Where it comes from
Harold Edwin Hurst was a hydrologist. Studying eight centuries of Nile flood records in the 1950s, he noticed that wet years clustered with wet years and droughts with droughts — far more persistence than an independent random process should show. The statistic he built to quantify that persistence, later formalized by Benoît Mandelbrot, is the Hurst exponent (H).
The classic estimator is rescaled-range (R/S) analysis. Split a return series into chunks of size n; for each chunk compute the range of the cumulative demeaned series (R) divided by its standard deviation (S); then observe how the average R/S grows with chunk size. For a self-similar process, E[R/S] ∝ nH — so the slope of log(R/S) against log(n) is the Hurst exponent.
How to read H
- H ≈ 0.5 — random walk. Increments carry no memory. Neither trend-following nor mean-reversion has a structural edge.
- H > 0.5 — persistent (trending). Up moves are more likely to be followed by up moves. Momentum and trend-following strategies are in their element.
- H < 0.5 — anti-persistent (mean-reverting). Moves tend to be reversed. Fading extensions and trading ranges works better than chasing breakouts.
In practice nobody trades the full-sample H of a fifty-year index — the useful object is a rolling window. The dashboard computes H over a rolling 100-trading-day window of SPY closes, so the number describes the market's current character, not its long-run average. Bands at 0.45 and 0.55 separate the three readings: below 0.45 mean-reverting, 0.45–0.55 effectively random, above 0.55 trending.
Why traders care
Most strategy failures are regime failures. A breakout system doesn't stop working because its parameters are wrong — it stops working because the market stopped trending. A rolling Hurst gives you a principled, model-free way to ask "which playbook does the current tape reward?" before deploying one. It pairs naturally with a direction filter (price vs. a long moving average): trending and above trend is a very different regime from trending and below it.
Caveats worth respecting
- It's an estimate, and a noisy one. On 100-day windows the sampling error is material; treat the bands as zones, not switches, and expect whipsaw near the boundaries.
- Window choice is a lag/noise trade-off. Short windows react fast but jitter; long windows are stable but recognize regime change late.
- H is descriptive, not predictive by itself. It tells you the character of the recent past. Combining it with independent inputs (volatility term structure, breadth, credit) is what turns it into a regime framework.