it is not called one |
| Check residual autocorrelation | Long runs of positive or negative residuals mean far fewer independent observations |
Forecast-evaluation practice adapted from standard time-series methods, such as cross-validation on a rolling forecast origin.
Common mistakes
- Treating a curve’s “fair value” as a price the market owes. It is a fitted line, and it moves when refitted.
- Quoting R² as evidence of predictive power for a trending series.
- Judging a top indicator on the two cycles it was designed around, and not on the cycles after.
- Calling a weekly-band crossing a regime change without counting how many crossings reversed within weeks.
- Attributing every cycle tool to one well-known analyst. The Pi Cycle Top and 2-Year MA Multiplier are Philip Swift’s, and log-growth curves predate the current debate.
- Showing a model with today’s parameters across past cycles and calling it a track record.
Is Bitcoin’s long-run trend a law, or a curve fitted to one history?
The case for a structural trend. The log-log relationship has held, within a wide corridor, for more than 15 years and across several very different market regimes. Baquero and Menezes found that it beat standard baselines at 12–24 month horizons in walk-forward tests, and it was the only asset they examined where no single-component growth curve improved on it. Supporters argue that adoption processes in networks often follow power-law growth, which gives a reason for the form. The case against. The exponent depends heavily on the chosen time origin, which is not what a physical law looks like. Other curves, including ones with a ceiling, fit as well. Residuals are highly persistent, so the effective sample is small, and regressions of trending series on time can produce a high R² with no real relationship. Stock-to-flow also had a high R² and an appealing story, and it failed. Where that leaves a student. Both sides accept that the curve has described the past well and that its residuals are large. The disagreement is about whether the past constrains the future. That can only be settled by forecasts fixed in advance and scored later, which is why the pre-registered 2027–2030 forecasts in the November 2026 paper are more informative than any new chart of the existing data.
Problem set
- A 21-week EMA is 50,000. The next three weekly closes are 44,000, 46,000 and 48,000. Compute the EMA after each week with k = 2/22.
Show solution
k = 0.09091. Week 1: 0.09091 × 44,000 + 0.90909 × 50,000 = 4,000.0 + 45,454.5 = 49,454.5. Week 2: 0.09091 × 46,000 + 0.90909 × 49,454.5 = 4,181.8 + 44,958.6 = 49,140.4. Week 3: 0.09091 × 48,000 + 0.90909 × 49,140.4 = 4,363.6 + 44,673.1 = 49,036.7. All three closes were below the EMA, so it fell each week, but slowly: about 1.9% in total. - Show that the 111-day and 350-day windows in the Pi Cycle Top give a ratio close to π, and explain why the factor of 2 is not part of that ratio.
Show solution
350 / 111 = 3.1532, versus π = 3.1416, a difference of about 0.4%. The factor of 2 multiplies the 350-day average’s price level: the condition is SMA111 > 2 × SMA350, meaning the short average is more than double the long one. Doubling the window to 700 days would be a different indicator. The closeness to π is a naming coincidence, not evidence of anything. - A log-log fit gives a = −17.0 and b = 5.8, with time in days since January 3, 2009. Compute the curve value at day 6,000. If the residual standard deviation is 0.30, what range covers price within one standard deviation?
Show solution
log10(6,000) = 3.7782. y = −17.0 + 5.8 × 3.7782 = −17.0 + 21.913 = 4.913. Curve value = 104.913 ≈ 81,800. One standard deviation is a factor of 100.30 = 1.995, so the range is about 41,000 to 163,000. The top of that range is four times the bottom, which is what “a good fit” means here. - Use the diminishing-returns table: peak-to-prior-peak multiples of 39, 17, 3.4 and 1.8. A student fits a straight line to the logs of these multiples against cycle number (1, 2, 3, 4) and extends it to cycle 5. What is wrong with treating the result as an expected value?
Show solution
The logs are 1.59, 1.23, 0.53 and 0.26. A line through them falls by about 0.47 per cycle, and extending it gives a log below zero, meaning a peak below the prior one, which shows how fragile the exercise is. Problems: four points give a slope with a huge standard error; the decay need not be linear in logs (it could level off or collapse); the points are not independent draws; and the choice of which peaks count (whether 2011 is included, whether 2019 is a peak) was made after seeing the data. The extension is arithmetic, not evidence, and it is not a forecast. - The code above shows that the curve value for January 1, 2026 was $857,442 when fitted through 2015 and $125,600 when fitted through 2026. Someone shows a chart of the 2026 fit and says the model “has tracked Bitcoin for 15 years”. Write a one-paragraph reply.
Show solution
The chart shows a curve fitted to all 15 years, so it tracks them by construction. The test is what the curve said before each period. Fitted through 2015, the same method put the January 2026 curve value near $857,000, nearly seven times the value from the full fit and nearly ten times the actual price of $88,684. The slope has fallen at each refit. A fair track record would show the curve as it stood at each date, scored on what came after, against a simple baseline. The 2026 walk-forward study by Baquero and Menezes does this and finds a modest advantage at 12–24 months, with typical errors of a factor of about 2.4. - A stock-to-flow regression has R² = 0.95 on 111 monthly points. Give three reasons the effective evidence is much weaker than 111 independent observations.
Show solution
(1) The explanatory variable changes in a few large steps at halvings, so there are only a handful of distinct levels of it, roughly one per halving period. (2) Both series trend upward over time, so much of the fit comes from the shared trend, the classic spurious-regression setting, and SF is close to a function of time. (3) Monthly residuals are strongly autocorrelated, so consecutive months add little new information. Together these leave something closer to three or four independent observations, which is consistent with the model failing in 2022.
References
- PlanB (2019). Modeling Bitcoin’s Value with Scarcity. Medium, March 22, 2019 (reposted copy). Source
- Asmakov, A. (2022). Vitalik Buterin: Stock-to-Flow Bitcoin Price Model ‘Really Not Looking Good Now’. Decrypt, June 21, 2022. Source
- Baquero, C. and Menezes, R. (2026). Bitcoin’s Power Law: Weak Structure, Strong Forecasts. arXiv:2605.21316 (preprint). 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). Source
- Brock, W., Lakonishok, J. and LeBaron, B. (1992). Simple Technical Trading Rules and the Stochastic Properties of Stock Returns. Journal of Finance 47(5), 1731-1764. 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
- Hyndman, R. J. and Athanasopoulos, G. (2021). Forecasting: Principles and Practice, 3rd edition. OTexts. Sections 5.10 (time series cross-validation) and 8.1 (simple exponential smoothing). Source
- Blockchain.com. Pi Cycle Top Indicator (chart and description, credits Philip Swift). Source
- mljones (2019). 2-Year MA Multiplier, a recreation of Philip Swift’s LookIntoBitcoin chart. TradingView open-source script. Source
- RiccoNewman (2025). Bull Market Support Band (20W SMA/21W EMA). TradingView open-source script. Source
- Richardson, A. (2023). Here’s a Potential Bitcoin Price Target for Late 2025, According to Crypto Analyst Benjamin Cowen. The Daily Hodl, November 13, 2023 (secondhand report of Cowen’s view). Source
- Lyons, C. (2026). The debate over Bitcoin’s four-year cycle is over: Benjamin Cowen. Cointelegraph Magazine, March 6, 2026. Source
- Cointelegraph (2024). Debate rages over Bitcoin power law as critics label it a ‘magic trick’. August 9, 2024. Source
- Into The Cryptoverse. About us (Benjamin Cowen’s company page, with a link to his YouTube channel). Source
- Coin Metrics. Community API documentation (daily PriceUSD used for all figures in this lesson). Source
Exercise
- Download daily Bitcoin PriceUSD from the free Coin Metrics community API, as in Lesson 9’s code, and compute the 111-day and 350-day SMAs and the 730-day SMA.
- Find every date the 111-day average crossed above twice the 350-day average, and compare each with the highest price within 60 days on either side.
- Fit the log-log regression on data through December 31, 2019 only. Record its curve value for every January 1 from 2020 to today, and compute the base-10 log error each year.
- Compare those errors with a naive forecast that uses the price one year earlier. Write two sentences on which did better and how many years the comparison rests on.
Check yourself
1. A 20-week SMA and a 21-week EMA have nearly the same average data age. Why do they still diverge at times?
2. The Pi Cycle Top crossed on April 12, 2021, and not near the November 2021 peak. Which conclusion is best supported?
3. A log-log fit has R² = 0.96. What does this establish?
4. Fitted through 2015, the curve value for January 2026 was about $857,000. Fitted through 2026, about $126,000. What is the main lesson?
5. Baquero and Menezes found the power law beat baselines at 12–24 months but lost to a no-change forecast at 1–3 months. Why might both be true?
6. Why did the stock-to-flow regression’s 95% R² overstate the evidence?
7. A report says Bitcoin’s peaks have reached a lower band each cycle and draws the next band. Which statement fits this lesson?
8. Which attribution is correct?
Common questions
What is the difference between an SMA and an EMA?
A simple moving average weights the last n prices equally and then drops each one entirely. An exponential moving average weights recent prices more, with weights shrinking geometrically, using k = 2/(n+1). Both describe the price about (n−1)/2 periods ago.
Who created the Pi Cycle Top and the 2-Year MA Multiplier?
Both were created by analyst Philip Swift of LookIntoBitcoin. The Pi Cycle Top compares the 111-day average with twice the 350-day average; the 2-Year MA Multiplier plots the 730-day average and five times it.
Why did the stock-to-flow model fail?
Its key variable changed in a few steps at halvings and trended with time, so its 95% fit rested on very few independent observations. In 2022 Bitcoin traded near $20,000 against a model range of about $100,000 to $110,000.
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.