Bitcoin's Power Law: Weak Structure, Strong Forecasts

Abstract: Bitcoin's price has been described as following a power law (PL) in time, $P \sim t^β$ with $\hatβ\approx 5.7$ over 2010-2026. We test this claim using the Clauset-Shalizi-Newman protocol applied to Bitcoin's tail-relevant distributional series, and develop three principled time-domain adaptations of the protocol. We find that (i) the distributional power law is rejected on UTXO balances and daily |returns|, with lognormal preferred decisively; (ii) the fitted time-domain exponent varies by nearly a factor of three across reasonable shifts of the time origin -- it is not specification-robust in the sense required for a shift-invariant structural reading; (iii) standard residual diagnostics and scale-invariance tests proposed in earlier work cannot distinguish a power law from a multi-component sigmoid stack fit to the same data; (iv) Bitcoin price stands apart in a cross-asset comparison spanning Bitcoin on-chain metrics and traditional asset classes: it is the only series in the nine-series in-sample test where no single-component growth curve improves on the power law, and the quarterly $K=3$ wave-stability bootstrap rejects the PL+AR(1) null on Bitcoin at $p = 0.015$ (strict 15% CV threshold) -- a clear cross-asset separation, although not a Bonferroni-robust rejection; and (v) walk-forward Diebold-Mariano evaluation against ten candidates -- including standard time-series baselines (RW with drift, auto-ARIMA, ETS, local-linear-trend) -- shows the in-sample winner (multi-sigmoid) is among the worst long-horizon forecasters, while the simple power law dominates 12-24 month horizons against every standard baseline at $p < 0.05$, precisely because it does not commit to specific wave shapes. The fit-prediction tradeoff is the practical counterpart of the descriptive findings.

• The paper finds that the conventional power-law model for Bitcoin's price is highly sensitive to time origin shifts, questioning its invariant structural interpretation.
• Methodologically, it compares distributional tests, residual bootstrapping, and multi-component sigmoid fits to reveal superior long-horizon forecasting performance for the simpler power law.
• The study demonstrates that Bitcoin's price exhibits a distinctive multi-wave structure with significant cross-asset differences and leads on-chain adoption metrics.

Bitcoin’s Power Law: A Specification-Robustness and Predictive Utility Analysis

Introduction and Motivation

The empirical observation that Bitcoin’s US-dollar price trajectory from 2010 to 2026 is well-approximated by a power-law relationship $P \sim t^\beta$ (with exponent $\beta \approx 5.7$) has attracted substantial interest in both academic and practitioner domains. On log-log axes, the daily price profile appears as a quasi-linear trend across nearly six orders of magnitude, yielding high $R^2$ in OLS regressions and motivating mechanistic rationales rooted in network adoption and Metcalfe-like scaling.

Figure 1: Bitcoin’s daily closing price, 2010–2026, on log-log axes; the power-law fit (dashed red, $\hat\alpha = 5.64$) yields $R^2 = 0.96$.

However, prior statistical examination of such power-law claims, especially in complex systems, underscore that apparent linearity on log-log scales is neither necessary nor sufficient for demonstrating genuine scale-free structure. The core objective of this paper is to rigorously evaluate the statistical robustness and predictive value of the power-law hypothesis for Bitcoin’s price evolution, employing both distributional and time-domain methodologies, cross-asset comparisons, and careful out-of-sample forecasting analysis (2605.21316).

Distributional Versus Time-Domain Power Laws

Distributional Tests Using the Clauset–Shalizi–Newman (CSN) Protocol

Application of CSN’s four-step protocol to two relevant Bitcoin series — UTXO balance cross-sections and the distribution of daily $|$returns$|$ — yields decisive rejections of the power-law hypothesis. Across seven yearly UTXO snapshots and four periods of daily returns, eight of eleven tests (bootstrap goodness-of-fit with $B=200$) reject the power law at $p < 0.05$, with lognormal forms substantially preferred by Vuong likelihood-ratio comparisons.

The two instances in which the test fails to reject the power law are direct consequences of sample size limitations in the extreme tails (truncation artifacts rather than statistical support). Thus, for distributional aspects of Bitcoin, lognormal models are consistently superior, congruent with established critiques in the empirical power-law literature.

Time-Domain Power Laws: Specification Sensitivity and Model Indistinguishability

The time-domain claim that $P \sim t^\beta$ is fundamentally a regression statement rather than a distributional tail-law. The authors adapt CSN logic to this setting through: (i) shift-sensitivity analysis (variation of fit with the choice of time origin), (ii) residual diagnostic bootstrapping, and (iii) Vuong tests against alternative functional forms.

Key findings:

• The fitted exponent $\beta \approx 5.7$0 is highly sensitive to the shift parameter $\beta \approx 5.7$1 in $\beta \approx 5.7$2: for $\beta \approx 5.7$3, $\beta \approx 5.7$4; for $\beta \approx 5.7$5, $\beta \approx 5.7$6 (Figure 2). No specification-robust invariant exponent exists.

  Figure 2: Fitted power-law exponent $\beta \approx 5.7$7 and AIC versus shift parameter $\beta \approx 5.7$8. Both grow monotonically, indicating high specification dependence.

• Standard residual diagnostics and scale-invariance tests are non-discriminative: when applied to synthetic data generated by a 3-component sigmoid stack (besides a power law), they “pass” identically, demonstrating that such tests only assert long-run average slope stability rather than structural uniqueness.
• Flexibility-controlled in-sample comparisons show that the multi-sigmoid ($\beta \approx 5.7$9) model outperforms the power law decisively ($R^2$0AIC = $R^2$1), with the superiority robust to parameter-matched polynomial and spline controls (which still fall short of the multi-sigmoid fit).

Cross-Asset Model Comparison and Structural Wave Detection

To discern whether Bitcoin’s apparent deviation from simple functional forms is idiosyncratic, the paper examines eight additional series: five Bitcoin on-chain metrics, NASDAQ, S&P 500, and gold.

Figure 3: Cross-series in-sample AIC deltas of various functional fits relative to power law. Only Bitcoin price exhibits no single-component model (e.g., K=1 sigmoid, pure exponential) that improves on power law.

• For all non-price series, a single-component model (sigmoid or exponential) surpasses the power law in AIC. Only for Bitcoin price does multi-component structure decisively dominate.
• The K=3 wave-stability bootstrap procedure is introduced to statistically test for persistent (across rolling windows) sigmoid components in K=3 fits, using a PL+AR(1) noise null model.

  Figure 4: Top—Bitcoin price and K=3 multi-sigmoid fit. Bottom—decomposed sigmoid components. Inflection dates demarcate the onset of each adoption “wave.”

• At quarterly resolution (2016Q1–2026Q1, $R^2$2), Bitcoin price consistently yields two stable (low-CV amplitude) sigmoid components, rejecting the PL+AR(1) null at $R^2$3 and $R^2$4 (Figure 5).

  Figure 5: Distribution of stable K=3 components under PL+AR(1) null (blue), with real Bitcoin (red) marking the tail outcome. Bitcoin exhibits statistically significant component stability.

• This rejection is unique to Bitcoin price across the asset comparison group; Ethereum is the only other series with marginal signal, but not significant at conventional levels.

  Figure 6: Parallel coordinates of per-wave amplitude CVs for synthetic nulls (blue) versus real Bitcoin price (red). Real Bitcoin amplitude stabilities are outliers in the null distribution.

Direction of Causality Between Price and Adoption Metrics

Using Granger causality and cross-correlation functions (on stationary, first-differenced log series), the paper establishes that price changes systematically precede corresponding changes in all five tested on-chain adoption metrics — with typical lead times of 2–8 weeks. The effect is consistent, though of moderate size (F-statistic asymmetry ratios $R^2$5), and not reversed in any window.

Figure 7: Cross-correlation functions between dlog(price) and dlog(on-chain metrics) for lags ±90 days. Peaks at negative lags show price leads adoption metrics.

Out-of-Sample Forecasting: Long-Run Utility of Power Law

The practical relevance of structural detection is evaluated via walk-forward forecasting, comparing the power law, K=3 and K=1 sigmoid models, flexible polynomials/splines, naïve (no skill), and four standard time-series baselines (RW-drift, auto-ARIMA, ETS, local-linear trend) over 11 yearly cutoffs and horizons from 1 to 24 months.

• At short horizons (1–3 months), the naïve “carry-forward” forecast is optimal (lowest RMSE), strongly winning against all parametric models (including standard time-series methods).
• At long horizons (12–24 months), the simple power law dominates, significantly outperforming every baseline (including K=3 multi-sigmoid, all time-series models) at $R^2$6 in Diebold–Mariano significance tests, with RMSE reductions exceeding $R^2$7 against alternatives.

  Figure 8: Walk-forward out-of-sample mean RMSE on log-price. Power law outperforms all competitors for $R^2$8 months; K=3 multi-sigmoid (in-sample winner) generalizes poorly.

• Notably, the in-sample advantage of the multi-sigmoid is not predictive: such models have no mechanism to anticipate subsequent “waves” not observed in the training data and revert to their asymptote, leading to poorer long-horizon generalization.
• This constitutes a concrete instance of a fit–prediction tradeoff: the model best describing historical structure does not ensure optimal predictive accuracy.

Implications, Theoretical Discussion, and Forward Directions

This work rigorously demonstrates that while a power-law envelope characterizes the long-run average slope of the Bitcoin price trajectory, such a relationship is not specification-robust (the exponent is highly origin-dependent) and is not structurally unique (multi-component sigmoid models with equivalent flexibility fit as well or better in-sample and are indistinguishable by standard diagnostics).

However, cross-asset and wave-stability bootstraps robustly indicate that Bitcoin price possesses distinctive multi-component “wave” structure—uniquely persistent compared to all benchmark metrics/assets tested, and not explainable by PL+AR(1) noise. Nonetheless, the forecast utility of such multi-wave description is limited: point forecasts at horizons relevant for valuation/portfolio allocation (1–2 years) are best grounded in simple power-law extrapolation, as models that attempt to fit individual waves lack predictive value due to their inability to anticipate future regime transitions.

Practically, this study provides a horizon-dependent prescription: use naïve methods for $R^2$9 months, switch to power law for $\hat\alpha = 5.64$0 months, and avoid flexible over-fitters for long-run forecasting. The results also clarify that mechanistic or speculative narratives (e.g., Metcalfe-derived exponentials, network models) must account for both the empirical multi-wave structure and the predictive irrelevance of highly flexible fits.

Work remains in developing envelope-based structural models that combine descriptive fidelity with out-of-sample stability; the present negative-control bootstraps and formal Bonferroni correction across assets leave open avenues for further theoretical and empirical reinforcement.

Conclusion

This research establishes that claims of a power-law generative mechanism for Bitcoin price evolution are not supported by rigorous statistical evidence: the exponent is specification-dependent rather than invariant, and standard tests fail to distinguish true power-law relationships from plausible multi-component alternatives. Yet, in terms of long-horizon predictive performance, the power law is pragmatically superior due to its effective averaging across cycles—solidifying its use as a baseline for valuation projections, conditional on continued environment (multi-wave) persistence. The framework, cross-series methodology, and bootstrapped testing implemented here present a robust template for specification testing in other high-dynamic-range financial and social systems (2605.21316).