The Hurst Exponent and Long Memory

An ARMA model, built from a handful of short-lag terms, assumes a shock's influence fades away quickly and predictably — an AR(1) coefficient of 0.5 means a shock is nearly forgotten within a few periods. But some real series — river flow, volatility, certain volume and liquidity series — show a stranger pattern: today is correlated with a shock from 50 periods ago far more than a short-lag model would ever predict, and that correlation decays extremely slowly rather than dying out geometrically. This is called long memory, and the Hurst exponent is the classic single-number measure of how much of it a series has.

An analogy: a river's flood memory versus a coin flip's memory

A fair coin flip has zero memory: whether you got heads five flips ago tells you nothing about the next flip. A pure random walk (like an unpredictable stock price) has a specific, moderate kind of memory: it accumulates all its past steps, but each individual step is still independent of the others, so its wandering grows at a textbook, predictable rate with the square root of time. Now consider a river's annual flood level: a multi-year drought tends to be followed by more dry years, and a multi-year wet spell tends to be followed by more wet years, in a way that persists far longer than you'd expect from year-to-year weather being roughly independent. Hurst originally studied exactly this — Nile River flood records — and found that rivers "remember" wet and dry spells over unusually long spans. The Hurst exponent $H$ measures exactly where a series sits on this spectrum: $H=0.5$ for the pure coin-flip-like random walk, $H>0.5$ for river-flood-like persistence, and $H<0.5$ for a series that tends to reverse itself more than randomness would.

The exponent, one piece at a time

The classic estimation method is rescaled range (R/S) analysis. For a time window of length $n$, compute the range of the cumulative deviations from the mean, $R(n)$, and divide by the sample standard deviation, $S(n)$. Hurst's empirical finding, later given theoretical grounding, is that this rescaled range scales with window length as a power law:

$$\mathbb{E}\left[\frac{R(n)}{S(n)}\right] \sim c \cdot n^H .$$

In plain English: as you look at longer and longer windows of the series, its rescaled range grows at a rate governed entirely by the exponent $H$ — plotting $\log(R(n)/S(n))$ against $\log(n)$ should trace a straight line whose slope directly estimates $H$. For a pure random walk, $H=0.5$ exactly, matching the familiar "variance grows with $\sqrt{n}$" scaling. When $H>0.5$, the series has persistent long memory — trends tend to continue, and the autocorrelation between distant points decays so slowly (as a power law, $\rho_k \sim k^{2H-2}$, rather than geometrically) that the sum of all autocorrelations across all lags is actually infinite, unlike a short-memory ARMA process where it's finite. When $H<0.5$, the series is anti-persistent — an up-move tends to be followed by a down-move more often than chance, a strong form of mean reversion.

Worked example 1: reading a straight-line slope

Suppose $R(n)/S(n)$ is measured at $n=10, 100, 1000$ and comes out to roughly 4.5, 14.2, and 45. Taking logs: $(\ln 10, \ln 4.5)=(2.30, 1.50)$, $(\ln 100, \ln 14.2)=(4.61, 2.65)$, $(\ln 1000, \ln 45)=(6.91, 3.81)$. The slope between the first and last points is $(3.81-1.50)/(6.91-2.30) = 2.31/4.61 \approx 0.50$ — this series is behaving like a pure random walk, $H\approx0.5$, no detectable long memory.

Worked example 2: detecting persistence in realized volatility

Applying the same method to a daily realized-volatility series gives $R(n)/S(n)$ values of roughly 6, 28, and 130 at $n=10,100,1000$. In logs: $(2.30, 1.79)$, $(4.61, 3.33)$, $(6.91, 4.87)$. The slope is $(4.87-1.79)/(6.91-2.30) = 3.08/4.61 \approx 0.67$ — well above 0.5, indicating strong persistence: high-volatility periods tend to be followed by more high-volatility periods over spans far longer than a standard GARCH(1,1) model's exponential decay would predict. This empirical regularity — $H$ around 0.6-0.8 for realized volatility across many markets — is precisely what motivates long-memory volatility models like FIGARCH and ARFIMA specifications for volatility (see ARFIMA Long-Memory Models).

What this means in practice

Realized volatility, trading volume, and some liquidity measures across many markets consistently show $H$ noticeably above 0.5, which is why long-memory models are standard for volatility forecasting rather than a niche curiosity. A quant seeing $H\neq0.5$ on a candidate signal should treat it as evidence the series has structure beyond what a short-lag AR or MA term can capture — worth modeling explicitly with a fractional-differencing approach (see ARFIMA Long-Memory Models) rather than papering over with a higher-order but still short-memory ARMA fit.