Bitcoin Price Model Glossary: Statistics Terms in Plain English

Every term here shows up in arguments about Bitcoin price models, and most of them get used loosely enough to hide the disagreement rather than settle it.

Each entry gets a plain-language definition, the axis of our six-part framework it serves, and one real worked example — no toy illustrations. Every number below comes from published work, cited at the bottom.

This page is permanent and gets updated as the series adds terms. Alphabetical.


AIC (Akaike Information Criterion)

AXIS 2 · AXIS 5

A score for comparing models that rewards fitting the data and penalizes using more parameters. Lower is better. It answers “which of these models best explains what I’ve seen,” which is not the same as “which is true” or “which will forecast well.”

Worked example: On Bitcoin’s price history, a three-component sigmoid model beats the power law by a huge AIC margin. Case closed? No — researchers generated fake data from a known, genuine power law and ran the same comparison. The sigmoid won by an even bigger margin on data they knew was a power law. AIC couldn’t tell the two apart. A test that flags the true model as wrong isn’t a test.1


Cointegration

AXIS 2

Two wandering series are cointegrated if some combination of them stays stable over time. It’s the standard defense against spurious regression — if two series are cointegrated, the relationship between them may be real rather than an artefact of both drifting.

Worked example — and the best live methodological dispute in the field. In 2022 Tim Stolte argued the power law fails the Engle-Granger cointegration test, so the fit is spurious.2 In 2024 Harold Christopher Burger and Giovanni Santostasi replied that the test doesn’t apply at all: “Both variables must be stochastic for cointegration to be possible. Time is emphatically not a random variable, but a deterministic variable.”3 That’s a serious objection — you can’t test whether the calendar wanders, because it doesn’t.

Then in 2026, cointegration reappears on the proponents’ side. Stephen Perrenod ran an Engle-Granger test on price against time, reported p = 0.026, and presented passing it as a major statistical validation.4 And the peer-reviewed Santostasi–Perrenod paper lists among its headline findings: “Stationary residuals, cointegration and out-of-sample fit rule out a spurious law.”10

So one argument says the test cannot apply to time; another uses passing it as evidence. Both come from the same side of the debate, and Santostasi is an author on both.

We are not calling that a contradiction — there may be a reconciliation in the full papers that we have not found, and the second of them cleared peer review. We flag it because it is exactly the kind of open question a reader should be able to see and chase down, rather than one that gets smoothed over in a summary. If you find the reconciliation, send it and this entry gets updated.


Diebold–Mariano test

AXIS 1

A test for whether one forecasting model is significantly better than another, rather than better by an amount that could be luck.

Worked example: The power law has lower forecast error than the naive baseline at 12, 18, and 24 months. But the Diebold–Mariano p-values are 0.196, 0.093, and 0.079 — none of them significant at the 5% level. The power law’s long-horizon advantage over doing nothing is real in the point estimates and not statistically established. The authors say why: with only eleven test windows, the test has limited power to detect anything.1

Anyone quoting the power law’s forecasting win should quote this too.


Exponential vs. power growth

AXIS 1 · AXIS 3

Exponential growth adds the same percentage every year, so it plots as a straight line on a chart with a log price axis and an ordinary time axis. Power growth adds a shrinking percentage: under a power law in time, each year’s growth depends on how old the thing already is, and the rate falls as it ages. Both curves can fit the past. They disagree sharply about the future.

Worked example: With Santostasi and Perrenod’s published exponent of 5.690,10 Bitcoin’s trend growth for the twelve months from October 2026, when it is 17.8 years old, works out to about 36.5%. Ten years on it is about 22.3%; thirty years on, about 12.5%. (Our arithmetic: ((age + 1) / age)5.690 − 1.) Matthew Mezinskis described the current trend rate as “about 40% CAGR at the moment” in August 2026.13 These are statements about the model, not forecasts. They hold exactly as long as the power law does.


Falsifiability

AXIS 4

Whether a claim can be shown wrong — and, in practice, whether its author said in advance what would do it. A model that survives by reinterpretation isn’t being tested.

Worked example: On 23 June 2021, PlanB wrote of stock-to-flow: “I will call s2f invalidated if we have not reached 100K by Dec this year.”5 Specific, dated, public, made while the outcome was genuinely uncertain — an exemplary pre-commitment.

Bitcoin closed 2021 at $46,211. On 18 December, six days before his own deadline, he wrote: “Will I ditch S2F model if this does not happen? Nah.”6

The pledge was the right kind of statement. What happened next is why we keep a Falsification Ledger rather than trusting anyone to keep their own.


Granger causality

AXIS 3

A test of whether past values of one series help predict another beyond what the second series already predicts about itself. It is not proof of causation — the name oversells it. It tests precedence: does A reliably move first?

Worked example: The power law’s proposed mechanism is that adoption drives price. Tested against five on-chain metrics — hash rate, active addresses, transactions, difficulty, UTXO count — price leads every one of them, by 13 to 61 days, with the Granger asymmetry running about 2.5 to 1 in the price → adoption direction.1

The honest caveat, which the authors state themselves: the peak correlations are only 0.04 to 0.10. Small. The direction is consistent; the strength is modest. That’s the finding, and both halves belong in any summary of it.


Lognormal distribution

AXIS 6

A distribution where the logarithm of the values is normally distributed. It looks a lot like a power law on a chart and is very commonly the better fit.

Worked example: Applying the standard Clauset–Shalizi–Newman protocol to Bitcoin’s tail-relevant data, eight of eleven tests rejected the power law outright — and a likelihood-ratio test preferred lognormal in all eleven, by margins the authors describe as overwhelming.1

This isn’t a Bitcoin embarrassment. Broido and Clauset applied the same protocol to about a thousand real-world networks and found 88% were fit at least as well by a lognormal.7 Power laws are the most over-claimed structure in applied statistics.


Naive baseline

AXIS 1

The “no model” forecast: tomorrow’s price is today’s price. It’s the bar every forecasting model must clear, and it clears a surprising number of them.

Worked example: Across ten competing models at a one-month horizon, the naive forecast had the lowest error — 0.087, against 0.367 for the power law and 0.894 for a degree-9 polynomial.1 This isn’t a crypto quirk. Meese and Rogoff found the same thing in currency markets in 1983.8

If someone shows you a short-horizon Bitcoin price model, the first question is whether it beats doing nothing.


Out-of-sample

AXIS 1

Data the model didn’t see when it was built. Performance on data it did see is called in-sample, and it’s nearly worthless as evidence.

Worked example: A three-component sigmoid model beats the power law decisively on Bitcoin’s price history. Then, forecasting forward, it’s roughly twice as inaccurate at 24 months.1 Best in-sample, near-worst out-of-sample — the same model, on the same data, judged two ways.


Overfitting

AXIS 1 · AXIS 5

Fitting the noise instead of the signal. An overfit model describes the past beautifully and falls apart the moment it steps past it.

Worked example — the best one in the literature. A degree-9 polynomial tracks Bitcoin’s history closely. Asked to forecast forward, its error goes from 0.894 at one month to 6,773 at twenty-four.1 Not a bad prediction — a number that isn’t a price.

When someone shows you a model hugging the historical curve, that closeness is the warning sign, not the evidence.


p-value

AXIS 1 · AXIS 6

Roughly: the probability of seeing a result this strong if nothing real were going on. Small means “unlikely to be luck.” The conventional cutoff of 0.05 is a convention, not a law of nature.

Worked example: At 24 months the power law beats a cubic B-spline at p = 0.051 and a multi-sigmoid at p = 0.049. One “fails,” one “passes.” The difference between them is nothing. Treat values near the threshold as what they are — inconclusive — rather than as a verdict that flips at a round number.1


Power law

THE SUBJECT OF MOST OF THIS SERIES

A relationship where one quantity scales as another raised to a fixed exponent. On a chart with both axes logarithmic, a power law is a straight line — which is why log-log plots are everywhere in this debate.

Worked example: Bitcoin’s price fitted against time since the genesis block gives an exponent of about 5.64, with R² = 0.96.1 Independent fits by others land at 5.674, 5.690, 5.723, and 5.88 — the spread comes from different date windows and different choices about where to start the clock.2,3,4,10


Quantile regression

AXIS 4 · AXIS 6

Ordinary regression fits one line through the middle of the data. Quantile regression fits a separate line for each percentile: the 5th, the 50th, the 95th. A price on the 5th-percentile line has been that low relative to trend about 5% of the time. Unlike sigma bands, it does not assume errors are symmetric or bell-shaped, which matters for an asset whose crashes and blow-off tops are not mirror images.

Worked example: Porkopolis Economics now draws its power-law bands this way. The chart page says the percentiles “are based on a pure quantile regression” and that “this analysis does not assume a bell curve.”9 In August 2026 Mezinskis said price sat “under 50% of the trend itself,” in what he called “this zeroth percentile.”13 In October 2026 he put it “a little bit below 8%, the eighth percentile,” and spelled out the reading: “that means 92% of the time of all observations that we’ve had in the Bitcoin world, you know, the price has been higher relative to trend.”14 Read that carefully. A historical percentile tells you how rare a reading has been. It cannot tell you the reading won’t go lower; a new low simply redraws the bottom band.


R² (coefficient of determination)

AXIS 2 · AXIS 6

How much of the variation in the data the model accounts for, from 0 to 1. Widely quoted, and the single most over-read number in this field. A high R² does not mean a model is correct, robust, or predictive.

Worked example: Shift the power law’s time origin across a range of reasonable choices and the fitted exponent runs from 5.65 all the way to 16.49 — nearly triple. R² stays above 0.91 across the entire range.1 Every one of those fits looks excellent. They cannot all be right.

R² tells you the line goes through the dots. It doesn’t tell you the line means anything.


Residual

AXIS 5 · AXIS 6

For one observation, what actually happened minus what the model said. On log-price models it is measured in logs, so a price at half the trend line has a residual of about −0.30 in base-10 logs whether Bitcoin is at $500 or $50,000. Sigma bands, percentile bands, and every “cheap versus trend” claim are statements about residuals.

Worked example: Critics point to Bitcoin’s residuals, with their long runs above and below trend, as evidence the power law is misspecified. Baquero and Menezes tested that argument by generating prices from a true power law plus realistic noise and running the same diagnostics. “Bootstrap p-values exceed 0.50 for each diagnostic,” and in over half of the synthetic runs the residuals showed “‘pathology’ of the same magnitude as observed on real Bitcoin.”1 On data this autocorrelated, ugly residuals are not evidence against the model.


Scale invariance

AXIS 3 · AXIS 5

A relationship is scale-invariant if stretching the input by a fixed factor always stretches the output by the same factor, wherever you start. Under a power law, multiplying Bitcoin’s age by some factor multiplies the trend price by that factor raised to the exponent, whether the age goes from 4 to 8 years or from 15 to 30. Exponential growth doesn’t work this way. It multiplies by a fixed amount per fixed span of time.

Worked example: With an exponent near 5.69, the trend doubles each time Bitcoin’s age grows by about 13% (21/5.69 ≈ 1.13). At age 5 that took about eight months; at age 17.8 it takes about 2.3 years. Proponents offer scale-invariance tests as evidence for the power law. Baquero and Menezes re-ran all four on real prices and on synthetic prices from a three-sigmoid model that is not a power law: pair-ratio slope 5.69 vs. 5.74, direct collapse 5.62 vs. 5.65. Every test came back the same for the rival.1 A test that a competing model also passes can’t count as evidence for either.


Sigma (σ) bands

AXIS 4

The channel drawn above and below a trend line, measured in standard deviations of the model’s error. If the errors were bell-shaped, about 68% of observations would sit inside the 1σ band and about 95% inside 2σ. Bitcoin’s errors are not guaranteed to be, which is why some power-law charts now draw percentile bands instead (see quantile regression).

Worked example — and this one matters right now. Santostasi and Perrenod’s published falsification criterion F1 puts model invalidation at 3σ below the fit, which they place near $10,000.10 The “power law floor” quoted in the headlines when this glossary opened on 2 August 2026 — Newhedge’s rail, then $59,386 — sat roughly 1.25σ below the trendline.11

Those are very different lines. A break of the second is not the event the first describes. If you take one thing from this glossary, take this: know which line is a chart band and which one is the model’s stated boundary condition.


Spurious regression

AXIS 2

When two unrelated series both trend over time, standard regression will report a strong relationship between them that isn’t there. R² climbs toward 1 as you add data, no matter how meaningless the pairing.

Worked example: This is Tim Stolte’s central objection to the power law: “Since both of the previously mentioned conditions hold, we are dealing with a so-called spurious regression… Hence, the increasing R-squared as reported in the article does not say anything about the fit of this model.”2 The counter-argument is under cointegration. Read both.


Stationarity

AXIS 2 · AXIS 3

A series is stationary if its statistical behavior doesn’t drift over time. Prices generally aren’t. Most regression tools assume stationarity, which is why applying them to raw price series causes so much trouble.

Worked example: Before testing whether price leads adoption, researchers first-differenced every series — converting levels into changes — specifically to remove the shared trend that would otherwise manufacture a relationship out of nothing.1 That single step is the difference between a causality result you can use and one you can’t.


Survivorship bias

AXIS 6

What got excluded, and whether excluding it flattered the result. Sometimes the exclusion is sound; sometimes it’s the whole finding.

Worked example: Stock-to-flow’s variable turns out to be 80.6% correlated with the log of time since the genesis block, and loses statistical significance once you control for time directly.12 It wasn’t measuring scarcity so much as measuring the calendar, with extra steps and a better story.


Walk-forward validation

AXIS 1

The honest way to test a forecasting model. Stand at a past date, use only data available then, forecast forward, check against what happened. Repeat at many dates.

Worked example: Eleven yearly cutoffs from 2014 to 2024, six horizons from one month to two years, ten competing models.1 It’s the procedure that produced nearly every result on this page — and it’s the one thing most published Bitcoin models have never been subjected to.


Missing a term?

This page grows with the series. If a term you’ve hit isn’t here, send it and it gets added with a worked example.

Related: How to Read a Bitcoin Price Model — the six-axis framework these terms serve · The Falsification Ledger — dated predictions and how they resolved.

Sources

  1. Baquero, C. & Menezes, R. (2026). Bitcoin’s Power Law: Weak Structure, Strong Forecasts. arXiv:2605.21316v1. Preprint; not peer-reviewed.
  2. Stolte, T. (2022). Bitcoin’s power-law corridor debunked. Amdax, 2 September 2022.
  3. Burger, H.C. & Santostasi, G. (2024). Bitcoin’s power-law really debunked? Quantodian Publications, 8 February 2024.
  4. Perrenod, S. (2026). Bitcoin’s Power Law Passes a Major Statistical Test. Substack, 5 June 2026.
  5. PlanB (@100trillionUSD), 23 June 2021.
  6. PlanB (@100trillionUSD), 18 December 2021.
  7. Broido, A.D. & Clauset, A. (2019). Scale-free networks are rare. Nature Communications 10, 1017.
  8. Meese, R. & Rogoff, K. (1983). Empirical exchange rate models of the seventies. Journal of International Economics 14(1–2).
  9. Mezinskis, M. The chart. Porkopolis Economics.
  10. Santostasi, G. & Perrenod, S. (2026). A mechanistic derivation of the Bitcoin price power law, §10. Nonlinear Science 8, 100172. DOI: 10.1016/j.nls.2026.100172, Elsevier. Peer-reviewed. Preprint version: Zenodo, DOI 10.5281/zenodo.19387099.
  11. Newhedge Bitcoin Power Law Model tracker, retrieved 2 August 2026.
  12. Shelton, A. (2024). Bitcoin return prediction. Journal of Risk and Financial Management 17(10):443.
  13. Mezinskis, M. Interview on TFTC #774, published 5 August 2026. Quotes taken from the episode transcript at tftc.io.
  14. Mezinskis, M. Interview on TFTC #804, published 27 October 2026. Quotes checked against the episode audio (8:57–9:19).

Educational content only. Nothing here is investment advice, a price forecast, or a recommendation to buy or sell any asset. Modern Wealth Model is not a registered investment adviser.

Written by Colin Reed

Colin Reed is an operations executive and independent consultant whose career spans foodservice distribution, airline catering, last-mile delivery, and hospitality, from restaurants and bars to luxury resort food and beverage programs. He has held distribution P&L with capital authority and run enterprise system selections across warehouse management, labor management, and ERP platforms, and advises funds, corporates, and management consultancies through expert networks and research firms. On Modern Wealth Model he writes on applied AI in the physical economy, Bitcoin, and long-horizon retirement planning — and builds the calculators.

Full bio · Contact · Consulting