Can the Hurst exponent tell when bitcoin is trending?
No. On rolling 60-day windows the standard rescaled-range estimate of the Hurst exponent reads about 0.59 even on shuffled bitcoin returns, which have no memory, so “H above 0.5” flags trends that are not there. Used as a switch between momentum and mean reversion on bitcoin and ether from 2017 to 2026, it did worse than plain momentum, and high readings preceded weaker momentum, not stronger.
The claim
The Hurst exponent measures memory in a series: 0.5 is a random walk, above 0.5 means persistence (trends), below it mean reversion. Caporale, Gil-Alana and Plastun applied rescaled-range (R/S) analysis and fractional integration to bitcoin and three other coins over 2013–2017, found persistence whose degree changes over time, and concluded that trend-following could earn abnormal profits. The trading version is widespread: compute H on a rolling window, follow trends when it is above 0.5 and fade moves when it is below.
What we tested
BTC and ETH daily closes, August 2017 to September 2026, with funding and trading costs. We computed H with five estimators — classic R/S, R/S with the standard small-sample correction, detrended fluctuation analysis, a variance-ratio form and an independent fractal-dimension measure — on 30- to 120-day windows, and first ran each on shuffled returns to see what it reports for a series with no memory. Then we tested 240 regime-switching rules per coin against plain momentum.
Result
Classic R/S averages 0.59 on shuffled returns over 60-day windows and labels 65% of days as “trending”. Corrected estimators centre near 0.46–0.48, and bitcoin’s and ether’s actual readings are indistinguishable from their own no-memory baselines. As a timing tool H works backwards: momentum made money over the following quarter 63% of the time after H above 0.55, and 83% of the time after H below 0.45. The median regime rule trailed plain momentum by 0.64 Sharpe points on bitcoin and 0.89 on ether; on ether, none of the 240 rules beat it. Even H computed on the future window did not help — the problem is the mapping from H to regime, not the estimate.
Rolling Hurst exponent as a regime filter, August 2017 – September 2026
| Measure | Result |
|---|---|
| H from classic R/S on shuffled returns (60-day window) | 0.59 |
| No-memory days labelled “trending” | 65% |
| Momentum profitable next quarter: after H > 0.55 / after H < 0.45 | 63% / 83% |
| Regime rule minus plain momentum, median Sharpe (BTC / ETH) | −0.64 / −0.89 |
What it means
On the window lengths traders use, the Hurst exponent mostly measures its own estimation bias, and where it carries any information it points the opposite way from the textbook reading. A regime filter built on it removes exposure at the wrong times.
What would change the verdict
Long-memory tests on much longer samples, or on intraday data where each window holds thousands of points, are a separate question from the trading switch tested here. A regime filter would have to beat plain momentum on data after this date, after costs.
This note re-tests a claim made by the cited authors on our own crypto market data; the summary of their method is ours. Tested · published · last reviewed . Research, not investment advice.