Cowen model finds Bitcoin cycle tops compressing

Benjamin Cowen published a quantitative paper titled “Asymmetric Tail Curvature in Bitcoin Price Quantiles” using 16 years of daily Bitcoin prices through May 2026. The paper reports that price peaks have compressed: the upper‑tail curvature measures about −0.33 while the lower tail is close to −0.02 and not distinguishable from flat.

Cowen replaces a single power‑law line with quantile regression bands across historical prices. The fitted bands form a fan in log‑log space: higher percentiles that track speculative peaks bend inward over time, while lower percentiles that track structural support follow a steady upward trend. The upper‑tail curvature is statistically different from zero and the difference between upper and lower tails is significant at the 1% level under a bootstrap test.

The paper benchmarks three well‑known price models against the 2019–2026 record and reports systematic overshoots. A historical power‑law fit calibrated through 2018 overshot prices on 77.2% of trading days with an average error about 32.1% above actual. The stock‑to‑flow model overshot on 94.9% of days with an average overshoot of 294.5%. The cross‑asset S2FX variant produced average overshoots of roughly 1,699%, implying modelled peak prices orders of magnitude higher than observed.

At the time of the analysis Bitcoin traded just below $70,000 with a market capitalization above $1.4 trillion. The paper notes an all‑time high near $126,080 in October 2025 and that price was about 44% below that peak at the time of the study.

Cowen identifies a mechanism he calls “diminishing reflexivity.” He describes that when market capitalization grows into the trillions, the same percentage price moves require far more capital than in Bitcoin’s early years, which reduces speculative amplitude while structural demand continues along a power‑law path.

The analysis reports stability of the upper‑tail curvature across 27 expanding historical windows, with values clustered near −0.3 and every window rejecting symmetry between tails. The model is presented as a probabilistic long‑run shape rather than a short‑term forecast.

The paper lists several caveats. Sample size is limited because Bitcoin has completed only four halving cycles. Parameter estimates vary on shorter sub‑windows. Results are sensitive to the data anchor: using January 2010 as the start date instead of January 2009 reduces the measured upper‑tail curvature toward zero, reflecting the influence of thin‑liquidity early years. Cowen also notes that a negative curvature in log‑log space does not imply literal behavior at infinity and that non‑rejection of lower‑tail flatness does not prove the floor cannot be breached. Historical stress events between 2010 and 2015 and the FTX collapse in November 2022 did pierce lower bands.

Cowen provides timing context, reporting the recent cycle top occurred within a week of its historical timing and giving a base case for the next cyclical low in October 2026, with a broader window in the fourth quarter of 2026. The paper recommends using the framework to set expectations about cycle amplitude rather than to generate precise market timing.