Hurst Exponent: Measuring Trend and Mean Reversion 2026

A Number Born on the Nile

In 1951, a British hydrologist named Harold Edwin Hurst published a study of a very practical problem: how large a reservoir would need to be to regulate the Nile. Hurst had spent decades in Egypt working on the river, and he had access to an extraordinary dataset, including flood records stretching back over 800 years to the Roda gauge near Cairo. What he found broke the statistics of his day. Wet years clustered with wet years and droughts with droughts, far more than independent random fluctuations could explain.

Hurst measured this clustering with a quantity that now carries his name. The Hurst exponent, H, describes how the range of a cumulative process grows with the observation window. For independent random steps it grows like the square root of time, H = 0.5. The Nile data gave Hurst roughly 0.7: the river had memory.

Benoit Mandelbrot picked the idea up in the 1960s, connected it to fractional Brownian motion and coined the term "the Joseph effect" for this long-range persistence, after the seven fat and seven lean years of Genesis. Quants adopted it for a simpler reason: H puts a single number on the most basic strategic question you can ask of a price series. Does it trend, does it revert, or does it just wander?

---

What H Measures

The Hurst exponent lives between 0 and 1 and partitions time series behaviour into three regimes.

H = 0.5: random walk. Increments are independent. Past moves tell you nothing about future moves, which is precisely the world the efficient market hypothesis describes and the world in which neither momentum nor mean-reversion strategies have any edge. A pure random walk diffuses with a range growing as the square root of time.

H > 0.5: trending (persistent). A move up is more likely to be followed by another move up. The series ranges further than a random walk over the same horizon, because moves reinforce. The closer H gets to 1, the smoother and more trend-dominated the series.

H < 0.5: mean reverting (anti-persistent). A move up is more likely to be followed by a move down. The series is choppier than a random walk and keeps getting pulled back on itself, so its range grows more slowly than the square root of time.

The compact way to say all three at once: the variance of the change over a lag k scales like k to the power 2H. At H = 0.5 you recover ordinary diffusion. Above it, super-diffusion. Below it, sub-diffusion. H is closely related to autocorrelation, but instead of measuring dependence at one specific lag, it summarises how dependence behaves across all horizons at once.

---

Rescaled Range Analysis, Step by Step

Hurst's original estimator is rescaled range (R/S) analysis. It sounds baroque; it is actually five small steps.

1. Take a window of n consecutive observations of your series (returns, not prices), and subtract the window's mean, so the window sums to zero.
2. Cumulate the demeaned values into a running total. This turns the window into a little random walk whose wandering you can measure.
3. Measure the range R: the maximum of that cumulative series minus its minimum. This is how far the walk strayed, peak to trough.
4. Rescale by the window's standard deviation S, giving the dimensionless ratio R/S. This is the step that lets you compare windows with different volatilities.
5. Repeat across window sizes. Compute the average R/S for many window lengths n, then plot log(R/S) against log(n). Hurst's empirical law says R/S grows like n to the power H, so the slope of that line is your estimate of H.

The logic mirrors the worked example from the Nile. For a reservoir, R is literally the storage you would have needed over that window to smooth the flow, which is why Hurst cared about it. For a trader, the interpretation is the useful bit: persistent series accumulate deviations and produce large ranges (steep slope, high H), anti-persistent series constantly cancel their own progress and produce small ranges (shallow slope, low H).

---

Estimating H in Python

R/S is the historical method. In practice many quants use the variance-of-lagged-differences version because it is shorter, faster and tends to be less biased on small samples. It reads the scaling law directly: the standard deviation of k-lag differences of log prices should grow like k to the power H.

import numpy as np


def hurst_exponent(prices: np.ndarray, max_lag: int = 100) -> float:
    """Estimate H from a price series via the scaling of
    lagged differences: std(x[t+k] - x[t]) ~ k**H.
    """
    x = np.log(prices)
    lags = np.arange(2, max_lag)
    sigma = np.array([np.std(x[lag:] - x[:-lag]) for lag in lags])

    # Slope of the log-log scaling line is the Hurst exponent
    slope, _ = np.polyfit(np.log(lags), np.log(sigma), 1)
    return float(slope)


rng = np.random.default_rng(3)
walk = 500.0 * np.exp(np.cumsum(rng.normal(0, 0.01, 2000)))

print(f"Random walk: H = {hurst_exponent(walk):.3f}")   # ~0.5

Run it on a simulated random walk and you should see values scattered around 0.5, typically between 0.45 and 0.55 for 2,000 observations. That scatter is your first caveat in numerical form: even on data that is a random walk by construction, the estimator does not return exactly 0.5. Anyone quoting H to three decimal places on six months of daily data is reporting noise.

Two implementation notes. Estimate H on log prices (equivalently, cumulated returns), not on raw returns, because the scaling law describes the cumulative process. And choose max_lag well below your sample length, since the largest lags have the fewest independent observations and dominate the fit with the least reliable points.

---

A Worked Interpretation

Suppose you compute H on rolling two-year windows of daily data and get these numbers:

The pair spread is the interesting row. H = 0.31 says deviations in the spread tend to undo themselves: stretch it and it snaps back more often than chance. That is precisely the property a pairs trade monetises, and a low Hurst estimate is a fast, model-free screen for such spreads before you commit to heavier machinery like cointegration tests and half-life estimation, both covered in our mean reversion guide.

The equity index row is equally instructive. Decades of research find large liquid indices hover close to 0.5 at daily frequency. Whatever edge exists there is too small for H to detect, which is a sobering benchmark for how efficient the big markets are.

---

How Quants Actually Use It

Almost nobody trades H directly. Its real job is as a regime filter that decides which family of strategy is allowed to run.

The logic: momentum strategies assume persistence and get chopped to pieces in anti-persistent markets, while mean-reversion strategies assume snap-back and get steamrolled in trends. A rolling Hurst estimate gives you a dial for which world you are currently in. A simple implementation computes H over a trailing window (say 250 to 500 observations), enables trend-following signals only when H is comfortably above 0.5, enables reversion signals only when it is comfortably below, and stands down in the ambiguous middle band, roughly 0.45 to 0.55, where the estimate cannot distinguish anything from a random walk.

Used this way, H is one input in a broader time series toolkit alongside autocorrelation tests, variance ratios and stationarity tests, all of which interrogate the same underlying question from different angles. When several of them agree, the regime call starts to mean something.

There is also a portfolio-level use: scanning. Computing H across hundreds of spreads or instruments is cheap, and ranking by H surfaces the strongest reversion or trend candidates for deeper investigation. As a first-pass filter it earns its keep; as a final answer it never should be.

---

Where the Hurst Exponent Misleads

H is a useful number wrapped in failure modes, and the failure modes deserve equal billing.

The estimator is noisy. Confidence intervals on H from realistic sample sizes are wide. With 500 daily observations, a reported H of 0.45 is entirely consistent with a true value of 0.5. Classical R/S is also biased upward on short samples, which is why Andrew Lo's 1991 modified R/S statistic exists, and why different estimators (R/S, variance scaling, detrended fluctuation analysis, wavelets) routinely disagree on the same data by 0.05 or more. Treat the second decimal place as fiction.

Markets are not stationary. A single H for "the S&P 500" assumes one scaling regime governs the whole sample. In reality persistence itself drifts: a series can trend for two years and revert for the next two, and a full-sample estimate averages the regimes into an uninformative 0.5. Rolling estimates help, but shorter windows buy responsiveness with even more estimation noise. You cannot have both.

Sample length and frequency change the answer. H measured on five-minute bars, daily bars and weekly bars of the same instrument will differ, because real markets have different dependence structures at different horizons (microstructure effects alone push short-horizon estimates around). The number is only meaningful alongside its frequency, window and estimator.

H is not a trading signal on its own. It tells you a series has tended to revert; it does not tell you the current deviation, the expected holding period, the transaction costs, or whether the relationship that generated the reversion still exists. A spread can show H = 0.3 over history and then break permanently the day one company in the pair changes its business. H describes the past shape of randomness. Position sizing, entries and risk still have to come from somewhere else.

Hurst spent his career on one river and gave time series analysis one of its most durable ideas: that memory has a measurable signature. The measurement is real. Just remember that a 75-year-old statistic computed on 500 noisy data points is an opinion with error bars, not a verdict.