Hurst Exponent

A statistical estimate of whether a series has tended to trend, mean-revert, or wander randomly — with wide error bars.

What it is

The Hurst exponent (H) is a statistic from time-series analysis, originally developed by hydrologist Harold Edwin Hurst in the 1950s to study Nile river flows, and later adopted in quantitative finance. It attempts to classify the memory of a series on a 0-to-1 scale: H above 0.5 suggests persistence (moves tended to be followed by moves in the same direction — trending behavior), H below 0.5 suggests anti-persistence (moves tended to be followed by reversals — mean-reverting behavior), and H near 0.5 is consistent with a random walk, where past direction carried no information. That is the honest answer to what it tells you: it describes which of those behaviors the sampled history statistically resembled. What it does not do is certify a regime or predict one. The estimate is exactly that — an estimate, with substantial sampling error at realistic data lengths — and a market that measured persistent last quarter is under no obligation to stay persistent. Much trading content presents Hurst values with false precision, quoting H = 0.63 as if the second decimal were meaningful; on a few hundred bars it usually is not. This entry is educational only, not financial advice, and all trading carries risk of loss.

How it works

The classical estimator is rescaled range (R/S) analysis. Take the returns of the series over a window of n observations and compute their mean. Build the cumulative sum of deviations from that mean — a running total of how far the series drifted from its average behavior. The range R is the difference between the maximum and minimum of that cumulative series, and S is the standard deviation of the returns. The rescaled range R/S measures how far the cumulative deviations wandered relative to the series' own volatility. Hurst's insight was that for many processes, R/S grows like n^H: for a random walk H ≈ 0.5, for a persistent (trending) series the cumulative deviations wander further (H > 0.5), and for an anti-persistent series they keep pulling back (H < 0.5). A proper estimate computes R/S at multiple window sizes and fits the slope of log(R/S) against log(n); the slope is H. Chart implementations, including the example below, often use a single window, estimating H as log(R/S)/log(n) — a rougher, biased shortcut, and we label it as such. The deeper caveat applies to every variant: R/S estimates are noisy and finite-sample biased (short samples of a true random walk routinely measure H well above 0.5), and volatility clustering in real markets contaminates the reading further. Treat any single number as a fuzzy indication, not a measurement.

How traders read it

• H > 0.5: over the sampled window, moves tended to follow through — the history statistically resembled a trending (persistent) series. It is a description of the sample, not a certified regime.
• H < 0.5: moves tended to be followed by partial reversals — the sample resembled mean-reverting (anti-persistent) behavior.
• H ≈ 0.5: the sample is consistent with a random walk. Note that 'consistent with' is weak — short samples of trending or mean-reverting processes can also land near 0.5.
• Respect the error bars: on typical chart lookbacks (a few hundred bars), the sampling error on H is large — often of order ±0.1 or worse. Differences like 0.55 versus 0.60 are usually noise, not signal.
• Watch stability rather than level: an H estimate that stays on one side of 0.5 across several window lengths and time periods is more informative than any single reading.
• Never trade the number directly: even a well-estimated H describes past memory structure. It assigns no probability to the next bar and is context, not a signal.

Common settings

Chart implementations typically estimate H over a rolling window of 100 to 500 bars, with a reference line at 0.5; longer windows give statistically steadier but slower estimates. Serious estimation uses multiple sub-window sizes with a log-log regression slope, and researchers often prefer alternative estimators (detrended fluctuation analysis, variance-ratio tests) because plain R/S is biased upward on short samples. Whatever the tool, the window length is the dominant setting: it decides both what period you are describing and how wide the error bars are. There are no 'best' settings, only trade-offs between recency and statistical reliability.

Strengths

• Asks a genuinely useful question — did this market's recent history look trending, mean-reverting, or random? — in a single standardized number.
• Grounded in real, documented statistics with decades of academic literature, not trading folklore.
• Regime description on a common scale allows like-for-like comparison across instruments and timeframes.
• Pairs naturally with strategy-class thinking: trend-following logic assumes persistence, mean-reversion logic assumes anti-persistence, and H makes that assumption explicit and checkable.
• Honest exposure to estimation uncertainty is itself educational — it teaches why single-number statistics from short samples deserve skepticism.

Pitfalls to watch

• False precision is the central trap: at realistic sample sizes the estimate carries wide error bars, so quoting H to two decimals — as much trading content does — dresses noise up as measurement.
• Finite-sample bias: classical R/S analysis measures H above 0.5 on short samples of a true random walk, so naive readings systematically overstate 'trendiness'.
• Regimes drift: even a well-estimated H describes the sampled window only; markets shift between trending and ranging behavior, and the estimate reacts with a full window of lag.
• Method sensitivity: R/S, DFA, and variance-ratio estimators can disagree noticeably on the same data, and single-window chart shortcuts (like the example below) are rougher still — know which variant your tool uses.
• Volatility clustering and fat tails violate the assumptions under which H has its clean interpretation, contaminating estimates on real financial data.
• No predictive power: H summarizes the statistical memory of past returns. It does not forecast continuation, reversal, or anything else, and no statistic removes the risk of loss.

Pine v6 example

//@version=6
indicator("Hurst Exponent (single-window R/S estimate)", overlay = false)

// Rough single-window estimate: H ~ log(R/S) / log(n).
// A proper estimate regresses log(R/S) on log(n) across many window
// sizes; this shortcut is biased and noisy — treat it as illustrative.
length = input.int(200, "Window Length", minval = 20)

ret     = math.log(close / close[1])
meanRet = ta.sma(ret, length)
sd      = ta.stdev(ret, length)

float cum    = 0.0
float maxCum = 0.0
float minCum = 0.0
for i = 0 to length - 1
    cum += ret[length - 1 - i] - meanRet
    maxCum := math.max(maxCum, cum)
    minCum := math.min(minCum, cum)

rs = sd != 0 ? (maxCum - minCum) / sd : na
h  = not na(rs) and rs > 0 ? math.log(rs) / math.log(length) : na

plot(h, "Hurst estimate", color = color.purple)
hline(0.5, "Random-walk reference", color = color.gray)

Pro tip: Before reading anything into a Hurst value, calibrate your skepticism: generate a pure random walk of the same length and run the estimator on it a few times — the spread of values you get on data with no memory at all is your effective error bar. If your market's reading falls inside that spread, the honest conclusion is 'indistinguishable from random', however tempting the 0.61 looks. Use H to frame which class of behavior recent history resembled, never as a signal. Educational content only, not financial advice; past statistical structure does not predict future prices, and all trading involves risk of loss.

Educational only — not financial advice, not a recommendation to trade. No indicator is predictive; trading involves substantial risk of loss.