Asymmetric Tail Curvature in Bitcoin Price Quantiles
This report develops an asymmetric quantile framework for Bitcoin's long-run price distribution and uses it to place a familiar observation, diminishing returns across cycles, on a formal statistical footing. After documenting that the dominant power-law and stock-to-flow models have carried systematic optimistic bias out of sample, the paper asks whether allowing the upper and lower tails of the price distribution to curve differently better describes Bitcoin's history. It finds a statistically significant asymmetry: the upper tail compresses meaningfully across cycles while structural support remains close to a straight-line power law. The estimate is a full-sample distributional summary across four halving cycles, intended as a characterization of how Bitcoin's conditional price distribution has evolved, not a forecast, a price floor, or a measure of portfolio loss risk.
Long-run Bitcoin price modeling has been dominated by two families of models: power-law regressions in log-log space and scarcity-based stock-to-flow frameworks. Both fit their calibration periods well, but their out-of-sample records have been poor in a consistent direction. Measured from their publication dates through early 2026, the OLS power law, the stock-to-flow model, and its cross-asset extension show geometric mean price errors of roughly +32%, +295%, and +1,699%, all optimistic, with the error growing as the sample lengthens. This pattern, already noted across the academic and practitioner literature, motivates a shift away from point prediction toward characterizing the full distribution of price given time, and in particular toward asking whether that distribution behaves the same way at its upper and lower edges.
It does not. Estimating the curvature of Bitcoin's conditional price distribution separately for its upper and lower tails reveals a clear asymmetry. The upper tail bends inward over time, with successive speculative peaks falling progressively closer to the structural trend, while the lower tail remains close to a straight-line power law. The difference between the two is statistically significant under block-bootstrap testing, and only the upper-tail curvature is distinguishable from zero. This is the formal counterpart to the diminishing-returns pattern long discussed by practitioners: rather than measuring how far one cycle's peak gain fell short of the last, the framework pools every cycle into a single estimated parameter for how steeply the upper band flattens as the market matures.
The result carries important limits, which the report states plainly. The curvature is identified on the full fifteen-year sample but not on individual sub-periods; with roughly four complete cycles in the record, it is best read as a long-run distributional summary rather than a stable, locally estimable coefficient. Out-of-sample, the asymmetric specification consistently improves on a linear baseline at the upper tail across every training window tested, but its lower-tail and central performance depends on where the training data ends, a window closing near a cycle peak temporarily distorts the structural-support estimate. The lower band describes where price has historically found support, not a guaranteed floor, and the framework characterizes price levels rather than return or loss probabilities.
Taken together, the analysis extends the existing power-law research program rather than replacing it: a near-linear lower tail remains an adequate description of structural support, and the contribution lies in identifying the single parameter the linear specification misses, the compression of the upper tail. A reduced-form model combining a fundamental monetization trend with a decaying speculative premium offers one mechanism consistent with the asymmetry, presented as a plausible explanation rather than the only one. The framework is best understood as a disciplined, distributional way to characterize how Bitcoin's price behavior has changed across cycles, with value that is descriptive rather than predictive.