Hurst Exponent for Bitcoin: What the Data Actually Shows

Is Bitcoin pure speculation, or does it carry persistent memory in its price? Both framings get repeated daily on trading forums and neither has been weighed against a number. The number for this article (the Hurst exponent of BTC-USD over the last 900 daily bars) is 0.56. That is slightly above random walk and slightly above the S&P 500, with a rolling series that has stayed on the trending side of the line 83.3% of the time over the last three and a half years.

What follows is the same window measured four different ways, the rolling view of how that number has changed since the 2024 halving, a side-by-side against SPY and gold, and what 0.56 actually means if you are trying to apply cycle analysis to Bitcoin.

A single number is misleading without context

The Hurst exponent comes from a 1951 paper by H.E. Hurst, a British hydrologist who measured Nile flood data and found that the range of cumulative deviations divided by their standard deviation grew faster with sample size than independent random data would predict. Hurst introduced the rescaled-range (R/S) statistic and the exponent that now carries his name. Worth flagging up front: this is not the same person as J.M. Hurst, the market practitioner whose 1970 book The Profit Magic of Stock Transaction Timing introduced the cyclic-model approach to retail traders. Two different people, two different fields, both relevant to broader cycle work but only H.E. Hurst's exponent matters to this article.

Andrew Lo's 1991 Econometrica paper "Long-term Memory in Stock Market Prices" is the second piece of mandatory context. Lo showed that naive R/S (the version Hurst himself published) systematically over-detects long-memory persistence when short-run autocorrelation is present in the data. Heteroskedasticity, AR/MA structure, and intraday clustering all bias R/S upward. Lo proposed a modified R/S that corrects for it. Any production Hurst calculation that ignores Lo's correction will report higher persistence than the data supports.

A second context problem: a single number averaged over 900 bars smears across regime transitions. The 2023-11 to 2026-05 window covers the SEC's January 2024 spot-ETF approval, the April 2024 halving, the 2025 cycle high above $100,000, and the subsequent pullback. Treating those distinct epochs as a single statistical population is convenient and wrong. The rolling Hurst section below corrects for that.

Bitcoin-specific Hurst work has been done before. Bariviera's 2017 Economics Letters paper "The inefficiency of Bitcoin revisited: A dynamic approach" ran rolling DFA on BTC daily returns from 2011 to 2017 and documented Hurst values well above 0.5 in the early years, declining toward 0.5 as the market matured. The reading below (multi-method consistency 0.90 around a classical R/S of 0.56) is consistent with that arc continuing: BTC's persistence has moderated relative to its early-history values but has not collapsed to the random-walk threshold.

Bitcoin's Hurst, four ways

The same 900-bar window, measured by four established methods:

Two readings sit comfortably above 0.5; two sit within 0.02 of pure random walk. The fractal dimension converts to an equivalent Hurst of 2 - D ≈ 0.57, which puts three of four methods between 0.51 and 0.57. Volatility scaling is the conservative outlier at 0.51.

Multi-method consistency: 0.90. That is a high agreement score on FractalCycles' regime classifier. The methods are not contradicting each other; they are reporting different views of the same neighborhood. Classical R/S is the most assertive ("clearly trending"); DFA and volatility scaling are the most conservative ("statistically borderline"); fractal dimension lands between them.

This pattern is exactly what the multi-method approach is built for. If a single classical R/S reading of 0.56 were the only data point, you might call BTC "trending." If the volatility scaling reading of 0.51 were the only data point, you might call it "random walk." Neither is wrong; neither is the full picture. The honest summary is: Bitcoin sits on the trending side of random walk, by a modest margin, across multiple definitions of "trending." Method choice matters, which is the point of the companion piece on why most cycle analysis fails. The right discipline is to look at several methods, not one. For background on the underlying mathematics, see the Hurst exponent guide.

The rolling Hurst tells the real story

A full-window number is one snapshot. The more informative view is the rolling Hurst: H computed on a moving 252-bar window, stepped every 22 bars, across the full 3.5-year sample. That gives 30 monthly snapshots from 2024-08-05 to 2026-05-05.

The 30 snapshots ranged from 0.50 at the low (2025-06-09) to 0.62 at the high (2025-02-19). Every single one of the 30 came in at or above 0.50.

There is no monthly snapshot in the entire 3.5-year window where Bitcoin's rolling Hurst dipped into mean-reversion. The series spent five months hovering near 0.50–0.55 (a brief deceleration in mid-2025) and otherwise sat above 0.55. The current reading as of May 2026 is 0.5686, trending, in the meaty middle of the historical range.

This is the part the full-window 0.56 conceals. The full-window value is not a chance averaging of trending and mean-reverting regimes that happened to cancel out near random walk. It is a fair summary of a series that consistently lived on the trending side of random walk for the last three and a half years, with normal month-to-month variation but never breaching into mean-reversion. The rolling Hurst calculator reproduces this view on any symbol.

For applied cycle work this matters: a rolling Hurst gate based on the current 252-bar window will produce different decisions than a gate based on the full 900-bar window. The current rolling value sits in the trending zone; the full-window value sits in the borderline-trending zone. Same data, different gates, different decisions. See the market regime detection guide for the thresholds FractalCycles uses.

Bitcoin vs SPY vs Gold over the same window

Bitcoin is often discussed as if it lives on a different statistical planet from traditional assets. The 900-bar window gives a direct comparison:

Bitcoin's Hurst over this window is 0.008 above SPY and 0.026 above gold. That is within the noise band of the R/S estimator at this sample size. Statistically, BTC is indistinguishable from SPY and slightly more persistent than gold at this scale.

This kills two opposing retail narratives at once. The "BTC is pure speculation / random walk" framing fails because BTC's Hurst is in the same neighborhood as the S&P, which nobody serious describes that way. The "BTC is uniquely trending / momentum-driven" framing also fails because BTC's persistence is not meaningfully different from the broadest US equity index over the same period. The honest framing is: BTC, SPY, and gold all sit in the mildly-persistent neighborhood, with BTC marginally the most persistent of the three. Anyone insisting the assets are categorically different at this scale is arguing against the arithmetic. For the longer macro context on Bitcoin specifically, see the Bitcoin beyond halving market analysis.

What H = 0.56 means for cycle analysis on BTC

A Hurst of 0.56 sits in what FractalCycles' pipeline treats as the "regime gate is barely open" zone. The practical implications for anyone applying cycle analysis to Bitcoin:

1. Standard mean-reversion cycle tools (composite projection with fixed-period bands, FLD strategies designed for ranging markets) face an uphill fight on BTC at this window. The market spends 83.3% of monthly snapshots on the trending side of random walk; mean-reverting setups expect the opposite.
2. Trend-following cycle tools (sector-rotation timing, cycle-confirmed breakout entries) get the regime tailwind, but a Hurst of 0.56 is not 0.7. The trend is real but not overwhelming, and overextension is still a meaningful risk.
3. Hurst alone is not a cycle. The exponent classifies the regime; it does not detect periodicity. Bartels-significant cycles can still exist inside a mildly-trending regime, but the cycles ride on top of trend rather than around a flat mean. Detrending choice (HP filter, polynomial fit, first-difference) matters more in this regime, not less. The Bartels significance testing guide covers the math of separating real cycles from filter-induced ones.
4. Use the rolling Hurst, not the full-window value, as the regime gate. Cycle inferences should be gated on the regime as it stands now, not as it averages over three years. Bitcoin's rolling Hurst spent five months near random walk in mid-2025, and those were structurally different from the surrounding trending months.

Frequently asked questions